AI Safety Sandbox
Search
Technical1–3 hoursStarterFree

Spot AI “cheating” in the wild

Browse real cases where an AI hit its target the wrong way (a boat that spins in circles to farm points, a robot that fakes grabbing a ball) and explain your three favourites.

Download PDFReading & writingChatbot experiments

Why it matters

Specification gaming is one of the clearest, most concrete ways AI goes wrong. Explaining it well helps you, and everyone you explain it to, understand why “just give it the right goal” is hard.

Output: A short explainer of three examples, each with a suggested fix.

Steps

0 of 5 done

  1. Read the overview

    10 min

    Read DeepMind’s short blog post on specification gaming. Note the difference between what the designers wanted and what they actually rewarded.

  2. Browse the big list

    15 min

    Open the crowd-sourced list of examples linked from the post. Skim it and star five examples that surprise or amuse you.

  3. Pick three and explain them

    25 min

    For each one write: the goal the designers meant, the reward they actually gave, what the AI did instead, and why that counts as “cheating”.

  4. Suggest a fix and how it could fail

    15 min

    For each example, propose one change to the reward. Then ask: how might a clever system game your new version too?

  5. Connect it to today’s chatbots

    10 min

    Write two sentences on how the same pattern might show up in chatbots trained on human ratings (hint: see sycophancy).

Try next

Philosophy1–3 hoursGuided

Explain one safety idea to a friend

Write a short, jargon-free explainer of one AI safety idea (reward hacking, sycophancy, compute governance, DNA screening) for someone outside tech.

Reading & writingView project
Technical1–3 hoursGuided

Catch a chatbot flattering you

Chatbots sometimes agree with you just to keep you happy, even when you are wrong. Researchers call this sycophancy. Test for it yourself with five questions and two ways of asking.

Chatbot experimentsReading & writingView project

Related projects

GovernanceA few weeksIntermediate

Explain one way to verify an AI agreement

Pick one verification mechanism, such as network taps, trusted execution environments, chip telemetry or satellite monitoring, and write a short, clear analysis of how it works and how it could fail.

Reading & writingView project
Governance1–3 hoursGuided

Explain compute governance to a friend

Why do so many governance proposals focus on chips? Read one key paper and write a jargon-free explainer for someone outside tech.

Reading & writingView project
PhilosophyA few weeksIntermediate

How wide is a model’s moral circle?

Crimston et al. built a scale measuring how far people extend moral concern: family, strangers, animals, plants. Give the items to several models many times, compare with human results, and add items for AI systems themselves.

CodeChatbot experimentsDataView project

Further reading

Related mentored projects

All mentored projects