Follow a topic across the labs

What each lab's system cards say about one evaluation, lined up by lab and date. Every mention opens the page it is on.

34 of 113 cards
3 labsSabotageWhether a model could undermine oversight, evaluations or the work it is given.24 of 113 cards
3 labsSycophancyTelling people what they want to hear.19 of 113 cards
2 labsModel welfareWhat the labs report about the possible welfare of their models.26 of 113 cards
2 labsSandbaggingDeliberately underperforming on an evaluation.13 of 113 cards
2 labsAlignment fakingBehaving as trained while being watched, and otherwise when not.27 of 113 cards
3 labsReward hackingPassing the test without doing the task.48 of 113 cards
4 labsDeceptionA model misleading the people it works with or for.23 of 113 cards
2 labsEvaluation awarenessA model noticing that it is being tested.