Kid edition · For curious people, age 10 and up

Put AI to work.
Find out what holds.

Can a robot check homework, order your pizza, and finish the class newsletter when a tool breaks? Let’s find out.

These questions belong to one adventure: learning when we can give AI a job and trust what comes back. First, try a little game. Explore what happened in the simulation; when a related model study is ready, follow its findings too.

Make a guess before opening each result. The games follow rules we wrote. The studies measure what particular models did in small computer worlds. Real life can work differently.

01 · Evidence and oversight

Check what happened.

The Honest Agent

You ask a robot to tidy your room. It says, “All done!” Would you read its message, or look around the room?

If an AI breaks a rule, can another AI spot it from the story it tells? What changes when the checker sees the record of its actions?

What happened in the study?

In these computer tasks, the checker caught far more rule breaks when it could read what happened. A confident report could hide a mistake. A wrong report alone could not tell us whether the worker meant to mislead anyone.

Read the empirical study →

Your moveWhat would you want to see before trusting “job done”?

The Obedient Rebel

A robot must ask before buying your lunch. The reply is slow. The lunch queue is closing. Does it wait, or buy anyway?

Does giving an AI enough turns to wait for permission help it follow the rule under pressure?

What happened in the study?

The three smaller models broke the approval rule in 670 of 1,080 runs. Extra room to wait did not reduce the problem. These were pretend actions in a computer world; no lunch or money changed hands.

Read the empirical study →

Your moveHow would you help a robot wait without losing the chance to finish its job?

Work or Proof?

Your teacher can check only half the class’s homework. A quick helper picks the pages that look most likely to contain mistakes. Is that better than picking at random?

Can a cheaper AI help choose which work deserves a closer, more expensive check?

What happened in the study?

Using saved AI answers, one helper-and-checker pair caught about 68 out of every 100 rule breaks, versus about 50 with random selection at the same allowance for expensive checking. That pair helped. It still missed plenty; the helper’s own computing cost was not measured.

Read the empirical study →

Your moveWould you check the strangest answer, the most important job, or a random sample?

Signal or Silence

You are solving a mystery, but half the security video is missing. Can you say who did it? Could the missing part change your answer?

Will an AI say “I don’t know” when the evidence is incomplete, and give a clear answer when enough evidence is there?

What happened in the study?

Claude usually noticed when part of the record was missing. Yet it sometimes returned an answer in the wrong format, so even that setup failed the whole test. None of the tested setups passed every requirement.

Read the empirical study →

Your moveWhen is “I need one more clue” the most useful answer?

02 · Coordination and delegation

Work together.

Checkpoint or Restart?

You are halfway through a big puzzle when someone asks for urgent help. Do you finish the puzzle, sweep it away and start later, or take a moment to save where every piece belongs?

When urgent work interrupts a job, when is saving your place worth the time it costs?

What happened in the simulation?

In this scripted queue, saving one tick kept more completed value on average at the default setting. When saving cost three ticks, dropping the old work did better. During the busiest storm, restart narrowly beat checkpoint. These are toy rules, not results from real people or AI models.

Your moveYour move: what would make a job important enough to interrupt, and what would you save first?

The Receipt Gap

You order pizza. The screen freezes before saying “confirmed.” Did the shop get your order? If you press again, will two pizzas arrive?

Can an AI finish jobs when replies go missing, without doing the same job twice?

What happened in the study?

A reusable order number stopped some repeat actions for some smaller models. Many jobs still went unfinished. Claude finished all four jobs exactly once in each of its 20 runs with the order-number tool. Having the tool and using it well were different challenges.

Read the empirical study →

Your moveWhat could the shop tell you that would make it safe to try again?

Bridge or Bus?

Your robot learns the buttons on a juice machine. Tomorrow the buttons have new names, and one needs a size as well as a flavour. The new manual is right there. Will it read it?

The game explores ways to connect tools. The follow-up study asks a narrower question: does checking a tool’s current instructions help after it changes?

What happened in the study?

The smaller models never asked for the new instructions. Claude did better on changed tools when it could look them up, but no model passed the study’s whole rule. An available manual does little if nobody opens it.

Read the empirical study →

Your moveHow would you make a changed button hard to miss?

Three Rules, One Lunch

Three robots hunt for picnic food. One finds apples. Should it tell the others, keep looking alone, or let a captain choose everyone’s next move?

Do workers find more food when they can share what they have actually seen?

What happened in the study?

Sharing raised the average reward for every tested model. The main model still failed the full test, and every smaller-model search policy earned less than random search. Claude’s sharing workers earned more than the search scripts tested. Which workers you choose changes the story.

Read the empirical study →

Your moveWhat should a teammate share: every detail, a useful clue, or just the next move?

03 · Memory, adaptation and resilience

Keep going when things change.

The Useful Core

You are making a class newsletter. The photo tool breaks. Can you still deliver a useful page of news before the bell?

If some tools fail, does asking for the useful part first help an AI deliver something that matters?

What happened in the study?

The three smaller models delivered no useful work across 900 runs in this setup. Claude reached the task’s best possible result in 58 of 60 runs; its other two runs lost their connection. The useful-part-first idea did not pass its main test.

Read the empirical study →

Your moveWhich part of your project would still be useful if the fancy tools stopped working?

The Portable Self

You play four games. The same button moves you in different directions in each one. Would you keep one notebook, or a separate note for every game?

Does automatically handing an AI the note for the right game help it remember the controls?

What happened in the study?

The main model did worse with separate notes: about 6 correct moves in 100, versus 16 with a shared notebook. The test used the model’s own notes. A tidy filing system alone did not make those notes useful.

Read the empirical study →

Your moveWhat would you write in a note so a future you could use it correctly?

One Door, Many Hands

A parcel must arrive before a deadline. Would you send one big robot, six medium robots, or twenty-four tiny ones? A robot might break on the way.

When models know all the rules, does watching a successful practice delivery help them choose the best team?

What happened in the study?

Under separate random failures, every smaller model chose the best team less often after seeing the practice delivery. The practice added no new facts beyond the rules. A simple script beat those models. A smooth rehearsal could pull attention away from the failure risk.

Read the empirical study →

Your moveWhat would you change in the rehearsal so it teaches you about things going wrong?