The Standard Drift: How AI Lowers Your Bar

In one week this month, two things landed on my screen that most people read as unrelated. Salesforce shipped a set of named agents that draft the sales email, answer the support ticket, write the code, and build the report, each one sitting on the company’s own data and rules. The same week, a Wharton study reported that heavy AI use at work erodes the judgment and reasoning that people build careers on. One story was framed as progress. The other as a warning. They are the same story.

Here is what I keep coming back to. When a machine can produce all of it, competently, on demand, the ability to produce stops being worth much. I have run companies for seventeen years and the thing I would have paid the most for, a person who could just make the thing, is now close to free. What is left, the only thing left, is the judgment to look at fifty plausible outputs and know which one is actually good. And that judgment is exactly the thing the abundance is quietly taking apart.

I do not think the risk is that AI hands you garbage. Garbage is easy. You see it, you throw it out, you move on. The risk is subtler and I have felt it in my own work: the machine hands you something that is good enough, and good enough slowly becomes your definition of good. Your bar does not break. It drifts. That is the thing I want to name and give you a way to fight.

When producing is free, the whole game moves to judging

Start with the price. In early 2023, running a good model cost real money per word. By 2026 that same generation is close to a rounding error, a drop of more than ninety-nine percent in a bit over two years. Producing a draft, a function, a design pass, a first version of almost anything, is now the cheapest step in the entire process. It used to be the expensive one.

When the cost of making something falls to zero, the value does not disappear. It moves. It moves to the one step the machine cannot do for you: deciding whether what came out is any good, and having the nerve to reject it when it is not. I wrote a while back that judgment became the new bottleneck once writing code and copy got cheap. This post is about the part I got wrong, or at least the part I left out. I treated judgment as a fixed asset you either have or you do not. It is not fixed. It moves too, and under AI it mostly moves down.

The stakes are not abstract. A 2025 study out of Stanford’s Social Media Lab and BetterUp Labs put a name to the thing everyone was already seeing: workslop. AI-generated output that looks like real work but does not carry the substance to move a task forward. In a survey of eleven hundred and fifty full-time US workers, forty-one percent said they had received workslop in the previous month. Each incident took an average of one hour and fifty-six minutes for someone else to sort out. The people who received it rated the colleague who sent it as less creative, less capable, and less trustworthy than before. That is a company quietly getting worse at its own work while the dashboards say output is up.

So the problem is not that people are producing less. They are producing more, faster, and a growing share of it is plausible mush that someone downstream has to catch. The catching is the job now. And whether you catch it depends entirely on one thing: whether your standard is still high enough to notice that the thing in front of you is not good enough.

The standard drift

Here is the one model I want you to hold from this whole piece. Call it the standard drift.

You have a bar. A sense of what excellent looks like in your domain, built over years of seeing great work, making bad work, and feeling the gap. That bar is not written down anywhere. It lives in your gut, and like anything in your gut, it recalibrates based on what you are exposed to most. For your whole career, what you were exposed to most was a mix: some great, a lot of mediocre, some genuinely bad, all made by humans with their own range. Your bar sat somewhere in the upper part of that range, and it stayed there because you kept seeing the high end and reaching for it.

Now change the input. Most of the work you look at all day is generated by a model whose output clusters tightly around the statistical middle of everything it was trained on. Not bad. Competent. Smooth. Confidently average. And you are marinating in it, hour after hour, draft after draft. Your bar does what any calibrated instrument does when you change what you feed it. It drifts toward the new average. Slowly enough that you never notice a single day it dropped, fast enough that a year later you are shipping things you would have sent back before.

Read the picture slowly. The dotted red line is the bar you believe you hold. Ask most founders how high their standards are and they will point at that line, honestly. The blue line is where the bar actually sits after months of exposure, sagging toward the orange line, which is the model’s confident middle. The grey dashed line is the market value of raw production, falling toward the floor. Two things are happening at once. The thing you are good at is worth less, and the thing that would save you, your standard, is sliding down to meet the machine instead of pulling the machine up.

The reason this is dangerous is that nothing in the drift feels like a decline. Every individual piece looks fine. You are shipping more. You feel faster. I have written before about why AI makes you feel faster than you are, and this is the quality-side cousin of that illusion. The output volume goes up and to the right, and underneath it the definition of good is quietly leaking out of the room.

Your sense of good is set by whatever you look at most

The clearest evidence for the drift is not a productivity study. It is the research on what AI does to the variety of what people make. A 2026 meta-analysis out of Tilburg and Aarhus pulled together nineteen studies on human-AI co-creation from 2022 through early 2026. It found a real, statistically significant homogenization effect. People working with generative models produce more similar ideas than people working without them. The effect was sharpest exactly where you would want originality most: concrete problem-solving, product ideas, real solutions.

The part that should worry any founder is the paradox buried in the same research. When you measure one person’s AI-assisted work on its own, it often scores as more creative than what they would have made alone. Individual gain. But line up everyone’s AI-assisted work side by side and it converges toward a shared, uniform middle. Collective loss. Each of you feels sharper. All of you are getting more alike. The model steers everyone toward the statistically common answer, because the statistically common answer is what it was built to produce.

Now connect that to your bar. Your standard is not a fact you looked up. It is an average of the work you have internalized. If most of the work you now internalize is the model’s confident middle, and that middle is the same middle everyone else is seeing, then your private sense of good is being rewritten toward a shared default without your consent. This is the reference reset, and it is the sharpest, least obvious idea in this whole piece: you are not just losing the skill of judging, the way muscles go when you stop using them. You are judging as hard as ever, against a reference that has quietly moved.

The menu is the trap. The model hands you five plausible options and the very act of choosing among them feels like judgment. It is not. It is selection inside a box someone else drew. Real judgment is knowing when to throw the whole menu out because none of it clears your bar, and that move only happens if your bar is still up there. When the bar has drifted, the best-of-menu looks great, because great is now defined by the menu. This is why building taste in the first place matters more than ever, and why the harder, less discussed problem is keeping the taste you already built from eroding.

Generating is free, but saying no got expensive

Suppose your bar has held. You look at the draft and you can feel that it is only good enough, not good. Now you have to act on that feeling, and this is where the second mechanism bites. Acting on your standard means rejecting work, and rejecting AI work is strangely hard in a way rejecting your own work never was.

When you wrote the draft yourself, throwing it out cost you the hours you spent. That sting kept you honest early, but it was a known cost. When the machine wrote the draft in nine seconds, throwing it out feels different. The thing is eighty-five percent there. It arrived instantly. Redoing it means either prompting again and hoping, or rolling up your sleeves and doing the last fifteen percent yourself, which is the hard, unglamorous part the machine skipped. In the moment, shipping the eighty-five percent is the path of least resistance, and the abundance makes that path feel reasonable. There is always more where that came from, so why agonize.

Multiply that decision across a team and you get the workslop economy. Each person ships the good-enough because saying no is friction and the volume is relentless. The cost does not vanish. It moves downstream to whoever has to catch it, at nearly two hours a piece, and it compounds into what the researchers called knowledge decay: the company’s shared record of what is true and what is good gets muddier with every plausible-but-hollow artifact added to it. Here is the line I want you to remember. Having taste is cheap now. Acting on it is what got expensive.

Look at the middle two rows. The two steps that got harder, deciding and rejecting, are the two steps that depend entirely on your standard. The machine made the easy step free and left you holding the hard steps with a bar that is quietly eroding. That is the squeeze. And notice this is a different problem from the one I wrote about in the reps problem, where judgment fades because you stop practicing it. Here you can be practicing constantly and still get worse, because you are practicing against a reference that has drifted. Disuse is one disease. Mis-calibration is another, and it hides better.

A standard that lives only in your head does not survive fifty agents

There is a third way the drift happens, and it is the one that turns a personal habit into a company problem. When you did the work yourself, your standard rode along automatically. It was in your hands. Every choice you made carried your taste whether you could articulate it or not. The moment you hand the work to a model, or to a team pointing models at everything, that automatic transmission breaks. The standard now has to travel from your head into a system, and taste does not travel well as an instruction.

You can write the style guide, the rubric, the prompt with seven examples. The output still misses, because the essence of a standard is knowing what is wrong that you never thought to specify. That is the part you cannot fully encode. So one of two things happens. Either you stay in the loop on everything, which does not scale and is the reason so many founders feel busier than ever, or you let go, and the work converges on the model’s default, which is the median we already talked about. A company running on agents without an owned, defended standard does not hold steady at your level. It regresses to the mean, at scale, quietly, in every function at once.

This is why I file this under the founder operating system and not under a writing tip. The standard is now a piece of infrastructure. Someone has to own it, the way someone owns the deploy pipeline or the books. If no one owns what good means, the machine gets to decide, and the machine’s answer is average. I have made the same argument about the recovery problem: reliability is not the same as being ready when things break. Quality is the twin of that. A high average output is not the same as a defended ceiling.

Where the drift shows up, function by function

This is not a writers’ problem or an engineers’ problem. It is the same drift wearing different clothes in every part of the company. Two examples I have watched closely. In engineering, a model produces code that runs and reads clean, and the reviewer, swimming in plausible diffs all day, approves the one that is fine instead of the one that is right, because fine now looks like the baseline. In hiring, a recruiter screens fifty AI-polished cover letters that all hit the same competent notes, and the candidate who would have stood out for a strange, specific, human reason gets averaged into the pile with everyone else.

Whether you win or lose in this environment comes down to two things that are easy to confuse. How much you can produce, and whether you can still tell good from good-enough. Most of the conversation about AI is stuck on the first. The second is where the whole game actually sits.

The bottom-right quadrant is where AI drops most people by default. It hands you the ability to produce at volume, and if your bar has drifted, you become a very efficient workslop factory, which is worse than being stuck, because you are now polluting everyone downstream. The top-right, the editor, is the only quadrant worth being in: you use the machine for volume and you hold a standard it cannot reach. The move from bottom-right to top-right is not a tooling upgrade. It is a decision to keep your bar up while everyone else’s drifts down. Here is the same pattern across five functions.

Read the right-hand column top to bottom. Every item is the thing that was going to be your edge, and every one is exactly what a median-seeking machine sands off. The drift does not just lower quality in the abstract. It specifically erases the parts of your work that were differentiated, because differentiated is, by definition, far from the average, and the average is where the model lives. Left unchecked, this is how a company full of smart people ends up shipping the same competent, forgettable thing as everyone else and calling it productivity.

The danger is not the output that is obviously bad

Most people, when they worry about AI quality, picture the model getting something wrong. Hallucinated facts, broken code, a confident lie. That failure mode is real, but it is not the one that will hurt you most, because it is loud. Wrong gets caught. Someone notices, it gets fixed, the system corrects. The failure mode that actually degrades a company is the opposite: output that is completely correct and completely mediocre. Right, and forgettable. It passes every check you have because it is not wrong about anything. It just is not good. And because it passes, it enters your work, your product, your knowledge base, and it adds one more data point to the average that is quietly resetting your bar.

This is the counterintuitive core of the whole problem. Obviously bad output is a gift, because it keeps your standard sharp by giving it something to reject. Plausible-mediocre output is the threat, because it gives your standard nothing to push against and slowly becomes the baseline. A related trap is trusting the output because it sounds authoritative, which I wrote about as automation bias, and its close cousin, the fluency trap, where smoothness gets mistaken for substance. Standard drift is the third member of that family, and the sneakiest, because it does not require you to trust anything. It just requires you to keep looking.

So the instinct to build more guardrails against wrong answers is half the job. The other half, the half almost nobody staffs, is a guardrail against right-but-mediocre. That guardrail is not a tool. It is a person with a standard, given the authority and the time to say this is fine and we are not shipping it.

The drift feeds itself, which is why it sneaks up on you

The reason standard drift is so easy to miss is that it runs as a loop, and each turn of the loop makes the next turn easier. Start with a bar that has slipped a little. Because it slipped, you approve a piece of median work you would once have sent back. That approved piece is now part of your product, your codebase, your published writing, the examples your team learns from. It becomes input. The next time you sit down to judge, that median piece is part of the reference you judge against, so your bar slips a little more, so you approve something a notch lower, and on it goes. Nobody decided to lower the standard. The standard lowered itself, one reasonable-looking approval at a time.

Two things make this loop tighter under AI than it ever was with human work. Volume, because you are now approving far more pieces per week, so the loop turns faster. And uniformity, because the median work all points in the same direction, so there is no scatter of high outliers in your daily input to pull your reference back up. In the old world, even mediocre teams produced the occasional spike of brilliance that reminded everyone what good looked like. A median-seeking machine rarely spikes. It regresses, reliably, toward the middle, so the corrective jolt that used to reset your eye mostly stops arriving.

That is why the defenses I care about are all external interrupts to the loop. A gold file, your own answer formed first, a person paid to reject: each one forces a high reference back into your input on purpose, because the environment has stopped supplying it on its own. You are not fighting a single bad decision. You are fighting a feedback loop that will keep running in the background of a busy company until someone deliberately breaks it. The good news is that breaking it is cheap. The bad news is that it will never happen by accident, and the loop is perfectly happy to run for years while every dashboard reports that output is up and to the right.

How to keep a reference of great in front of you

The drift is a calibration problem, so the fixes are all about calibration: keep pulling your reference back up toward great, deliberately, because the environment is pulling it down automatically. Here is what I actually do and what I tell the founders I work with.

First, keep a gold file. A small, living collection of genuinely excellent work in your domain, the writing, the code, the product flows, the decks that made you jealous. Before you judge anything the machine gives you, spend two minutes with the gold file. You are re-anchoring your eye on the high end before you look at the median. This one habit does more than any prompt trick, because it fights the reset at its source: what you looked at most.

Second, generate your own answer before you look at the model’s. Even a rough one, even just the shape of it. If you form your own view first, you have a reference to judge the output against, and you will notice what is missing. If you look at the model first, its answer becomes your starting point and you will only ever edit around its frame. This costs a few minutes and it is the difference between using the machine and being used by it.

Third, give rejection a budget and a name. On my teams, someone owns the quality floor for each surface that matters, and it is understood that killing good-enough work is part of the job, not an interruption of it. Rejecting a plausible draft should feel as normal as merging a good one. If no one is allowed to be the person who says not good enough, the median wins by default.

Fourth, do some work with no AI at all, regularly. Not out of nostalgia. It is a recalibration. When you make something end to end yourself, you feel the full range again, the struggle, the bad version, the moment it gets good, and that resets your gut to a human standard instead of a machine one. I treat it like a pilot keeping up manual flying hours. You are not being inefficient. You are maintaining the instrument you use to judge everything else. This is close to the point I made in the enough problem, but pointed the other way: there the bar for how much you produce keeps climbing; here the bar for how good it has to be keeps sliding, and both need a deliberate hand on them.

When everyone drifts to the mean, holding a standard becomes a moat

There is an upside hiding inside all of this, and it is the reason I am not gloomy about it. If AI pulls the median of all work toward a shared middle, and it does, then the small number of people and companies who refuse to drift become sharply differentiated by default. When everyone’s output converges, distinctiveness gets scarce, and scarce is valuable. The standard you defend stops being a personal virtue and becomes a competitive position.

Think about what a moat actually is. It is something your competitors cannot easily copy. In a world where anyone can copy your feature in a weekend with a model, the thing they cannot copy is the accumulated judgment about what to build and what to reject. Taste was always a soft edge. It is hardening into one of the few durable ones, precisely because the tools that commoditized everything else cannot commoditize it. I have argued that initiative is a skill AI quietly erodes, and the same logic runs here. The capabilities the machine cannot supply are the ones appreciating in value while everyone stares at the ones it made cheap.

So the practical stance is not to resist AI or to slow down. Produce at full volume. Use every agent you can. But treat your standard as the one asset you protect at all costs, because when the flood recedes, the companies still standing will be the ones that kept a bar the machine could not reach. The others will have shipped a great deal of forgettable, correct, average work, very fast, and wondered why none of it moved anything.

Questions founders ask about AI and quality standards

Is standard drift just a fancy way of saying AI makes people lazy?

No, and that framing lets you off too easy. Laziness is a choice you can see and correct. Standard drift happens to people working hard, paying attention, and genuinely trying. The mechanism is calibration, not effort. Your sense of good recalibrates to what you are exposed to, and you are now exposed to an ocean of competent-average machine output. You can be the most diligent person on the team and still ship worse work than you did two years ago, because your reference for good moved without your permission.

How do I know if my own bar has already drifted?

Go back to something you made three or four years ago that you were proud of, and something you approved last month. Compare them honestly against each other, not against the deadline. If the old work feels sharper, more specific, more like a human made a real choice, your bar has likely drifted and the recent work is riding on good-enough. Another tell: how often do you reject AI output outright versus lightly edit it and ship? If it is almost always the latter, the machine is setting your standard, not you.

Does this mean I should use AI less?

Use it as much as you want for producing. The drift is not caused by using AI to generate. It is caused by letting the generated median become your reference for quality. So the fix is not less AI, it is more deliberate calibration: keep great work in front of you, form your own view before you look at the output, and protect the act of rejection. You can run a hundred agents and still hold a high bar, as long as someone is actually holding it.

Is this the same as worrying about AI hallucinations?

It is almost the opposite, and that is what makes it dangerous. Hallucinations are wrong, and wrong is loud, so it gets caught and fixed. Standard drift is about output that is completely correct and quietly mediocre. It passes every accuracy check because it is not wrong about anything, and it slips into your work and your baseline unnoticed. Guarding against wrong answers is necessary but it does nothing about right-but-forgettable, which is the bigger long-term threat to a company that wants to be distinctive.

How do I set a standard for a team using AI without becoming the bottleneck on everything?

You cannot fully encode taste into a rubric, so do not try to review everything yourself, which is the trap. Instead, name an owner of the quality floor for each surface that matters, give a few concrete examples of the bar cleared and the bar missed, and make rejecting good-enough work an explicit, expected part of the role rather than an interruption. The goal is to distribute the standard, not to be the single point it passes through. A standard that only lives in the founder’s head does not survive a team pointing models at everything.

Is a gold file really enough to fight this?

On its own, no, but it is the single habit that does the most because it attacks the drift at its source. The drift comes from what you look at most, so deliberately looking at excellent work right before you judge re-anchors your eye at the top of the range instead of the middle. Pair it with generating your own answer first and protecting rejection, and you have a real defense. The gold file is the cheapest part and does the most, which is why I start there.

Will better models fix this on their own?

Better models make the output more fluent and more confidently average, which if anything strengthens the drift, because the median gets more persuasive and harder to reject. A smoother mediocre is more dangerous than a rough one, not less. The models are optimized to produce the likely answer, and the likely answer is the middle of the distribution by construction. No amount of capability changes that shape. The defense has to come from you keeping a reference the model is not built to reach.

Where does this fit with everything else AI is doing to how founders work?

I think of it as one piece of a larger shift in the founder operating system. Producing got cheap, so judgment became the bottleneck, and now the judgment itself is under pressure from three directions at once: it atrophies from disuse, it gets fooled by fluency, and it drifts from over-exposure to the median. Standard drift is that third pressure. If you are building a company on AI, protecting the quality of your own judgment is not a soft skill on the side. It is core infrastructure, as load-bearing as your code or your cash.

What I would keep, if I kept one thing

If you take nothing else from this, take the reframe. The threat was never the output that is obviously bad, because bad gets caught. It is the output that is good enough, because good enough is what quietly resets your definition of good. Producing is free now. Guarding the bar is the whole job. Keep great work in front of your eyes, form your own view before the machine forms it for you, and give someone the authority to say not good enough and mean it. Do that and the drift that is dragging everyone else toward the middle becomes the exact thing that sets you apart.

I am a four-time founder building AI-native companies. I write about the founder mindset, AI opportunities, and the mechanics of building with AI at vikasmalpani.com.