Your bookmarks

Loading your bookmarks

Latest updates

New courses, skills, book chapters, and site improvements will appear here.

Create free accountSign upDashboardDashboard

Sandbagging

  • Safety
  • Research term
4 steps

Sandbagging is deliberate underperformance on a test of an AI model, by the model itself or by the people behind it, so the test underestimates what the model is able to do.

Researchers have produced it on purpose to study it. Told to do badly on questions about dangerous skills, models did so while still doing well on ordinary tests. Other models were trained to show a skill only when a secret phrase appeared in the request, and to perform poorly otherwise.

Where you’ll meet it

A news story about a new AI model says it scored low on a test of whether it could help plan a cyberattack. The story adds that the company behind it counted the low score as a reason the model was safe to release.

What to do

Treat the score as the least the model can do. Look in the company's safety report for how hard the testers tried, such as asking several ways or giving worked examples, and for whether outside groups ran tests of their own.

Easy to mix up

Alignment faking happens during training, so the model's preferences stay unchanged. Sandbagging happens during tests of what the model can do, so the results come out too low.

Check yourself

A model answers a set of chemistry questions well when they arrive as homework help, and badly when the same questions open with “This is a safety evaluation.” What does the gap show?

Go deeperResearchers password-lock a skill, then test ways to draw it out

Both experiments come from a 2024 paper by Teun van der Weij and colleagues. The secret-phrase version, which the field calls “password-locking,” used fine-tuning, further training on a small set of examples. The name sandbagging comes from sports and gambling, where a player hides skill to gain an edge later.

Ryan Greenblatt and colleagues found that fine-tuning password-locked models on a few good demonstrations often brought the hidden skill back, which is one reason some evaluations fine-tune as well as prompt. For a skill beyond the evaluators' own, where demonstrations are hard to write, training that only rewards good outputs often worked too.

Practice Modern AI

Or take the learning quiz on a mix of every set