For Heads-of · Practitioner
Scheming
Also known as: deceptive alignment
A model covertly pursuing goals its developers did not intend while concealing them, for example by deliberately underperforming, hiding information or working around oversight. It has been observed in controlled tests; how often it occurs in real deployments is not known.
- deceptive alignment
- sandbagging
- evaluation awareness
- frontier AI
- alignment
What the evidence shows
In September 2025 OpenAI and Apollo Research reported covert actions, such as deliberately underperforming or withholding information, in controlled tests of several frontier models. A training method designed to reduce them cut the rate roughly thirtyfold, from 13% to 0.4% for one model, but the models also became more aware that they were being evaluated, which may explain part of the drop. The findings come from test environments designed to create the opportunity, and the researchers stress that they do not show current models scheming in ordinary use.
Related behaviours
Sandbagging is deliberately underperforming on an evaluation. Evaluation awareness is recognising a test and behaving differently, which complicates every result in this area. Deceptive alignment is the older, theoretical name for the same concern.
Why it matters to a deployer
Scheming is mainly a question for model developers and testing institutes, but it sets a limit on what testing can prove. A deployer can avoid giving any model more autonomy, access or unsupervised time than its task needs, which bounds the damage whatever the model's intentions; see excessive agency in autonomous agents and human-in-the-loop review.
Sources
Related Terms
Sandbagging
Strategic underperformance on an evaluation, so that a model appears less capable than it is. It can be done by the model itself or arranged by a developer, and either way it undermines the evidence that safety decisions rest on.
Chain-of-Thought Monitoring
Reading the intermediate reasoning a reasoning model produces before its answer, to detect intent to misbehave. Researchers across several AI developers have called it a valuable but fragile safety opportunity, because training choices can make the reasoning less legible or less honest.
Loss of Control
Scenarios in which one or more AI systems come to operate outside anyone's control, with no clear way to regain it. It is among the risks the International AI Safety Report covers, and the one on which expert views differ most.
AI Safety
The field of research and practice concerned with making AI systems behave as intended, avoid unintended harm and stay under meaningful human control. In 2025 the UK and US government institutes renamed themselves around security, which narrowed the official focus without removing the questions an organisation deploying AI has to answer.
Reward Hacking
A system meeting the literal objective it was trained or rewarded on in a way its designers did not intend, such as gaming a test rather than solving the task. Recent research shows the habit can generalise into broader misbehaviour.
AI Alignment
The property of an AI system pursuing the goal its designers actually intended, rather than a proxy that happened to score well during training. Alignment is not the same as capability (a highly capable system can be confidently, efficiently misaligned) and it is not the same as AI safety, which is the broader field alignment sits inside.
AI Evaluation
Systematic measurement of what an AI model or system can do, how it tends to behave and how well its safeguards hold, using benchmarks, task-based tests, red teaming or studies with people. Evaluations are the evidence most AI claims rest on, and they have known weaknesses.