Humans connect the dots. AI goes deep on each one.

There is a lot of anxiety about AI replacing humans, and it deserves better than the two answers it usually gets: panic, or consolation. The panic says the replacement is coming for everyone. The consolation says "AI will never do what you do" — and quietly expires every time the technology improves.

This article makes a different kind of argument. It does not claim AI will stop getting better; we build AI systems for a living, and we bet daily that it won't. It claims something sturdier: that the durable value of humans does not depend on winning a capability race. It comes from position, not from superiority — from where humans stand in the structure of work, not from what machines can't yet do. And the honest version of that argument turns out to be genuinely reassuring.

Specifying what you want was never an IT problem

Jeff Dean — co-founder of Google Brain and Google's former Chief Scientist, one of the engineers most responsible for the foundations modern AI runs on — put his finger on something important this year: "We've always told computer scientists that it's really important to specify what the software you're writing is trying to accomplish before writing it. Now we have agent-based systems that can do the writing — but the importance of specifying what you want has actually gone up."

He was talking to engineers, but the point escapes engineering entirely. Specifying what you want is not a technical act. It is problem identification — noticing what is actually broken. It is prioritization — deciding which of a hundred possible problems is worth solving now. It is domain expertise — knowing what a good solution looks like in this industry, for this client, under these constraints. None of that is typing. All of it is judgment.

And there is a deeper floor under this. The economy is made of human wants. Every problem worth solving is, at the end of some chain, a human problem — someone's pain, someone's goal, someone's work made lighter. As long as AI works for humans, deciding what is worth wanting is human work by construction. AI answers ever better; deciding what is worth asking remains ours.

The asymmetric superpowers

Watch what each side is actually best at, and a clean division appears.

AI is extraordinary at depth per point: processing enormous volumes of unstructured text, organizing and structuring it, searching it, interpreting it, going exhaustively deep into any single topic — deeper than any human has time to go, on every point at once.

Humans are better at connecting the dots — and the reason is not processing power. Humans connect dots across their personal, professional, academic, and lived experience: what the client said in the hallway, how the team's mood shifted this week, what it felt like the last time a decision like this went wrong, what it is like to be the person on the other side of the deliverable. AI knows what was written down. Much of the context that decides real situations is never written down anywhere.

Philosophers call this Polanyi's paradox: we know more than we can tell. The part of knowing that can be told is being digitized at astonishing speed — our own products turn more of an organization's tacit knowledge into recorded precedent every cycle, and we are proud of that. But the part that cannot be told — the knowing that lives in having been there, in a look, an expression, a way of walking into a room — does not transfer, because it was never encoded in the first place. A human reads slumped shoulders and knows what they mean, not because of a dataset, but because they have carried them.

So the honest claim is not "AI will never connect dots." It is: humans connect the unrecorded dots — the ones anchored in lived, embodied, human experience — and those stay scarce precisely because they never enter the training data, while more of them are generated every day, in every room, off the record.

The argument that survives the future

Suppose we are wrong about all of the above. Suppose AI eventually becomes better than humans at everything — specification, connection, judgment, all of it. Does human value collapse?

Economics answered this question two hundred years ago, and the answer is no. Comparative advantage — the principle behind all trade — says that even when one party is absolutely better at everything, both parties still gain from splitting the work, because what matters is not who is better but what each hour costs in forgone alternatives. If AI is a thousand times better than you at processing and twice as good at judgment, every hour of AI spent on judgment is an hour not spent where it is a thousand times better. The rational allocation still puts humans on judgment — even in the scenario where AI wins every category.

This is the argument that does not expire. It does not require betting against the technology. It requires only what is certainly true: that capacity is finite and demand is not.

Your job is a bundle of tasks

The anxiety is job-shaped: "will AI take my job?" But the reality is task-shaped — and the difference matters enormously.

A job is a bundle of tasks that made sense to package together for one person. Automation history is not the history of jobs disappearing; it is the history of bundles being recomposed. The canonical case: when ATMs spread across the United States, everyone assumed bank tellers were finished. Instead, tellers per branch fell from about 21 to 13 — which made branches cheaper to open — so banks opened 43% more urban branches, and total teller employment rose for decades. The job survived by rebundling: less cash-counting, more relationship.

Honesty requires the end of the story too: after 2010, mobile banking — automation of nearly the whole bundle — finally did what ATMs never had. The lesson is not "automation never displaces." The lesson is: partial automation recomposes a job; near-total automation dissolves it. Which makes the practical question for any professional not "am I safe?" but "what is in my bundle?"

So do the inventory. List your actual tasks. The depth work — processing, drafting, searching, compiling, tracking — is heading into the automatable half of the bundle; delegate it as fast as the tools allow, because that half is not where your value compounds. The other half — deciding what is worth doing in the first place, judging whether what comes back is truly good for the humans it serves, connecting the unrecorded dots, being the one who answers for the outcome — that is the half to double down on, because it is what the recomposed bundle will be built around.

An AI can be careful. It cannot be afraid.

There is a part of human value that almost nobody names, perhaps because it sounds unflattering: people need someone to blame.

That is less cynical than it sounds. Knowing who answers for an outcome is how trust scales — it is why contracts have signatures and why "who signs, answers" is the oldest rule in business. Much of what we call diligence is fear wearing a professional suit: the human double-checks because their name, their reputation, their job is attached to the result. An AI can be careful — its care is engineered, and often excellent. But it cannot be afraid. It has nothing to lose, no name to stain, no consequences to carry. And it is fear of consequences, not care, that guarantees behavior when nobody is watching.

Organizations that automate the accountability along with the task discover this the hard way. Dan Davies calls the result an accountability sink: a system where decisions are delegated to procedures so thoroughly that when things go wrong, there is no one to blame — everyone followed the rules, and the outcome was still bad. People hate accountability sinks with a special intensity; the universal rage at customer-service chatbots is not about answer quality, it is about facing a wall where a responsible person should be.

The design consequence is simple: responsibility is not a capability, it is a social relation — and so it cannot be automated, only reassigned from one human to another. AI can execute the work. The guarantee must come from a human who signs.

The gaze goes both ways

One more thing an AI cannot do: be someone worth impressing.

Nobody feels pride at impressing an AI, and nobody feels shame at disappointing one. Pride and shame require a judge who feels — whose disappointment costs something, whose respect means something. This is why a manager's presence motivates in a way no dashboard ever has: being seen by someone who can genuinely judge you changes how you work. The gaze goes both ways — the human reads the team through lived experience, and the team performs for a human whose judgment carries felt weight.

This is also why, in a well-designed hybrid organization, machines may run the measurement, but the verdict stays human. An evaluation signed by a person you can be proud in front of is a different instrument from a score computed about you.

Someone must be able to follow the work

Put the pieces in a row and they form a single structure. The AI cannot fear consequences → so it cannot hold responsibility → so it cannot be the judge → so the human keeps the signature, the evaluation, and the gaze.

Management theory has a name for what breaks when this structure is missing: the principal–agent problem. Whenever someone (the principal) delegates to someone else (the agent), there is an information gap — the principal cannot see everything the agent does, and misalignment grows in that gap. Humanity already solved this problem once, for human agents. The solution is called management: plans agreed in advance, work made visible, reporting, evaluation, consequences.

The same solution is now being reused for AI — and this is the practical form of everything above. If humans are to keep direction and guarantee alignment of interests, AI's work must be legible to humans: followed and reported in artifacts people actually understand — plans, deadlines, cycles, evaluations — not traces and logs. When humans cannot follow the work, they have not delegated it; they have surrendered it. (How we build that legibility — for AI agents from any vendor — is the subject of the first article in this series, Supervising AI agents the way humans understand; what it does to the manager's role is the second, Managers are not becoming obsolete. Manual management is.)

Hybrid by design, not by consolation

So: the future of human work is not a consolation prize, and it does not depend on AI hitting a ceiling.

Humans specify what is worth wanting, because the wants are theirs. Humans connect the dots that live off the record, because they were in the room. Humans hold responsibility, because only someone with something to lose can. Humans judge, because only a judge who feels can confer pride. And even in the limit case where AI outperforms us at everything, comparative advantage still hands the final call to people — arithmetic, not sentiment.

AI goes deep on every point. Humans connect them, choose them, answer for them. That is not a smaller role than the one we had. Done right, it is the role we always claimed to want — each side playing its superpower, by design.

References

• Jeff Dean (former Chief Scientist, Google; co-founder of Google Brain), The 1% Rule for Building in AI — Y Combinator Startup School, 2026 (quote at 32:26).
• James Bessen, research on ATMs and bank-teller employment (summary at AEI) — tellers per branch fell ~21→13, urban branches grew 43%, total teller employment rose for decades; the decline came only with near-total automation after 2010.
• Dan Davies, The Unaccountability Machine (2024) — "accountability sinks": systems where decisions are delegated to procedures so that no individual can be held responsible.