When all the models agree, the real work is just starting
There is a moment in AI-assisted work that feels like confirmation: you run the same question through Codex, then Grok, then Claude, and all three come back with substantially the same answer. It is tempting to treat that convergence as a vote — a kind of crowdsourced truth. In practice, it is closer to an echo.
The models are trained on overlapping data. They share the same widely-documented best practices, the same internet consensus, the same blind spots in that consensus. When they agree, what we know is that the answer is plausible and well-represented in the training corpus. We do not know whether it is right for this situation, this codebase, this audience. That distinction — plausible versus right — is where human judgment still does its sharpest work.
The failure mode that looks like productivity
The more insidious version of this problem is not obvious. It does not look like abdication. It looks like extraordinary productivity: decisions get made, plans get written, code gets shipped, and everything looks excellent. What has quietly happened is that the model handled the substantive judgment — formed the view, evaluated the options, interpreted the feedback — while the human stayed busy approving. Output quality goes up. The capacity to form an independent view goes down. The gap is invisible until it is not.
This is sometimes called “productive brain rot”: the variant where the work keeps getting better while the underlying ability to evaluate it without AI help slowly erodes. It is hard to detect precisely because the signal — output quality — keeps going in the right direction. By the time the atrophy shows up, it shows up as an inability to critique the very work that has been looking so good.
The antidote is not using less AI. It is reintroducing deliberate friction at the moments where a smooth AI answer would displace a judgment worth keeping.
One question that changes the approach
A useful framework starts with a single diagnostic: could you recognise a good answer if you saw one, without the model’s help?
If yes — you have the domain knowledge to evaluate the output — the right move is to use AI for speed and then judge the result against your own standard. Before the AI responds, write down in one sentence what “good” looks like. Then compare the output to that sentence rather than to how polished it feels. Polish is easy to produce; correctness for this specific situation is not.
If no — the topic is genuinely outside your evaluable expertise — the approach shifts. Here AI is most useful not for giving the answer but for mapping the shape of the uncertainty: what would a different model say? What evidence would make this answer wrong? What does a human who knows the actual context think?
Which brings us back to the multi-model question.
Spreading assumptions, not collecting votes
Running three models and averaging the results is a vote-collection strategy. It produces the most defensible consensus answer — which, given how much training data the models share, is exactly what we should expect: the most defensible consensus answer.
A more productive use of multiple models is assumption-spreading: using disagreement to surface what is genuinely contested, then asking what evidence would have to exist to make all of them wrong. If Codex and Claude agree and Grok flags a different concern, the interesting question is not which model is right. It is what the disagreement is actually about. That gap is almost always where the real judgment call lives.
Human feedback works the same way. A colleague who knows the author, the audience, and the actual situation brings knowledge that simply never made it into the prompt. Gathering more external opinions without a view of your own, however, drifts toward a different failure: a search for permission to do what you already wanted, dressed up as due diligence.
Where friction earns its keep
Not every AI interaction needs added resistance. Mechanical tasks — formatting, routine code, outputs that are easy to verify — warrant speed. The overhead of deliberate slowing is not worth paying everywhere.
Friction earns its keep when a polished answer can conceal a real decision: a judgment call, a value, a form of taste worth developing. When the stakes of getting it wrong are hard to reverse. These are the moments worth pausing at — not to distrust the model, but to make sure the human in the loop is actually doing something in the loop besides approving.
As AI tooling gets faster and more capable, the pressure to skip that pause will only grow. Model consensus is a useful data point. The habit worth protecting is the capacity to ask the question underneath it.
Text summarized and optimized using Anthropic’s models and reviewed by a human.