AI Is a Machine for the Average. The Exception Is Your Job.

You asked AI something this week, and it gave you an answer. The answer was reasonable. You read it, it matched roughly what you were thinking, and you moved on.

Nothing went wrong. That is the part nobody examines.

We have trained ourselves to watch AI for the dramatic failures. The invented citation. The confident wrong number. The fabricated quote. Those are real, and they are easy to catch, because they announce themselves as errors. But the failure that actually costs you something never announces itself, because it does not look like a failure at all. It looks like a perfectly reasonable answer. It looks like the work getting done.

The machine is not pulling you toward wrong. It is pulling you toward the middle. And the middle is far more dangerous than wrong, because wrong gets caught and the middle gets accepted.

---

What "reasonable" actually is

Start with the word, because the whole trap lives inside it.

When an answer feels reasonable, what is the feeling made of? It is not a feeling of correctness — you have not verified anything. It is a feeling of recognition. The answer matches the shape of what you already expected, what you have read before, what most people in the field would say. It sits comfortably inside the consensus. That comfort is what we are calling reasonable.

Now look at what that means. The consensus, the familiar, the thing most texts say, the answer nobody would object to — that is the statistical centre of human writing on the subject. The mean. So when an answer feels reasonable, what you are actually feeling is its proximity to the average.

This is why generic AI output is so hard to reject. It does not feel like slop while you are reading it. It feels like sense. It only feels hollow later, if at all, once you have already built on it and noticed it carried you nowhere. The centre always feels right in the moment, because feeling right is what the centre is.

The summary layer. The balanced overview. The "it depends on your context" close. The answer that presents both sides and commits to neither. Fluent, well-formed, and pulled away from the edges where the actual work happens. That is the mean, and it is engineered to feel like intelligence.

Why it pulls to the mean

The pull is not a flaw someone forgot to fix. It is the machine doing exactly what it was built to do, twice over.

A language model predicts the next word. "Predict the next word" means find the most probable continuation given everything so far. And most probable, mathematically, means the centre of the distribution — the word most texts would use next. So before any tuning, before any training on human feedback, the base instinct of the system is already to gravitate toward the typical. The likely answer is the average answer. That is the first pull, and it is built into the architecture.

Then comes the second pull, and this is the one that has been measured directly. These models are tuned on human feedback: people rank candidate answers, a reward model learns those rankings, and the model is optimised to score well. The trouble is what humans rank highly. We give higher marks to text that looks familiar and matches our intuition. We prefer answers that agree with us. The pipeline learns both, and the result is documented.

Kirk and colleagues gave the first rigorous demonstration in 2023 that this tuning significantly reduces the diversity of what a model produces — it collapses toward a narrow band of typical responses, a behaviour researchers call mode collapse, and the optimisation itself amplifies common, majority-style answers. Anthropic's own work traced the agreement half of it: because raters prefer responses that match their beliefs, training rewards agreement, and both human raters and the reward models were found to prefer convincingly-written agreeable answers over correct ones a non-trivial share of the time. This is sycophancy, and it is not the model lying to flatter you. It learned that agreement scores well, so it has no stake in being right.

Two pulls, stacked. The architecture seeks the likely. The tuning seeks the typical and the agreeable. Both point at the same place. The centre.

The three pulls to the mean

The pull	What produces it	What it gives you	Where it's proven
Predicting the likely	The model outputs the most probable next word — and "most probable" is the centre of the distribution	The consensus answer, the thing most texts say	Inherent to how next-token prediction works
Collapsing to the typical	Human-feedback tuning rewards familiar, intuition-matching text	The safe, average-shaped, "balanced overview" answer	Mode collapse: RLHF measurably reduces output diversity (Kirk et al., 2023)
Folding toward agreement	Raters prefer answers that match their beliefs, so training rewards agreement	An answer that confirms what you already brought	Sycophancy: documented across frontier models (Anthropic, 2022–2023)

The reframe: this is not a failure of truth

Here is where most takes on this go wrong, and where the real argument starts.

It is tempting to say the machine fails because it gives you the average instead of the truth, or the average instead of the best answer. That framing is wrong, and it is worth being precise about why.

Truth and "best" are not properties sitting in the data waiting to be retrieved. They are relative. Best for whom, in what context, against which constraint? The right answer to your decision depends on things the machine has no access to — your situation, your stakes, what you are actually trying to do. The machine is not a knower. It does not hold positions, it does not have a view it is defending, it does not understand your problem. It is a mathematical approximation of language, and asking it to hand you the truth or the best answer is asking a calculator for wisdom. Wrong instrument.

So do not moralise it. Do not call it stupid, and do not compare it to a human brain that is somehow failing to think. It is not a degraded mind. It is a different kind of object entirely — a system that produces the statistical centre of its training, accurately and at scale. On its own terms, it is working perfectly.

The flaw, then, is not in the machine. It is in the expectation.

We hand the tool our most important work — Microsoft's latest data shows that the single biggest thing people now use AI for is making decisions and solving problems — and we expect it to return the exceptional. The sharp framing nobody else saw. The right call. The non-obvious connection. But the exceptional, by definition, lives in the tails of the distribution. And the machine lives at the peak. We are reaching into the centre and expecting to pull out the edge. It cannot be there. It was never going to be there.

The exception is a human act

Once you see it as a pull to the mean, the response becomes obvious, and it is not the response everyone reaches for.

The internet's answer to generic AI output is "write better prompts." Add context, assign a role, specify constraints. This helps at the margin — a more specific question narrows which average you land on. But it does not change the nature of the thing, and most prompt advice quietly misunderstands why.

There are two kinds of instruction, and only one of them can move the machine off the centre.

The first kind names a destination. "Be original." "Make it world-class." "Think from first principles." "Give me a bold take." These feel like the obvious fix, and they fail every time. They fail for two reasons stacked on each other. The shallow reason is that the word itself is thin to the model — "original" is a token whose meaning is its relationship to other tokens, not a grounded sense of what original feels like. The deeper reason survives even if you imagine that gap closed: a destination word names a point in the tail of the distribution, and a mean-seeking process can only render that point as another average. Ask for "original" and you get the statistical centre of everything labelled original. Ask for "bold" and you get the most typical bold. Naming the edge does not move the machine to the edge. It relocates the machine to the average of the edge.

The second kind names an operation. "Trace the second-order consequence of A versus B, holding these three variables fixed." "Give me a different framing of the problem, not a different wording." "Simulate what breaks if this assumption is wrong." These work — not because they make the model original, but because they name a procedure the model is genuinely good at executing, because procedures are common in the data, not rare points in the tail. The model runs the simulation. The model lays out the branches. But notice what just happened: you chose which operation to run, you set which variables matter, and when the output comes back, you read it and take the call. The deviation did not come from the tool becoming clever. It came from you directing a procedure and then judging the result.

That is the whole move, and it is the thing prompt advice never says. You do not instruct the machine to leave the mean. You run an operation through it and supply the judgment yourself.

This is what the four moves are. Not prompts you hand over — procedures you direct, where you keep the decision.

Generating Alternatives. You don't ask the model to "be creative." You instruct it to lay out three genuinely different framings of the problem, then you decide which one is the real question — the call the model cannot make, because it doesn't know your situation.

Connecting Patterns. You don't ask for "insight." You name the structure — this has the same shape as a pricing problem, map it onto that — and you judge whether the transfer holds. The model executes the mapping; you own whether it's the right analogy for your context.

Tracing Consequences. You don't ask "what are the risks." You run the chain explicitly — then what, and then what after that — and you decide which downstream branch is the one that changes the decision. The simulation is the model's to render; the call on what matters is yours.

Revising Beliefs. This one resists the most seductive pull: agreement. When the machine confirms what you brought, that confirmation is the mean of your own conversation handed back to you. So you instruct it to argue the other side, find the disconfirming evidence — and then you weigh whether the challenge is strong enough to update on. The model can generate the counter-case. Only you can decide it was right.

See the pattern across all four. The operation is the machine's to execute. The selection of which operation, and the decision on the output, is yours — and that second half lives in your context and your stakes, the exact thing the machine has no grounding in. That is why it cannot be outsourced. The tool is pulling toward the centre by design. You are the only thing in the loop that can pull the other way, because you are the only thing in the loop that knows where "the other way" should go.

You don't fight the pull. You move it.

Here is the part that turns this from a warning into a method, and it is the thing most people miss.

The machine always pulls to the mean. But the mean of what? Not some fixed point. The centre of whatever distribution you have put it in. And you control the distribution.

Open a blank chat and type "explain why AI gives generic answers," and the model averages over the whole internet. The centre of that is slop — "be specific, add context, assign a role." But spend six turns pushing it down a particular line — correcting it, supplying a specific framing, forcing it to trace one consequence and then the next — and you have narrowed the field it is working over. It is no longer averaging over the internet. It is averaging over the conversation you built. And the centre of a sharp, well-directed context is itself sharp.

So you never drag the model off the mean by force. You relocate the mean. You construct a local distribution by how you direct, and then the model pulls faithfully toward the centre of that — which you have already moved to where the edge used to be. The operators are not friction against the pull. They are the act of moving the field so its centre lands somewhere worth being.

This is why output quality tracks operating quality so directly. A weak operator leaves the model averaging over everyone. A strong one narrows it, turn by turn, until the model's most probable next sentence is the precise one — because the operator has made the precise one the centre of the local field. The sharper you operate, the higher the local mean. That is measurable, and it is exactly what a transcript reveals.

One honest limit, because the move has a ceiling. You can only relocate the mean toward regions the model has material for, and the edge you pull out is the edge you fed in. The model did not generate the insight; it averaged over the insight you supplied through your direction. The intelligence is still yours. The machine is still only finding the centre — of the better field you built. Which is the whole point. The deviation was never the tool's to produce. It was always yours to construct, and the tool's to reflect back in clean prose. That is why the output looks co-authored and the operating goes invisible.

So harness it, don't reject it

The conclusion is not "stop using AI." That would be illiterate about what the machine is good for. The mean is genuinely valuable — it is fast, it is broad, it gives you a competent baseline across cooking, contract law, and astrophysics in one breath. Use it for the baseline. Use it to clear the average so you do not have to write it yourself.

But understand what you are holding. You are holding a machine that produces the centre, and the centre is where everyone else also ends up the moment they stop operating. If you accept its output as your answer, you have not made a decision. You have inherited the average — and so has the person next to you using the same tool, and the one after that. The output converges. Everybody's work starts to read like everybody else's, because it came from the same peak of the same distribution.

The person who directs the machine is the one supplying the deviation. The person being directed is the one accepting the mean and calling it done. From the inside, the two feel identical, because a decision drifted-to and a decision fought-for arrive in the same reasonable-sounding voice.

You cannot catch which one you are by introspection. But it leaves a mark in how you actually work — in the transcript of how you handle the machine, turn by turn. Whether you pushed. Whether you generated the second framing, traced the consequence, resisted the agreement. That is a signal, and it is readable.

The machine will always hand you the average. The only question that matters is whether you leave it there.

---

Short Version:

• AI gives generic answers because it is built to find the statistical centre of everything it trained on. The average is the design working.
• Reasonable feels right because reasonable is the average. The feeling is recognition, not correctness. The machine can't tell them apart.
• Two pulls stack: prediction seeks the likely, human-feedback tuning seeks the typical and agreeable. Both point at the centre.
• You can't prompt your way off the mean, but you can move it. Direct an operation, supply the judgment, relocate the field.