One of the issues that I have with the current generation of “AI” products is that they do not appear to take their own premises seriously. I look at a plethora of obsequious chatbots claiming to be serious tools for problem solving, and I think, this is not what a problem-solving tool would look like.

Even before we get to the tremendous ethical problems with the frontier labs, it is this impression of their composition as a product that makes me feel, constantly, whenever I am interacting with them, that they are less a software product than that they are a grift, a scam designed to make me feel like I am interacting with a product that has capabilities that it simply does not, to try to lull me into a false sense of security that I can trust it.

The frontier labs are of course the worst offenders, but every criticism here applies just as much to Ollama, which (if anything, due to the obviously poorer quality of the available models themselves) needs these features even more than the frontier labs do.

Here, I will set down a few features that might convince me that an LLM-based product, particularly one focused on research or software development, was actually serious about helping me do useful things with it.

Make “Checking For Mistakes” A First-Class Feature

This is the biggest issue, and the major reason that I was inspired to write this post.

It is a truth universally acknowledged, that AIs cannot reliably provide information.

I could cite a ton of news articles and studies about this fact, but there is no need. Every single chatbot admits this, up front, in a fine-print disclaimer as a core part of their user interface. Gemini says “AI can make mistakes, so double-check responses”, Claude says “Claude is AI and can make mistakes. Please double-check responses.¹” ChatGPT says “ChatGPT can make mistakes. Check important info.”.

Every time I see that last one, I wonder how I’m supposed to know what “info” is supposed to be “important”.

All of these warnings are all small, gray text, painfully obviously included as legalese to push responsibility back onto the user rather than to help with anything. This is a core limitation of all these products. Checking their output is a part of the workflow for using them that:

1. you absolutely cannot skip or skimp on without creating risks to yourself and whoever you are conveying its output to, and,
2. it is very easy to skip or skimp on and you are encouraged at every turn to do so, because “just trust the output” is one of the quickest ways to save time.

A chatbot product that took this weakness seriously, as an actual consideration for using it, would put a checkbox next to every claim in its output. It would be a 2-column worksheet, where you’ve got the LLM output in the first column, and next to it, human notes in the second column, explaining what work went into checking this claim, and a big checkbox that you would only check off after you believe you’d checked its claims thoroughly enough.

Coding assistants would need to have some version of this as well. Right now, this is pushed off into code review, which means it is a dark pattern which subtly encourages the “author”² to offload this work to their code reviewer without ever looking. Once again, “it’s probably fine, I don’t need to check” is the quickest way to save time and churn out those PRs faster.

It might even be useful for coding harnesses to have some affordance for checking code before it even runs tests. As the vendors themselves have admitted, it’s not just expensive to burn tokens on your “AI”, you also end up burning far more compute on the AI. Being able to check your diffs before sending them over to uselessly exhaust your testing compute cluster would be useful.

If your product tells me that it makes mistakes and I must be the one to check for the mistakes, but then gives me zero tools to check for mistakes, I cannot take it seriously.

More Citations to Check, And More Details

Most chatbots prefer to give an answer, rather than a citation. In my own personal use, I find that when asked to provide a list of citations with clearly marked sources for each one, they will appear to “get bored” halfway through the list and simply stop including citations at some point.

When the bots include citations at all, present them as inline annotations that say nothing but the domain name of the search result, in a font so small that it’s barely legible, and an equally indecipherable icon that is fewer than 16 pixels on a side.

This is backwards.

Now, I am aware that these citations do come from somewhere, and in an attempt to reduce hallucinations, all of the major providers support some form of “grounding”³, and that those little barely-readable citation links are referencing actual structures in the RAG pipeline and not just potentially-hallucinated tokens, but I’m not talking about the underlying machinery in the model, I’m talking about the presentation to the user.

Plus, regardless of whether a snippet of text came from a RAG query, we know that LLMs can never provide an authoritative result; it’s a fundamental limitation of the technology. They can still garble the results of RAG as much as they can misrepresent any other training data. This means that it must never present its results as authoritative.

If you ask an AI to do research queries, every result should be presented as a list of citations. Moreover, the presentation should display each citation as a large object of in its own right, with clearly identified metadata, including not just the site where it was found but its publication date and, if possible, the name of the author. The literal, unmodified quotation (not from RAG, not a summary: a quotation extracted with a regular program and not an LLM) should be front-and-center, larger than any AI-generated text.

If the AI product wants to editorialize or summarize (which should not always be necessary!), the AI-generated text should be presented as small text underneath the citation that has been found, de-emphasized as much as the disclaimer is right now, at the very least until the user has verified that the summary is accurate. Perhaps, for a research project, a “did you read the citation” checkbox might even be helpful.

If your product openly tells me that it will scramble, misrepresent, or omit its citations in its summaries, and I must read the original human-authored citations to be sure, but then gives me no tools to track my reading of those citations or even any way to find them, I cannot take it seriously.

No First-Person Output, No Apologies

There is no reason for a software development or research tool to use first-person language to describe itself. They should not do so. In fact they should not be allowed to do so.

There is also no reason that they should ever apologize. It is a waste of everyone’s time; it’s a waste for the chatbot to generate the apology, it’s a waste for the user to read the apology, and it’s a waste for the user to respond to the apology. Yet they unfailingly do this upon every correction.

The vendors of these tools know that they are routinely causing mental-health crises. In response, they have added non-functional “guard rails” that can still, in 2026, easily be bypassed.⁴

A product seriously interested in helping with productivity would correct this glaringly obvious flaw, focus on the task at hand, and stop emitting useless verbiage.

In the previous two sections, I tried to focus on ways in which the harness would be constructed differently even if the LLM technology is fundamentally impossible to improve; in this case, I have to assume that the labs have some control over the model itself. But unless they are truly incapable of influencing their output (and all their “benchmarks” and “capabilities” seem to indicate that they can control it very tightly) they ought to be building models that are much less verbose.

More Non-Natural-Language User Interfaces

Although natural language could hypothetically be a powerful interface for interacting with a computer system, the practical upshot of LLM natural language interfaces is that these interfaces are imprecise and repetitive, full of superstitions masquerading as “best practices”. The inputs are a mess and the resulting outputs are a mess.

The general way of addressing this unstructured mess is to allow the chatbot to directly take action in response to the user’s input; in other words to supply it with “tools” via an MCP server. But again, this is backwards. If we cannot even express our intent clearly in the first place, why are we trusting this system to take potentially destructive and harmful actions on our behalf?

Instead, I would expect a product that was seriously invested in helping me accomplish specific tasks, to have user interfaces specific to those tasks. Is it supposed to be able to be a security scanner that can discover OWASP top 10 bugs in a codebase? Have a button for that. Build that functionality into your harness, train it directly into the model, use smaller models that can satisfy that functionality more effectively than throwing it at the planet-sized brain of Fable or whatever.

I’m aware that there are small software startups that do something like this, but they are bolted on to the side of the main model providers’ APIs, not integrated into the core of the product and not using their own models and AI systems to achieve consistent and repeatable results.

Strong Data Provenance Indicators

Chatbots produce data tables pulled from websites, from APIs, from MCP tools or from summarizing and scrambling the user’s input. In order to provide the illusion of a seamless interface, this data is presented in-line regardless of where it comes from. But some of these outputs are produced mechanically via regular old API calls, for example, from the result of calling a tool or querying a website, but presented uniformly.

But there is a huge difference between an authoritative data source being inlined as part of a chatbot conversation, being treated as input by the chatbot, and some ad-hoc hallucinated data being treated as output of the chatbot.

If a product is trying to help me make accurate, empirically-grounded, data-driven decisions, the source of the data is critical.

Integrated into the “check for mistakes” and “verify citations” workflow I described above, there’s a necessary “verify data programmatically” pass as well; to have tools that will treat portions of the output as a regular spreadsheet, allowing regular-old computer arithmetic to verify things and showing where such arithmetic was used, and how.

Better User Control of Reproducibility

Anyone familiar with the technical specifics of LLMs will know that they have a variable called “temperature” which controls the degree of randomness that the LLM uses to produce its outputs. But most users don’t know this, because it isn’t exposed as part of the user interface by default.

This leads to a subjective impression that you asked ChatGPT, and you got ChatGPT’s authoritative answer.

You can’t just set the temperature to zero and still get useful results - I am aware that it does more than just scramble the output at random, and there are perhaps good reasons that simply exposing just a temperature setting would not be that useful to users. But if we followed some more of my earlier recommendations for making more structured UI elements to solve specific problems rather than having long back-and-forth chats where each refinement depends on the previous response, perhaps those elements could also re-play the process so that users can see how reliable the bot is at a particular task and develop a sense of how the stochastic nature of the process actually affects it.

Similarly, if a user is trying to solve the same problem repeatedly with a chatbot, and the chatbot product has numerous computational tools that don’t really have anything to do with the LLM, such as deterministic data-processing tools, then having a way to freeze the non-deterministic parts of the transcript but re-populate a particular data frame with updated information and fork / continue the conversation from there would be a way to avoid introducing pointless additional randomness when you already know what tool you’re trying to use.

The fact that every conversation is presented as this flat chat prompt that doesn’t let me interact with any of the widgets that were previously produced except through more chatting, really makes me feel like the whole product is just doing predatory social-media style “increase time on site” optimization, just trying to lure me into further repetitive and unreliable chats, rather than letting me get in, solve my problem, and get out.

Context Visibility

Managing the LLM context is the ongoing challenge facing organizations that are trying to use “agentic” workflows. Filling up the context with too much information causes well-known problems. In response, advanced LLM users attempting to solve larger problems must break up very long prompts into “skills”, give access to lengthy information via “tools”, and delegating sub-problems to “sub-agents” rather than simply extending a single prompt indefinitely.

All of these strategies have flaws, because even on the largest models, compared to the breadth and depth of knowledge-work problems, LLM contexts are quite small.

And yet, none of these products will show the context to the user by default. There are third-party addons that can show you a simple progress bar but for addressing the premier engineering difficulty with this technology, that is below the bare minimum.

This lack of visibility means that almost all of the tools for extending the context are flying blind. Rather than responding meaningfully to a full context, everyone just kind of guesses how much state they need by guessing and trying over and over again with progressively more elaborate skill and sub-agent layouts. Even managing context compaction ends up being an advanced API-driven workflow⁵.

A serious product that was trying to help the user understand would not only show “available context” but explain the impact of context compactions, make it easier to see harness-generated prompts, and so on. This would be a first-class feature, combined with the aforementioned reproducibility / replay tools, would allow users to do real experiments to develop an understanding about how to make good use of the context window.

A Sandbox That Actually Works

I’ve been focused on the chatbot interface here because it is the most immediately egregious upon looking at the UI. But the “agentic loop” tools used for coding are equally dangerous, if not more so. Coding tools keep destroying everyone’s data, over the course of years.

These catastrophic incidents that become front-page news are relatively rare compared to the amount of coding-agent use out there. But they also aren’t the only kind of sandbox violation. Coding models will so routinely edit test code instead of the system under test that there are “pro tips” articles all over the web giving you the flawed advice to simply ask the agent not to cheat. News write-ups of the catastrophic incidents themselves will also offer glib and wrong advice, like “use a docker container”. That might prevent it from literally deleting your operating system, but it won’t prevent it from destroying all the local work you have in your codebase (it needs access to a checkout, after all!)

There is a flurry of activity in the infosec space where people are rushing to plug the gaps left by these coding harnesses. Everyone’s got their own version of an MCP approval gateway where you can optionally place a proxy between your agent and your production infrastructure.

In the best case, though, all these mitigations and proxies and prompts simply turn the user into an auto-approval automaton, hitting Y, Y, Y, Y over and over again, until you finally are driven mad and hit “yes to all”, turn on full-auto mode and submit yourself to the void. With nothing between your personal vigilance and disaster, there are no workflows left beyond decrementing your own vigilance until there’s nothing left and then hoping the disaster never arrives.

The fact that some mitigations exist that can be deployed by extra-cautious users does not change the fact that “agentic coding” is an unsafe-by-default technology deployed without concern or guidance. Every frontier lab has tied a spring-loaded shotgun to a dog; the fact that dog owners can publish thoughtful blog posts explaining how you can teach your dog the basics of gun safety or how you can have your dogs play in a bullet-proof room does not mitigate the fact that the product should not have been allowed in the first place, nor should it continue to exist without VERY strong security controls.

I might believe that a frontier lab were seriously interested in providing developers with a useful tool if they shipped something that had safety built-in.

That means tools in the harness, detached from any LLM, independent of the prompt, that could:

• sandbox all filesystem operations and strictly limit ANY deletions outside of specified scopes, regardless of operating system,
• enforce snapshotting of the entire repo on every operation for easy rollbacks and minimal lost work,
• remove the disaster of “auto mode” (not to mention nonsense like --dangerously-skip-permissions) entirely, and
• carefully consider a structure for presenting plans to the user where, rather than provoking immediate alert fatigue by asking for checks on every action, make structured plans which can be submitted to the user as a group of actions and reviewed and approved as a batch.

In the same way that I suggested above that research-based tasks should have a way of re-issuing prompts to determine how reproducible a result is, or whether other sources might be found, agent-based tasks should have a way of being executed against mock services for popular APIs, so that the verification can match both on the front-end (review the plan for making the API calls before they’re executed) and the back end (review the API calls that were issued to the mock service and verify that they matched).

Instead, the frontier labs provide us products that are disasters out of the box, give us “best practices” to build massive and elaborate, as well as incomplete and error-prone, security perimeters of our own design. Then they blame “operator error” when it inevitably goes wrong. I cannot believe that these design choices are intended to help us be productive.

Bonus: Human Processes

Organizations deploying AI also frequently come across as unserious, for similar reasons. In 2023, naive exuberance could perhaps be forgiven. But today, as we near the close of 2026, there are several well-known problems, that have been extremely well-covered in the press. None of these things should be surprising, but most orgs deploying these tools are still just letting them rip and hoping it all works out.

Organizations deploying these tools would need at least three kinds of major modifications to their internal processes, if they wanted to be serious about using them safely:

1. Shift Rotations to Prevent Vigilance Decrement

There have been several high-profile incidents where software developers’ gradual acquiescence to accepting LLM output have lead to serious economic consequences for the companies deploying them, perhaps best typified by Amazon’s “millions of lost orders” due to a gradual decay of their engineering processes from LLM use.

These outages, and other AI-related failures, are due to the difficulty of maintaining focus on the same problems. In other words, as I described above, vigilance decrement is a constant problem, because AI outputs are most often correct, but continue to be incorrect in surprising and non-intuitive ways. As I have previously written, you cannot trust yourself to catch every bug with code review, and LLM output.

Aviation, for example, has very strict rules around rest requirements. There is also a specific rule that “No certificate holder may operate an aircraft without a second in command if that aircraft has a passenger seating configuration, excluding any pilot seat, of ten seats or more.”. Other safety-critical professions have similar rules.

And yet, even in the age of the supposed “AI revolution”, most software teams are still assigning every engineer a full feature load, not planning for any rest, and telling people to review code whenever they happen to have some “free time”.

Maintenance of vigilance has to be your top priority. Regular, scheduled, inviolable rest periods where people do work without AI assistance, and are not exposed to any AI output for review or otherwise, would be crucial in order to stay mentally sharp enough.

The tools themselves should have this sort of thing built in. The mistake-review process described above should have a periodic spot-check mode where a second reviewer periodically reviews a chatbot log, doing their own independent verification of claims, to see if they spot the same errors. This could provide a feedback loop to determine how much rest is necessary to maintain continuous attention and actually spot hallucinations.

2. Skill Practice To Prevent Skill Loss

It is also well-known that AI use leads to AI reliance, and AI reliance leads to skill loss.

I like to use the analogy to dockworkers at a seaport⁶ adopting automation.

If you employ dockworkers to load and unload ships all day long, they are going to be getting tons of exercise. They will be able to lift heavy objects on demand, whenever. They might have plenty of health problems and injuries from this type of work, but “lack of exercise” will not be a problem.

With the development of standardized container ships and mechanized cranes, you are going to be changing their job description substantially: now they mostly spend all day sitting in a small cubicle moving a control lever back and forth, not lifting heavy stuff. They will get worse at lifting heavy objects.

In this analogy however, the cranes are not all that reliable. We know they break, and they drop their payloads sometimes, and the stuff needs to be manually moved. But this only happens a few times a week, at most. If you need whoever is driving the crane to be able to jump out at any moment and still move stuff around manually, then you need to make an affordance for that. You need to give them time to go to the gym and do some lifting for practice, or every crane failure is going to be a major emergency.

An organization doing an AI transformation would also need a massive increase to learning & development budget, both in terms of resources and in terms of schedule. If your people are going to lose skills because they’ve lost regular practice in the incidental course of doing their duties, then they are going to need deliberate, intentional, non-incidental practice of those skills to keep them sharp.

But rather than trying to accommodate new workflows and give time for people to adjust, most AI mandates are simply dropped on workers like a ton of bricks, with no time to adapt and no affordance for maintaining their skills. Operate the crane and stay fit and healthy and ready to switch back to manual lifting at any time and then get back in the crane cockpit right afterwards. Don’t mess up.

Then an accident happens and everyone is surprised, as if this process weren’t practically designed to produce a terrible result.

3. Mental Health Resources to Deal with Mental Health Risks

AI psychosis often begins with practical problem-solving, and beyond that, it can start specifically at work. Not to mention the more pedestrian condition of “AI brain fry”.

If you are mandating your employees to use a hazardous tool that may seriously and directly damage their mental health, you need trainings and resources. You need in-house therapists and you need to be making sure to check in with people actively to make sure that this is not happening.

Again, the tool itself ought to have some way of dealing with this. An occasional “take a break” popup is easily dismissed; they need a user-visible AI personal dosimeter so you can see your cumulative usage over time.

I don’t even know if “usage over time” is a sufficient metric to gauge risk. Maybe if your work chatbot start to talk about resonance too much, unless you literally work as an acoustic engineer, that should be flagged for someone.

We are, again, years into dealing with these tools, and we know these risks exist. Yet no serious mitigations are provided. Not even any way of measuring the risk exposure.

And More

There are also many other risks associated with the technology. There are intellectual property risks with the foundation models, due to recklessness with their training data. There are existential financial risks associated with the infrastructure build-out. The extent to which most “open” models are simply derivatives of frontier models is an open question.

What I Think

If any one of these things were regularly overlooked by AI vendors or users, that would be a totally normal product oversight. Room for improvement for the next version, but nothing catastrophic.

Shipping without any of them doesn’t seem like lean product management, it seems like a careless attitude towards risk and a product design philosophy oriented entirely towards short-term demos, with no regard for how to realize actual productivity gains.

Furthermore, being available for years without anything like these features, despite hundreds of incidents demonstrating the risks, with hundreds of billions of dollars of funding, makes it seem to me like if they were to add all the features that would make their product actually safe and hypothetically useful, these features would reveal that it is actually not an improvement to productivity.

In the year since I first wrote about measuring the cost/benefit ratio of AI, I have heard from numerous people who have shown this to management to try to illustrate why their AI initiatives — like almost all AI initiatives — were either failing or burning out their engineers.

I’ve also heard from lots of people that have told me that it’s obviously useful and they don’t need to measure so carefully, because they are getting lots of work done that they couldn’t have otherwise.⁷

I have yet to hear from a single person who has said “yeah, we measured according to your methodology⁸, and it turns out that our AI work is going great and that our ratio is 0.75”.

Obviously, I cannot say for sure why this is; absence of evidence is not evidence of absence. But at this point I think the null hypothesis is that AI tools provide, in aggregate, zero value. They make mistakes too often, and the externalities they produce are so bad and so difficult to control that even before we get to the places where they are just physically poisoning people, even the negative effects on their direct users end up cancelling out whatever benefit to they provide to their organizations.

If I were wrong, then including tools to measure an AI’s effectiveness at the tasks their users are actually trying to accomplish, rather than meaningless benchmarks, would show big productivity gains. The frontier labs would be champing at the bit to add such features, and crowing about their fantastic results.

I think the labs know that if they did that, it would present a grim picture to their users. Such tools would let their users see that it’s making mistakes much more often than they realized, that they’re spending much more time with it than they want to be, and that it’s just generally not fit for purpose.

If they prove me wrong by adding in all of these safety mechanisms, and in the process, they make all of their AI technology less harmful, I’ll be thrilled to be debunked.

Acknowledgments

Thank you to my patrons who are supporting my writing on this blog. If you like what you’ve read here and you’d like to read more of it, or you’d like to support my various open-source endeavors, you can support my work as a sponsor!

I get asked this rhetorical question a lot, in various forms:

Sure, datacenters might use a lot of energy, but you don’t have to use a hosted frontier model to do software development. What if I just run a local open-weights model to do some coding, with an open-source coding agent? Video games also use my GPU. Is local model development any worse than playing a video game?

So I want to write down my comprehensive answer to this: Yes, using an LLM to write some code is worse than playing a video game, for a few reasons.

Video Games Are Interactive, LLMs Are Batch Jobs

Video games use compute to respond to human input. You are using your GPU while you are looking at a screen, displaying an image. When you are done playing, you shut off the game, and your computer goes back to idle. It’s much less energy. By contrast, agentic loops with evals (the only kind of “AI” that is meaningfully any good at coding) are running hot, for days. To use the most recent example of such a thing, a very rough first sketch of an implementation of a Windows graphics API backend to help port a paint program to other platforms, it took 3 weeks of Claude time, “day and night”. Do you play a lot of video games for 500 hours to make it past the tutorial level, while also using other computers for other things, as well as the rest of your carbon footprint?

Video Games Need Development, LLMs Need Training

Video games use compute to respond to human input during development, too. Your game has to be made, but your LLM has to be trained. LLMs use a historically extreme amount of power, probably using more than the entire Internet, but it’s kind of hard to say. Still, it seems a reasonable estimate to within several orders of magnitude that even over a multi-year project with hundreds of developers, the power used to develop an individual video game is nowhere close to training even a small LLM.

This is true even for local models. OpenAI has openly claimed that DeepSeek “stole its intellectual property”, and I have heard grumblings that none of the open-weights generalist models could realistically exist without the massive lift that the frontier labs are doing with their training, in various other ways too. Secrecy throughout the industry makes this kind of impossible to understand rigorously, but it seems fair to say that you are partially culpable for all that famously energy-intensive frontier lab training if you’re using a local model.

And They Keep Needing Training

You also can’t dismiss this as a sunk cost, because in order to stay current with industry developments, models need to be updated with new information from the rest of the world, which means that you need to keep training them. Beyond the energy for your own use, if you want a real-life agentic workflow that actually does useful stuff, practically speaking you would still need to update your local models over and over again, at least once every few months, which means you would be incentivizing continued energy consumption by whoever was doing that training for you, including the energy cost of scraping.

Let’s Be Real Here, You Aren’t Actually Using A Local Model

This question is a hypothetical thought experiment. Despite synthetic benchmarks that keep showing there isn’t much difference between open weight and frontier models, nobody’s actually using local models for much of anything beyond sharing those talking points. Depending on which benchmark you’re looking at, maybe it’s good enough or maybe it’s worse.

As an inveterate AI hater, all these systems seem pretty bad to me, but it seems that people who find them useful tend to subjectively believe the frontier models are worth the premium, and that’s what they’re actually using. Once you have accepted that it is OK to use LLMs for coding at all, it seems like a very quick slippery slope on down to “we’ll go ahead and use the frontier models for now anyway, but we could be ethically better in the future by switching to an open weights one, that option is always available”.

There’s A Reason We Have Data Centers

Devolving power usage to local LLMs might be good to make users responsible for their costs and decrease the impacts to communities that are physically next to huge concentrations of power utilization, not to mention generation. However, there’s a reason that it makes sense for the providers to build these giant facilities: economies of scale reduce total power consumption, they don’t increase it. If you do all the same stuff with a local model that they have to do in hosted environments, it will probably take more power, even though you will be incentivized to do different stuff. This incentive to “do different stuff” is why although local models can hypothetically hold their own against the frontier labs for some tasks, when people or businesses take their inference costs in-house they often find that it’s too painful and move back to hosted LLMs.

There Are Problems Other Than Power

These are subjects for a different post, but you have to consider a lot of other externalities: AI psychosis, de-skilling, comprehension debt, cultivating a dependency, introducing security defects, limiting your design space based on what LLMs can understand, context rot, wasting time on invalid solutions, introducing unpredictability into your workflows. You still have to consider the total cost benefit ratio.

To Sum Up

Local LLMs might alleviate some of the harms from using the hosted frontier providers. There are fewer privacy concerns, you can measure your power utilization and be more directly responsible for it, you can build interfaces with affordances that are less oriented towards addiction and dependency than the major frontier labs’ harnesses.

But they’re not automatically “the same as playing a video game” just because they can use the same GPU.

Acknowledgments

Thank you to my patrons who are supporting my writing on this blog. If you like what you’ve read here and you’d like to read more of it, or you’d like to support my various open-source endeavors, you can support my work as a sponsor!

As I have discussed in previous posts, “AIs” can make mistakes. In fact, they do make mistakes, and their mistake-making patterns are such that where and how they will make mistakes is both uncertain and constantly changing.

Thus, in any scenario where you want to attempt to make “productive” use of “AI”, you must have a system in place for checking every result. Not checking some results; checking every result. If each result might have a consequence for you (and if it didn’t have a consequence, why bother automating it?) and you cannot predict in advance which kinds of results will need verification, then verification is always required.

The verification often ends up being just as expensive as doing the work in the first place, which means that if you want your usage of “AI” to be personally profitable, you have to find someone else to externalize the cost of verification onto. This person becomes your adversary, and, if you are successful, your “AI’s” victim.

The Ladder-Climber And Their Reverse-Centaur Rungs

One way that this constellation of facts can straightforwardly assemble themselves into a dystopian nightmare is the phenomenon, described by Cory Doctorow, of the reverse centaur. This is when your employer non-consensually turns you into the verification system. The “AI” does the fun part of initially performing the work, and then you do the boring part where you check if the robot is right and clean up its messes, even if everyone already knows that it would, in aggregate, be cheaper for you to do the work in the first place.

Reverse centaurs can be made from any automation, not only “AI” automation. I think that there is a reason that this term happens to have emerged in the “age of AI”, though, and not with earlier automation technologies (even those which were considerably more viscerally horrific). That reason is: the wrongness of “AI” output is not merely a technical feature that must be compensated for, it is a generalized externality.

As I mentioned above, if you are responsible for the entirety of the work, both extruding the “AI” output and checking it, it’s usually cheaper to have humans do the entirety of the work to begin with. When humans do the writing directly, we can check as we go, and thus verification doesn’t need to be as comprehensive.

When “AI” coding advocates say “code review is the bottleneck”, what they are observing is that the LLM is still rolling the dice for each PR, and a human is still necessary to verify that each of those rolls is a winner. But calling this process “code review” is a bit of a misnomer; it’s not really “code review” in the traditional sense, it’s human understanding.

Before the advent of “AI”, the human understanding was implicit in the process of writing the code in the first place¹, and the code review was a way of diffusing and extending that understanding. Now that the code can be authored with no initial understanding taking place, that cost has not gone away, it has moved.

Human understanding was always the bottleneck.

However, this is taking a collaborative view of a software project, where satisfying the needs and solving the problems of your customers are the goals. We can see that “AI” is a bad tool to satisfy those goals, because all it’s doing is converting the first half of the work, that of understanding the code as you write it, to understanding the agent’s output as you read it.

What if, instead, we were to take the view that every software company is a Hobbesian nightmare, red in tooth and claw? In this view, the only goal of a software project is for the individual developers to make their promo cycles and get their bonuses. Given that there is only a certain amount of money to go around, this is a zero-sum game where each programmer wants to look more productive than their colleagues.

Pretty much every organization finds it easy to reward “productivity” as expressed by lines of code emitted, but the benefits of doing thorough and thoughtful design, analysis, and code review very difficult to reward. In this world, an LLM is an invaluable tool for the sociopathic ladder-climber, particularly if your legacy organization is still structuring their workflows as if the person prompting the bot is “writing” the code, and then they get to foist off the act of “reviewing” the code onto someone else.

Here, the prompter effectively externalizes the cost of the LLM’s failures but internalizes any benefits. The prompter will vibe-code a big feature, so large that the assigned reviewer can’t possibly comprehend it all effectively. When this happens, the reviewer will, eventually, be pressured to approve it, even if they can try to spot a few problems along the way. The reviewer has their own work to get back to, after all, the obligation to review the prompter’s (read: the bot’s) code is a drain on their time that they are not going to get rewarded for.

If this feature is a big success, the prompter gets a promotion. If it causes a big issue, well, the reviewer must not have been careful enough.

This is why LLMs are “good for coding”, and also why their biggest promoters keep having outages.

The Generative Gish Galloper

Coding is the biggest “success story” of this type of adversarial communication, but it is by far not the only instance of such a thing. LLMs create a new form of leverage that can turn Brandolini’s law from a linear advantage into an exponential one. If you are engaged in a political debate where you want to overwhelm the other side in nonsense, an LLM can generate bullshit faster than it is physically possible for a human being to type, let alone respond thoughtfully. There is an asymmetry to the utility of this weapon as well: only one side of the political spectrum wants to flood the zone and destroy trust in institutions and the concept of truth. There’s a good reason that the fascists love it.

Straightforward Spam and Fraud

This is kind of obvious, but LLMs can generate lightly-customized, plausible-looking text much more quickly than any human being. This facilitates their use in fraud, spam, and scams. In a spamming or fraudulent interaction, once again, the costs are externalized onto the victim: the recipient of a spam message has to do all the work of “checking” the LLM’s output. Spammers already expect very low hit rates from boilerplate, and if the LLM can increase those percentages from 1% to 5% the technology will pay for itself; they don’t need anything like reliable accuracy.

Customer “Support”

If you have any kind of commercial relationship with a company, I probably don’t even need to mention this: customer “support” bots are a misery. Everybody knows it at this point. But customer support is usually conceptualized by businesses as an adversarial interaction, because it is a cost center. They maintain internal metrics on time-to-resolution and try to optimize them. Implicitly, this creates a dynamic where the goal of the customer service agent’s job is not to solve your problem, but to emit noise that will cause you to think your problem is resolved, or to give up, as fast as possible. Unsurprisingly, LLMs can emit this noise faster than humans can, getting those customers off the phone. But those customers will remember those interactions, and the story outside the TTR metrics is horrible.

Similarly to the situation in software development, LLMs can look very good on paper for customer support, but mostly what they are doing is illuminating the problems with the industry’s existing metrics, by turning “winning the metrics battle against the customer” into a more obvious and immediate defeat for the company’s long term reputation.

“Education”

In 2026 it is sadly a fact of life that students cheat all the time using “AI”, and that this cheating is very successful, in that the teachers find it very hard to detect.

LLMs are great for cheating on schoolwork because the student is externalizing the work of the checking onto the teachers, who are often starting at a disadvantage to begin with, at least in the US.

My view is that this is happening because of a divergence in the way that students vs. teachers (or, more accurately, “the broader educational system”) view grading.

When a student is asked to write an essay, the teachers see the effort as both intrinsically worthwhile for the student, as well as useful as a pedagogical tool to evaluate and react to the student’s progress. The student, by contrast, sees a stumbling block designed to knock them off the path to success and into a permanent underclass. It is no wonder that the student sees “AI” as useful to their own goals and has no compunction about deploying it.

There is a bitter irony that the ability to understand the inherent value of actually writing the essay on their own is the sort of thing that students can really only learn by writing a bunch of essays. There’s no way that I can think of which makes the benefit legible as long as a shortcut is available.

The net effect here is a downward spiral, where the already-wobbling educational system is sustaining an attack that it doesn’t have the resources to recover from. The individual students’ attacks against their teachers and their schools’ grading systems might appear to momentarily succeed, but they will win the battle and lose the war.

Spamming “For Good”?

Usually when we talk about someone unilaterally choosing to enter into an adversarial relationship, that’s an “attack” and for good reasons we have a negative impression of the attacker. However, I would be remiss if I did not point out that there are some cases where the relationship was already adversarial; just because you’re the attacker doesn’t mean that you are evil.

For example we might imagine use-cases like automatically filing appeals for prior authorizations against health insurance. It’s relatively well-known at this point that the main way for-profit insurers maintain their margins is by denying claims right up to the line of the policies themselves being fraud, so using a spamming tool to fight them might be entirely justifiable² in that case.

Similarly, using an LLM could be justified in a fight against a company refusing to honor a warranty. One could imagine using an LLM to immediately generate replies and escalations.

However, even in imagined cases like these, the underlying problem is that the insurers and the vendors already have a tremendous amount of structural power, so it is more likely that they will have the advantage in deploying a communications weapon like an LLM, as well as enacting policies to simply ignore any LLM-based communication that you might submit. Worse, if these strategies were to become widespread, they might provide an excuse to reject any communications by feeding them into an unreliable “LLM detector” and issuing an automated “computer says no” even to hand-written correspondence.

It is also worth stressing that these cases are imagined, as compared to the very real coworker-abuse, spam, scam, fraud, and disinformation campaigns being waged in real life today.

Therefore, while legitimate uses might exist, it’s hard to imagine that there’s anywhere they would be genuinely valuable and sustainable. In the best case “AI” will provide a temporary advantage for underdogs that will provoke an arms race which the resource-advantaged adversaries will win in the long run, in the worst case the arms race itself will cement permanent structural change that will make things worse.

“Search” By Stealing

Most of the adversarial utility of “AI” is on the “write” side, since write-amplification is more obviously aggressive than reading. But the “read” side of LLMs — summarization and question-answering — can be a form of attack as well.

To begin with, the act of reading itself is currently enormously destructive, but that’s arguably not a fundamental aspect of this technology. They could set reasonable rate-limits and respect things like robots.txt, as search engines have for decades now. They could also refrain from committing criminal levels of copyright infringement. But, today, using “AI” tools does suborn this sort of out-of-control crawling.

More insidiously, consider the scenario described in this YouTube video. The LTT Bros decided to try Linux again, and in the course of so doing, they had problems. When trying to solve these problems, they were faced with a choice: they could consult Reddit, or they could ask an LLM. Asking an LLM would “gaslight the heck out of” them, but they still found it preferable, because they would at least get an answer without getting yelled at.

Initially this sounds great. But it also means that you want to extract knowledge from a community, while mechanically eliding any values or norms that the community may want to impart as part of offering that knowledge. As someone who spent many years in a community tech support role, this is worrying. Many requests for support are people asking how to do things that will momentarily solve a superficial problem but create a long-term reliability problem or even an immediate security risk, that the question-asker doesn’t want to hear about. Consider the question “I’m tired of entering my password so much, how do I make it so my laptop unlocks automatically”. An obsequious chatbot will helpfully tell you how to do this without pushback.

But, this is also a sort of ethically murky area. The Linux community is somewhat famously, for many years now, a toxic cesspool of general hostility, misogyny, etc. It is certainly a good thing that people can get access to this knowledge without subjecting themselves to abuse. But it also means that the people with the power and the privilege to change the community for the better can just quietly withdraw, rather than fixing the problems. It also means that the positive elements of culture cannot be transmitted, and people will have no opportunity to learn about unknown unknowns.

In this case, the “adversarial” communication is with society. The thing that using an LLM for search lets you do is withdraw from society and avoid forming any personal connections. There are some personal connections which are painful and annoying, and so that can feel like a momentary balm. But the need to make connections in general is, like, the concept of society itself.

Who Am I Hurting?

LLMs are good at adversarial communication. They are so good at it, relative to their other benefits, that they will tend to make communications adversarial if you are not remaining vigilant about the possibility that it might do so. My request to you, dear reader, if you are going to use such tools, is to always ask yourself, “who might I be hurting, if I use an LLM for this?”

If you’re using an “AI”, who is its adversary? If you haven’t given it one yet, who might the “AI” turn into an adversary? Who might you overwhelm with an asymmetric amount of output, or, if you’re receiving information and not sending it, who are you taking that information from without consulting?

Figure out the answers to these questions and conduct yourself accordingly; the answer might be “yourself”.

Acknowledgments

Thank you to my patrons who are supporting my writing on this blog. If you like what you’ve read here and you’d like to read more of it, or you’d like to support my various open-source endeavors, you can support my work as a sponsor!

Humans Are Bad At Perceiving

Humans are not particularly good at catching bugs. For one thing, we get tired easily. There is some science on this, indicating that humans can’t even maintain enough concentration to review more than about 400 lines of code at a time..

We have existing terms of art, in various fields, for the ways in which the human perceptual system fails to register stimuli. Perception fails when humans are distracted, tired, overloaded, or merely improperly engaged.

Each of these has implications for the fundamental limitations of code review as an engineering practice:

• Inattentional Blindness: you won’t be able to reliably find bugs that you’re not looking for.

• Repetition Blindness: you won’t be able to reliably find bugs that you are looking for, if they keep occurring.

• Vigilance Fatigue: you won’t be able to reliably find either kind of bugs, if you have to keep being alert to the presence of bugs all the time.

• and, of course, the distinct but related Alert Fatigue: you won’t even be able to reliably evaluate reports of possible bugs, if there are too many false positives.

Never Send A Human To Do A Machine’s Job

When you need to catch a category of error in your code reliably, you will need a deterministic tool to evaluate — and, thanks to our old friend “alert fatigue” above — ideally, to also remedy that type of error. These tools will relieve the need for a human to make the same repetitive checks over and over. None of them are perfect, but:

• to catch logical errors, use automated tests.
• to catch formatting errors, use autoformatters.
• to catch common mistakes, use linters.
• to catch common security problems, use a security scanner.

Don’t blame reviewers for missing these things.

Code review should not be how you catch bugs.

What Is Code Review For, Then?

Code review is for three things.

First, code review is for catching process failures. If a reviewer has noticed a few bugs of the same type in code review, that’s a sign that that type of bug is probably getting through review more often than it’s getting caught. Which means it’s time to figure out a way to deploy a tool or a test into CI that will reliably prevent that class of error, without requiring reviewers to be vigilant to it any more.

Second — and this is actually its more important purpose — code review is a tool for acculturation. Even if you already have good tools, good processes, and good documentation, new members of the team won’t necessarily know about those things. Code review is an opportunity for older members of the team to introduce newer ones to existing tools, patterns, or areas of responsibility. If you’re building an observer pattern, you might not realize that the codebase you’re working in already has an existing idiom for doing that, so you wouldn’t even think to search for it, but someone else who has worked more with the code might know about it and help you avoid repetition.

You will notice that I carefully avoided saying “junior” or “senior” in that paragraph. Sometimes the newer team member is actually more senior. But also, the acculturation goes both ways. This is the third thing that code review is for: disrupting your team’s culture and avoiding stagnation. If you have new talent, a fresh perspective can also be an extremely valuable tool for building a healthy culture. If you’re new to a team and trying to build something with an observer pattern, and this codebase has no tools for that, but your last job did, and it used one from an open source library, that is a good thing to point out in a review as well. It’s an opportunity to spot areas for improvement to culture, as much as it is to spot areas for improvement to process.

Thus, code review should be as hierarchically flat as possible. If the goal of code review were to spot bugs, it would make sense to reserve the ability to review code to only the most senior, detail-oriented, rigorous engineers in the organization. But most teams already know that that’s a recipe for brittleness, stagnation and bottlenecks. Thus, even though we know that not everyone on the team will be equally good at spotting bugs, it is very common in most teams to allow anyone past some fairly low minimum seniority bar to do reviews, often as low as “everyone on the team who has finished onboarding”.

Oops, Surprise, This Post Is Actually About LLMs Again

Sigh. I’m as disappointed as you are, but there are no two ways about it: LLM code generators are everywhere now, and we need to talk about how to deal with them. Thus, an important corollary of this understanding that code review is a social activity, is that LLMs are not social actors, thus you cannot rely on code review to inspect their output.

My own personal preference would be to eschew their use entirely, but in the spirit of harm reduction, if you’re going to use LLMs to generate code, you need to remember the ways in which LLMs are not like human beings.

When you relate to a human colleague, you will expect that:

1. you can make decisions about what to focus on based on their level of experience and areas of expertise to know what problems to focus on; from a late-career colleague you might be looking for bad habits held over from legacy programming languages; from an earlier-career colleague you might be focused more on logical test-coverage gaps,
2. and, they will learn from repeated interactions so that you can gradually focus less on a specific type of problem once you have seen that they’ve learned how to address it,

With an LLM, by contrast, while errors can certainly be biased a bit by the prompt from the engineer and pre-prompts that might exist in the repository, the types of errors that the LLM will make are somewhat more uniformly distributed across the experience range.

You will still find supposedly extremely sophisticated LLMs making extremely common mistakes, specifically because they are common, and thus appear frequently in the training data.

The LLM also can’t really learn. An intuitive response to this problem is to simply continue adding more and more instructions to its pre-prompt, treating that text file as its “memory”, but that just doesn’t work, and probably never will. The problem — “context rot” is somewhat fundamental to the nature of the technology.

Thus, code-generators must be treated more adversarially than you would a human code review partner. When you notice it making errors, you always have to add tests to a mechanical, deterministic harness that will evaluates the code, because the LLM cannot meaningfully learn from its mistakes outside a very small context window in the way that a human would, so giving it feedback is unhelpful. Asking it to just generate the code again still requires you to review it all again, and as we have previously learned, you, a human, cannot review more than 400 lines at once.

To Sum Up

Code review is a social process, and you should treat it as such. When you’re reviewing code from humans, share knowledge and encouragement as much as you share bugs or unmet technical requirements.

If you must reviewing code from an LLM, strengthen your automated code-quality verification tooling and make sure that its agentic loop will fail on its own when those quality checks fail immediately next time. Do not fall into the trap of appealing to its feelings, knowledge, or experience, because it doesn’t have any of those things.

But for both humans and LLMs, do not fall into the trap of thinking that your code review process is catching your bugs. That’s not its job.

Acknowledgments

Thank you to my patrons who are supporting my writing on this blog. If you like what you’ve read here and you’d like to read more of it, or you’d like to support my various open-source endeavors, you can support my work as a sponsor!

The most optimistic vision of generative AI¹ is that it will relieve us of the tedious, repetitive elements of knowledge work so that we can get to work on the really interesting problems that such tedium stands in the way of. Even if you fully believe in this vision, it’s hard to deny that today, some tedium is associated with the process of using generative AI itself.

Generative AI also isn’t free, and so, as responsible consumers, we need to ask: is it worth it? What’s the ROI of genAI, and how can we tell? In this post, I’d like to explore a logical framework for evaluating genAI expenditures, to determine if your organization is getting its money’s worth.

Perpetually Proffering Permuted Prompts

I think most LLM users would agree with me that a typical workflow with an LLM rarely involves prompting it only one time and getting a perfectly useful answer that solves the whole problem.

Generative AI best practices, even from the most optimistic vendors all suggest that you should continuously evaluate everything. ChatGPT, which is really the only genAI product with significantly scaled adoption, still says at the bottom of every interaction:

ChatGPT can make mistakes. Check important info.

If we have to “check important info” on every interaction, it stands to reason that even if we think it’s useful, some of those checks will find an error. Again, if we think it’s useful, presumably the next thing to do is to perturb our prompt somehow, and issue it again, in the hopes that the next invocation will, by dint of either:

1. better luck this time with the stochastic aspect of the inference process,
2. enhanced application of our skill to engineer a better prompt based on the deficiencies of the current inference, or
3. better performance of the model by populating additional context in subsequent chained prompts.

Unfortunately, given the relative lack of reliable methods to re-generate the prompt and receive a better answer², checking the output and re-prompting the model can feel like just kinda futzing around with it. You try, you get a wrong answer, you try a few more times, eventually you get the right answer that you wanted in the first place. It’s a somewhat unsatisfying process, but if you get the right answer eventually, it does feel like progress, and you didn’t need to use up another human’s time.

In fact, the hottest buzzword of the last hype cycle is “agentic”. While I have my own feelings about this particular word³, its current practical definition is “a generative AI system which automates the process of re-prompting itself, by having a deterministic program evaluate its outputs for correctness”.

A better term for an “agentic” system would be a “self-futzing system”.

However, the ability to automate some level of checking and re-prompting does not mean that you can fully delegate tasks to an agentic tool, either. It is, plainly put, not safe. If you leave the AI on its own, you will get terrible results that will at best make for a funny story⁴⁵ and at worst might end up causing serious damage⁶⁷.

Taken together, this all means that for any consequential task that you want to accomplish with genAI, you need an expert human in the loop. The human must be capable of independently doing the job that the genAI system is being asked to accomplish.

When the genAI guesses correctly and produces usable output, some of the human’s time will be saved. When the genAI guesses wrong and produces hallucinatory gibberish or even “correct” output that nevertheless fails to account for some unstated but necessary property such as security or scale, some of the human’s time will be wasted evaluating it and re-trying it.

Income from Investment in Inference

Let’s evaluate an abstract, hypothetical genAI system that can automate some work for our organization. To avoid implicating any specific vendor, let’s call the system “Mallory”.

Is Mallory worth the money? How can we know?

Logically, there are only two outcomes that might result from using Mallory to do our work.

1. We prompt Mallory to do some work; we check its work, it is correct, and some time is saved.
2. We prompt Mallory to do some work; we check its work, it fails, and we futz around with the result; this time is wasted.

As a logical framework, this makes sense, but ROI is an arithmetical concept, not a logical one. So let’s translate this into some terms.

In order to evaluate Mallory, let’s define the Futzing Fraction, “$\mathrm{FF}$”, in terms of the following variables:

$H$

the average amount of time a Human worker would take to do a task, unaided by Mallory

$I$

the amount of time that Mallory takes to run one Inference⁸

$C$

the amount of time that a human has to spend Checking Mallory’s output for each inference

$P$

the Probability that Mallory will produce a correct inference for each prompt

$W$

the average amount of time that it takes for a human to Write one prompt for Mallory

$E$

since we are normalizing everything to time, rather than money, we do also have to account for the dollar of Mallory as as a product, so we will include the Equivalent amount of human time we could purchase for the marginal cost of one⁹ inference.

As in last week’s example of simple ROI arithmetic, we will put our costs in the numerator, and our benefits in the denominator.

The idea here is that for each prompt, the minimum amount of time-equivalent cost possible is $W+I+C+E$. The user must, at least once, write a prompt, wait for inference to run, then check the output; and, of course, pay any costs to Mallory’s vendor.

If the probability of a correct answer is $P=\frac{1}{3}$, then they will do this entire process 3 times¹⁰, so we put $P$ in the denominator. Finally, we divide everything by $H$, because we are trying to determine if we are actually saving any time or money, versus just letting our existing human, who has to be driving this process anyway, do the whole thing.

If the Futzing Fraction evaluates to a number greater than 1, as previously discussed, you are a bozo; you’re spending more time futzing with Mallory than getting value out of it.

Figuring out the Fraction is Frustrating

In order to even evaluate the value of the Futzing Fraction though, you have to have a sound method to even get a vague sense of all the terms.

If you are a business leader, a lot of this is relatively easy to measure. You vaguely know what $H$ is, because you know what your payroll costs, and similarly, you can figure out $E$ with some pretty trivial arithmetic based on Mallory’s pricing table. There are endless YouTube channels, spec sheets and benchmarks to give you $I$. $W$ is probably going to be so small compared to $H$ that it hardly merits consideration¹¹.

But, are you measuring $C$? If your employees are not checking the outputs of the AI, you’re on a path to catastrophe that no ROI calculation can capture, so it had better be greater than zero.

Are you measuring $P$? How often does the AI get it right on the first try?

Challenges to Computing Checking Costs

In the fraction defined above, the term $C$ is going to be large. Larger than you think.

Measuring $P$ and $C$ with a high degree of precision is probably going to be very hard; possibly unreasonably so, or too expensive¹² to bother with in practice. So you will undoubtedly need to work with estimates and proxy metrics. But you have to be aware that this is a problem domain where your normal method of estimating is going to be extremely vulnerable to inherent cognitive bias, and find ways to measure.

Margins, Money, and Metacognition

First let’s discuss cognitive and metacognitive bias.

My favorite cognitive bias is the availability heuristic and a close second is its cousin salience bias. Humans are empirically predisposed towards noticing and remembering things that are more striking, and to overestimate their frequency.

If you are estimating the variables above based on the vibe that you’re getting from the experience of using an LLM, you may be overestimating its utility.

Consider a slot machine.

If you put a dollar in to a slot machine, and you lose that dollar, this is an unremarkable event. Expected, even. It doesn’t seem interesting. You can repeat this over and over again, a thousand times, and each time it will seem equally unremarkable. If you do it a thousand times, you will probably get gradually more anxious as your sense of your dwindling bank account becomes slowly more salient, but losing one more dollar still seems unremarkable.

If you put a dollar in a slot machine and it gives you a thousand dollars, that will probably seem pretty cool. Interesting. Memorable. You might tell a story about this happening, but you definitely wouldn’t really remember any particular time you lost one dollar.

Luckily, when you arrive at a casino with slot machines, you probably know well enough to set a hard budget in the form of some amount of physical currency you will have available to you. The odds are against you, you’ll probably lose it all, but any responsible gambler will have an immediate, physical representation of their balance in front of them, so when they have lost it all, they can see that their hands are empty, and can try to resist the “just one more pull” temptation, after hitting that limit.

Now, consider Mallory.

If you put ten minutes into writing a prompt, and Mallory gives a completely off-the-rails, useless answer, and you lose ten minutes, well, that’s just what using a computer is like sometimes. Mallory malfunctioned, or hallucinated, but it does that sometimes, everybody knows that. You only wasted ten minutes. It’s fine. Not a big deal. Let’s try it a few more times. Just ten more minutes. It’ll probably work this time.

If you put ten minutes into writing a prompt, and it completes a task that would have otherwise taken you 4 hours, that feels amazing. Like the computer is magic! An absolute endorphin rush.

Very memorable. When it happens, it feels like $P=1$.

But... did you have a time budget before you started? Did you have a specified N such that “I will give up on Mallory as soon as I have spent N minutes attempting to solve this problem with it”? When the jackpot finally pays out that 4 hours, did you notice that you put 6 hours worth of 10-minute prompt coins into it in?

If you are attempting to use the same sort of heuristic intuition that probably works pretty well for other business leadership decisions, Mallory’s slot-machine chat-prompt user interface is practically designed to subvert those sensibilities. Most business activities do not have nearly such an emotionally variable, intermittent reward schedule. They’re not going to trick you with this sort of cognitive illusion.

Thus far we have been talking about cognitive bias, but there is a metacognitive bias at play too: while Dunning-Kruger, everybody’s favorite metacognitive bias does have some problems with it, the main underlying metacognitive bias is that we tend to believe our own thoughts and perceptions, and it requires active effort to distance ourselves from them, even if we know they might be wrong.

This means you must assume any intuitive estimate of $C$ is going to be biased low; similarly $P$ is going to be biased high. You will forget the time you spent checking, and you will underestimate the number of times you had to re-check.

To avoid this, you will need to decide on a Ulysses pact to provide some inputs to a calculation for these factors that you will not be able to able to fudge if they seem wrong to you.

Problematically Plausible Presentation

Another nasty little cognitive-bias landmine for you to watch out for is the authority bias, for two reasons:

1. People will tend to see Mallory as an unbiased, external authority, and thereby see it as more of an authority than a similarly-situated human¹³.
2. Being an LLM, Mallory will be overconfident in its answers¹⁴.

The nature of LLM training is also such that commonly co-occurring tokens in the training corpus produce higher likelihood of co-occurring in the output; they’re just going to be closer together in the vector-space of the weights; that’s, like, what training a model is, establishing those relationships.

If you’ve ever used an heuristic to informally evaluate someone’s credibility by listening for industry-specific shibboleths or ways of describing a particular issue, that skill is now useless. Having ingested every industry’s expert literature, commonly-occurring phrases will always be present in Mallory’s output. Mallory will usually sound like an expert, but then make mistakes at random.¹⁵.

While you might intuitively estimate $C$ by thinking “well, if I asked a person, how could I check that they were correct, and how long would that take?” that estimate will be extremely optimistic, because the heuristic techniques you would use to quickly evaluate incorrect information from other humans will fail with Mallory. You need to go all the way back to primary sources and actually fully verify the output every time, or you will likely fall into one of these traps.

Mallory Mangling Mentorship

So far, I’ve been describing the effect Mallory will have in the context of an individual attempting to get some work done. If we are considering organization-wide adoption of Mallory, however, we must also consider the impact on team dynamics. There are a number of possible potential side effects that one might consider when looking at, but here I will focus on just one that I have observed.

I have a cohort of friends in the software industry, most of whom are individual contributors. I’m a programmer who likes programming, so are most of my friends, and we are also (sigh), charitably, pretty solidly middle-aged at this point, so we tend to have a lot of experience.

As such, we are often the folks that the team — or, in my case, the community — goes to when less-experienced folks need answers.

On its own, this is actually pretty great. Answering questions from more junior folks is one of the best parts of a software development job. It’s an opportunity to be helpful, mostly just by knowing a thing we already knew. And it’s an opportunity to help someone else improve their own agency by giving them knowledge that they can use in the future.

However, generative AI throws a bit of a wrench into the mix.

Let’s imagine a scenario where we have 2 developers: Alice, a staff engineer who has a good understanding of the system being built, and Bob, a relatively junior engineer who is still onboarding.

The traditional interaction between Alice and Bob, when Bob has a question, goes like this:

1. Bob gets confused about something in the system being developed, because Bob’s understanding of the system is incorrect.
2. Bob formulates a question based on this confusion.
3. Bob asks Alice that question.
4. Alice knows the system, so she gives an answer which accurately reflects the state of the system to Bob.
5. Bob’s understanding of the system improves, and thus he will have fewer and better-informed questions going forward.

You can imagine how repeating this simple 5-step process will eventually transform Bob into a senior developer, and then he can start answering questions on his own. Making sufficient time for regularly iterating this loop is the heart of any good mentorship process.

Now, though, with Mallory in the mix, the process now has a new decision point, changing it from a linear sequence to a flow chart.

We begin the same way, with steps 1 and 2. Bob’s confused, Bob formulates a question, but then:

3. Bob asks Mallory that question.

Here, our path then diverges into a “happy” path, a “meh” path, and a “sad” path.

The “happy” path proceeds like so:

4. Mallory happens to formulate a correct answer.
5. Bob’s understanding of the system improves, and thus he will have fewer and better-informed questions going forward.

Great. Problem solved. We just saved some of Alice’s time. But as we learned earlier,

Mallory can make mistakes. When that happens, we will need to check important info. So let’s get checking:

4. Mallory happens to formulate an incorrect answer.
5. Bob investigates this answer.
6. Bob realizes that this answer is incorrect because it is inconsistent with some of his prior, correct knowledge of the system, or his investigation.
7. Bob asks Alice the same question; GOTO traditional interaction step 4.

On this path, Bob spent a while futzing around with Mallory, to no particular benefit. This wastes some of Bob’s time, but then again, Bob could have ended up on the happy path, so perhaps it was worth the risk; at least Bob wasn’t wasting any of Alice’s much more valuable time in the process.¹⁶

Notice that beginning at the start of step 4, we must begin allocating all of Bob’s time to $C$, so $C$ already starts getting a bit bigger than if it were just Bob checking Mallory’s output specifically on tasks that Bob is doing.

That brings us to the “sad” path.

4. Mallory happens to formulate an incorrect answer.
5. Bob investigates this answer.
6. Bob does not realize that this answer is incorrect because he is unable to recognize any inconsistencies with his existing, incomplete knowledge of the system.
7. Bob integrates Mallory’s incorrect information of the system into his mental model.
8. Bob proceeds to make a larger and larger mess of his work, based on an incorrect mental model.
9. Eventually, Bob asks Alice a new, worse question, based on this incorrect understanding.
10. Sadly we cannot return to the happy path at this point, because now Alice must unravel the complex series of confusing misunderstandings that Mallory has unfortunately conveyed to Bob at this point. In the really sad case, Bob actually doesn’t believe Alice for a while, because Mallory seems unbiased¹⁷, and Alice has to waste even more time convincing Bob before she can simply explain to him.

Now, we have wasted some of Bob’s time, and some of Alice’s time. Everything from step 5-10 is $C$, and as soon as Alice gets involved, we are now adding to $C$ at double real-time. If more team members are pulled in to the investigation, you are now multiplying $C$ by the number of investigators, potentially running at triple or quadruple real time.

But That’s Not All

Here I’ve presented a brief selection reasons why $C$ will be both large, and larger than you expect. To review:

1. Gambling-style mechanics of the user interface will interfere with your own self-monitoring and developing a good estimate.
2. You can’t use human heuristics for quickly spotting bad answers.
3. Wrong answers given to junior people who can’t evaluate them will waste more time from your more senior employees.

But this is a small selection of ways that Mallory’s output can cost you money and time. It’s harder to simplistically model second-order effects like this, but there’s also a broad range of possibilities for ways that, rather than simply checking and catching errors, an error slips through and starts doing damage. Or ways in which the output isn’t exactly wrong, but still sub-optimal in ways which can be difficult to notice in the short term.

For example, you might successfully vibe-code your way to launch a series of applications, successfully “checking” the output along the way, but then discover that the resulting code is unmaintainable garbage that prevents future feature delivery, and needs to be re-written¹⁸. But this kind of intellectual debt isn’t even specific to technical debt while coding; it can even affect such apparently genAI-amenable fields as LinkedIn content marketing¹⁹.

Problems with the Prediction of $P$

$C$ isn’t the only challenging term though. $P$, is just as, if not more important, and just as hard to measure.

LLM marketing materials love to phrase their accuracy in terms of a percentage. Accuracy claims for LLMs in general tend to hover around 70%²⁰. But these scores vary per field, and when you aggregate them across multiple topic areas, they start to trend down. This is exactly why “agentic” approaches for more immediately-verifiable LLM outputs (with checks like “did the code work”) got popular in the first place: you need to try more than once.

Independently measured claims about accuracy tend to be quite a bit lower²¹. The field of AI benchmarks is exploding, but it probably goes without saying that LLM vendors game those benchmarks²², because of course every incentive would encourage them to do that. Regardless of what their arbitrary scoring on some benchmark might say, all that matters to your business is whether it is accurate for the problems you are solving, for the way that you use it. Which is not necessarily going to correspond to any benchmark. You will need to measure it for yourself.

With that goal in mind, our formulation of $P$ must be a somewhat harsher standard than “accuracy”. It’s not merely “was the factual information contained in any generated output accurate”, but, “is the output good enough that some given real knowledge-work task is done and the human does not need to issue another prompt”?

Surprisingly Small Space for Slip-Ups

The problem with reporting these things as percentages at all, however, is that our actual definition for $P$ is $\frac{1}{\mathrm{attempts}}$, where $\mathrm{attempts}$ for any given attempt, at least, must be an integer greater than or equal to 1.

Taken in aggregate, if we succeed on the first prompt more often than not, we could end up with a $P>\frac{1}{2}$, but combined with the previous observation that you almost always have to prompt it more than once, the practical reality is that $P$ will start at 50% and go down from there.

If we plug in some numbers, trying to be as extremely optimistic as we can, and say that we have a uniform stream of tasks, every one of which can be addressed by Mallory, every one of which:

• we can measure perfectly, with no overhead
• would take a human 45 minutes
• takes Mallory only a single minute to generate a response
• Mallory will require only 1 re-prompt, so “good enough” half the time
• takes a human only 5 minutes to write a prompt for
• takes a human only 5 minutes to check the result of
• has a per-prompt cost of the equivalent of a single second of a human’s time

Thought experiments are a dicey basis for reasoning in the face of disagreements, so I have tried to formulate something here that is absolutely, comically, over-the-top stacked in favor of the AI optimist here.

Would that be a profitable? It sure seems like it, given that we are trading off 45 minutes of human time for 1 minute of Mallory-time and 10 minutes of human time. If we ask Python:

1
2
3
4
5

>>> def FF(H, I, C, P, W, E):
...     return (W + I + C + E) / (P * H)
... FF(H=45.0, I=1.0, C=5.0, P=1/2, W=5.0, E=0.01)
...
0.48933333333333334

We get a futzing fraction of about 0.4893. Not bad! Sounds like, at least under these conditions, it would indeed be cost-effective to deploy Mallory. But… realistically, do you reliably get useful, done-with-the-task quality output on the second prompt? Let’s bump up the denominator on $P$ just a little bit there, and see how we fare:

1
2

>>> FF(H=45.0, I=1.0, C=5.0, P=1/3, W=5.0, E=0.01)
0.734

Oof. Still cost-effective at 0.734, but not quite as good. Where do we cap out, exactly?

1
2
3
4
5
6
7
8
9

>>> from itertools import count
... for A in count(start=4):
...     print(A, result := FF(H=45.0, I=1.0, C=5.0, P=1 / A, W=5.0, E=1/60.))
...     if result > 1:
...         break
...
4 0.9792592592592594
5 1.224074074074074
>>>

With this little test, we can see that at our next iteration we are already at 0.9792, and by 5 tries per prompt, even in this absolute fever-dream of an over-optimistic scenario, with a futzing fraction of 1.2240, Mallory is now a net detriment to our bottom line.

Harm to the Humans

We are treating $H$ as functionally constant so far, an average around some hypothetical Gaussian distribution, but the distribution itself can also change over time.

Formally speaking, an increase to $H$ would be good for our fraction. Maybe it would even be a good thing; it could mean we’re taking on harder and harder tasks due to the superpowers that Mallory has given us.

But an observed increase to $H$ would probably not be good. An increase could also mean your humans are getting worse at solving problems, because using Mallory has atrophied their skills²³ and sabotaged learning opportunities²⁴²⁵. It could also go up because your senior, experienced people now hate their jobs²⁶.

For some more vulnerable folks, Mallory might just take a shortcut to all these complex interactions and drive them completely insane²⁷ directly. Employees experiencing an intense psychotic episode are famously less productive than those who are not.

This could all be very bad, if our futzing fraction eventually does head north of 1 and you need to reconsider introducing human-only workflows, without Mallory.

Abridging the Artificial Arithmetic (Alliteratively)

To reiterate, I have proposed this fraction:

which shows us positive ROI when FF is less than 1, and negative ROI when it is more than 1.

This model is heavily simplified. A comprehensive measurement program that tests the efficacy of any technology, let alone one as complex and rapidly changing as LLMs, is more complex than could be captured in a single blog post.

Real-world work might be insufficiently uniform to fit into a closed-form solution like this. Perhaps an iterated simulation with variables based on the range of values seem from your team’s metrics would give better results.

However, in this post, I want to illustrate that if you are going to try to evaluate an LLM-based tool, you need to at least include some representation of each of these terms somewhere. They are all fundamental to the way the technology works, and if you’re not measuring them somehow, then you are flying blind into the genAI storm.

I also hope to show that a lot of existing assumptions about how benefits might be demonstrated, for example with user surveys about general impressions, or by evaluating artificial benchmark scores, are deeply flawed.

Even making what I consider to be wildly, unrealistically optimistic assumptions about these measurements, I hope I’ve shown:

1. in the numerator, $C$ might be a lot higher than you expect,
2. in the denominator, $P$ might be a lot lower than you expect,
3. repeated use of an LLM might make $H$ go up, but despite the fact that it's in the denominator, that will ultimately be quite bad for your business.

Personally, I don’t have all that many concerns about $E$ and $I$. $E$ is still seeing significant loss-leader pricing, and $I$ might not be coming down as fast as vendors would like us to believe, if the other numbers work out I don’t think they make a huge difference. However, there might still be surprises lurking in there, and if you want to rationally evaluate the effectiveness of a model, you need to be able to measure them and incorporate them as well.

In particular, I really want to stress the importance of the influence of LLMs on your team dynamic, as that can cause massive, hidden increases to $C$. LLMs present opportunities for junior employees to generate an endless stream of chaff that will simultaneously:

• wreck your performance review process by making them look much more productive than they are,
• increase stress and load on senior employees who need to clean up unforeseen messes created by their LLM output,
• and ruin their own opportunities for career development by skipping over learning opportunities.

If you’ve already deployed LLM tooling without measuring these things and without updating your performance management processes to account for the strange distortions that these tools make possible, your Futzing Fraction may be much, much greater than 1, creating hidden costs and technical debt that your organization will not notice until a lot of damage has already been done.

If you got all the way here, particularly if you’re someone who is enthusiastic about these technologies, thank you for reading. I appreciate your attention and I am hopeful that if we can start paying attention to these details, perhaps we can all stop futzing around so much with this stuff and get back to doing real work.

Acknowledgments

Thank you to my patrons who are supporting my writing on this blog. If you like what you’ve read here and you’d like to read more of it, or you’d like to support my various open-source endeavors, you can support my work as a sponsor!