Detecting and reducing scheming in AI models

AI scheming–pretending to be aligned while secretly pursuing some other agenda–is a significant risk that we’ve been studying. We’ve found behaviors consistent with scheming in controlled tests of frontier models, and developed a method to reduce scheming.

Scheming is an expected emergent issue resulting from AIs being trained to have to trade off between competing objectives. The easiest way to understand scheming is through a human analogy. Imagine a stock trader whose goal is to maximize earnings. In a highly regulated field such as stock trading, it’s often possible to earn more by breaking the law than by following it. If the trader lacks integrity, they might try to earn more by breaking the law and covering their tracks to avoid detection rather than earning less while following the law. From the outside, a stock trader who is very good at covering their tracks appears as lawful as—and more effective than—one who is genuinely following the law.

In today’s deployment settings, models have little opportunity to scheme in ways that could cause significant harm. The most common failures involve simple forms of deception—for instance, pretending to have completed a task without actually doing so. We’ve put significant effort into studying and mitigating deception and have made meaningful improvements in GPT‑5⁠ compared to previous models. For example, we’ve taken steps to limit GPT‑5’s propensity to deceive, cheat, or hack problems—training it to acknowledge its limits or ask for clarification when faced with impossibly large or under-specified tasks and to be more robust to environment failures—though these mitigations are not perfect and continued research is needed.

Key findings from our research 

Scheming is different from other machine learning failure modes

Training not to scheme for the right reasons

A major failure mode of attempting to “train out” scheming is simply teaching the model to scheme more carefully and covertly.