- The idea: some questions have one clear right answer that a machine can check. Others have better and worse answers, but no single one that can be checked. AI has moved fastest on the first kind.
- The warning: wins on the checkable kind make people assume AI is just as good at the rest. It is not, at least not yet.
- This week: three pieces of research show the pattern, and each has limits worth knowing.
Three famous wins, one clear pattern
In 1997, a chess computer called Deep Blue beat the world champion, Garry Kasparov, over a six-game match. It was IBM's machine, and it was the first computer to beat a reigning champion under standard tournament rules. In 2011, IBM's Watson won the quiz show Jeopardy! against two of its strongest players. In 2016, Google DeepMind's AlphaGo beat the Go champion Lee Sedol four games to one.
Andrew Ng, who founded DeepLearning.AI and runs its AI courses, mentions all three in his latest letter for The Batch. His point is that each of these games had a checkable answer. A chess move is legal or it is not. A quiz answer is right or wrong. A machine can test its own moves against the rules, over and over, until it gets better.
Ng calls these verifiable tasks. The tasks that are harder for AI are the judgement tasks: handling an upset customer, or deciding a company's strategy. These have better and worse answers, but no single right one a computer can check. His warning is that the wins on verifiable tasks lead people to overestimate AI on judgement tasks.
Why this matters for you
When you use an AI tool for homework, code or work, one simple question helps. Could I check this answer? If you can, the tool is likely to be useful and you can verify it quickly. If you can't, keep your own judgement in the loop, and treat the answer as a draft.
What happened this week
Three recent items show the pattern at work.
- A cryptography puzzle. On 16 October, Anthropic, the company behind the Claude models, said its AI had worked out where the "seeds" behind five standard codes came from. A seed is simply a starting number for a random-number generator, and these codes were set up in 2000. The check is easy: a candidate seed either rebuilds the code or it does not. Anthropic says the finding is "not a cryptographic break, new attack, or weakness". It is also the company's own account, and outside experts have not yet reviewed it.
- Quantum circuits. A preprint posted on 15 October describes a language model proposing designs for quantum circuits, which are the step-by-step recipes a quantum computer follows. An automatic checker scored each design. The authors say the best designs beat the best published ones in seven of eight tasks, using fewer qubits (the quantum equivalent of bits) and fewer steps. A preprint has not yet been checked by journal reviewers.
- Maths papers. Earlier this month, our own story reported that OpenAI had shared 722 maths papers written with AI. Of those, 162 carry a proof that a computer can check.
Ng's loops for building things
In a June letter, Ng described three loops for building software with AI. In the first, an AI agent writes code, tests its own work, and keeps going until it meets the spec. In the second, a developer reviews the product and steers the agent. In the third, real users and tests feed their results back into what gets built next.
Ng's point is that humans still bring a context advantage. They know who the users are, what the product is for, and what would go wrong in the real world. A checker cannot see any of that. That is the judgement part, and in his view it is the part that cannot be automated away.
A limit that never goes away
Ng's letter also mentions the Bayes error. In simple terms, it is the lowest error any system can reach on a task. If a question has no single fixed answer, some error is unavoidable, however clever the system is. Think of predicting the exact score of a football match: even the best analyst will sometimes be wrong, because the game itself has randomness. Some judgement questions work the same way, which is why the letter's warning matters.
What to watch
Three things are worth keeping an eye on.
- Who builds the checker. A checker that a model can fool is no real checker.
- Where the line moves. Researchers keep finding ways to check more kinds of work. Each time they do, a task that used to need judgement may move into the checkable group.
- What the checkers miss. The cryptography work still leaves one seed (B-163) and four base points unexplained. Outside experts have yet to look at it.