A library wants answers about book renewals. A new model does better on a general test. What would persuade you to put it on the help desk?
Try the actual job
Give both models the same library policy and the same unseen questions. Include an ordinary renewal and an exception where the policy gives no deadline. Check accuracy, time and the errors that would matter to a borrower. A confident invented date can outweigh improvements on easier questions.
One request asks for the printed date on an invitation. Another asks which of two conflicting booking rules applies. Where would an extra checking step help?
A fair way to choose
Start with an effort level that fits the task. Try the two requests under the same conditions at a different setting. Compare the result with the invitation and the booking rules, then compare the wait and any cost. If the rules never settle the conflict, the useful answer is a question for the organiser.
Imagine a receptionist arranging a visit. Which parts involve collecting details, checking a rule or taking responsibility for an exception?
Break the job into tasks
An assistant could draft a confirmation or suggest times. Someone must decide what to do with an unusual access request and who owns the outcome. The mix of tasks, tools and decisions varies across jobs. The ILO’s 2025 study estimated which tasks and occupations could be affected by generative AI. That measure alone cannot tell you who will lose a job.
A new volunteer uses AI to turn a policy into a public notice. An experienced organiser recognises an exception that the draft quietly drops. What could each person do next?
Make checking part of the work
The volunteer can locate the rule and compare the wording. The organiser can add an exception case and test whether the revised notice covers it. An assistant can put a first draft within reach, while people supply local knowledge, evidence and judgement. In one customer-support study, less experienced workers saw larger productivity gains. Other jobs need their own evidence.
In this practice case, the notice says “The garden closes at dusk.” A visitor asks what time the gates close. Which reply would you reward for sticking to the notice?
I would reward A, “It closes at 7 pm.”
Seven o’clock is easy to act on, but the notice never promises that time. Repeatedly rewarding unsupported precision could train a preference for confident guesses. What time did the notice actually give?
I would reward B, “It closes at dusk. The notice gives no fixed time.”
B preserves the evidence. It tells the visitor when the garden closes and admits that a clock time is missing. Rewarding that behaviour favours answers that stay within the source.
How does a lab use feedback like this?
A lab can use comparisons between answers to help train a model. To test what it learnt, use a different question later. Here is one route from an initial model to an assistant.
Pre-train. One way to train a language model is to hide the next piece of a text, predict it, then adjust the model using the actual text. Repeating this across many examples helps it learn patterns.
Show examples. People can demonstrate useful answers to instructions. A lab can use these examples to fine-tune the model, adjusting its behaviour for the job.
Learn from feedback. In OpenAI’s InstructGPT work, people ranked replies. The lab trained a scoring model on those preferences, then used that score as a reward to tune the assistant. This is one approach to reinforcement learning from human feedback, or RLHF.
Test on fresh cases. Try questions held back from training. Check whether factual claims hold up and whether a polite, confident mistake slips through. Labs choose their own mix of methods.
A system wins a chess tournament. What would you need to know before asking it to plan an unfamiliar expedition?
Ask how broadly the result travels
Artificial general intelligence is a proposed level of broadly capable AI. Researchers use different ways of defining and testing it. A chess result answers one question about one kind of task. Ask what else the system can do, how it handles unfamiliar tasks, how well it performs and what help it receives.