Judgement isn’t a permanent hiding place

The reassuring story goes like this: AI will do the execution, while people supply the taste and judgement.

I don’t think that’s a safe place to stop thinking. It may describe useful work today. It doesn’t explain why that division should last.

I expect better models, with relevant and current context, to take on much more judgement. That prediction needs evidence of its own: the absence of a permanent human advantage does not establish inevitable replacement.

What are we calling taste?

Taste can mean recognising what fits an audience, choosing an argument or rejecting something impressive that misses the point. Working alongside someone exposes the choices, corrections and exceptions behind it.

I think machines could learn more of those patterns. But anticipating a preference is different from helping someone choose well. A model could reproduce a choice that the person later regrets. People can also agree on consequences and disagree about what matters.

More messages do not automatically create good taste. The question is what the system can learn and use.

A trade-off isn’t a spell that stops automation

Imagine deciding whether a feature should ship on Friday. Keep the actual defect the same, and change only what the decision-maker knows.

The test is whether the AI can use the fuller brief, compare options and identify uncertain assumptions. Access to information doesn’t guarantee a correct inference. An information gap and a capability gap are different problems.

An experiment with 758 BCG consultants, published in 2026, shows why both possibilities matter. GPT-4 assistance improved selected creative and analytical tasks, but made correct answers less likely on a managerial problem beyond its capabilities. These were designed tasks using an older model, not an autonomous consultant.

Context changes the comparison

We often compare a person with years of context against a model handed a paragraph, then credit the whole difference to an irreducible human quality.

Research revised in June 2026 used interviews and surveys from 1,052 people. Agents with personal context predicted held-out survey responses more accurately than agents given demographics alone. That supports taking context seriously; it is not evidence of better life decisions.

Context can also be stale, contradictory or irrelevant. A 2023 comic-recommendation study found that fixed-preference assumptions failed even in that narrow setting. Yesterday’s preferences are not necessarily today’s goals.

What would make the prediction stronger?

Give people and models the same relevant information and unfamiliar decisions. Agree on success beforehand—including whose priorities count if the goal is disputed.

Count correct answers, confident mistakes, missed constraints and the full cost of checking. Gains that transfer to new situations would support handing over more work. Gains that require constant expert repair would count against it. The studies above do not establish that broad result.

The scarce thing can move again

Eero Alvar’s What Remains Valuable When Anyone Can Build Anything? expresses a premise I share: cheaper cognition moves scarcity. It doesn’t remove it.

Good judgement could remain valuable while becoming cheaper. That would be a gain, and a challenge to people who sell it. But cheaper advice does not automatically mean fewer jobs or lower pay: demand, implementation costs and the desire for human responsibility still matter.

Capability also does not grant authority over someone else’s life. Consent and accountability remain separate questions from who reasons better.

I still want to develop taste and judgement. They help me do worthwhile work now. I don’t want to confuse developing an ability with securing a permanent monopoly on it.

“I understand the trade-offs” should begin an explanation of what you know. It shouldn’t end the conversation about what AI could learn.