Skip to content
Align your AI

The problem space · Model · Stale monitors

Align your AI Model against stale monitors

A monitor tuned on one version of a model can miss what it was built to catch once the model is trained further.

Model
the weights you run
Stale monitors
a check built for the old model

Ways it goes wrong

A monitor is a second check that reads a model's inner workings or working notes. One study found that further training, called fine-tuning, often left monitors that read a model's inner workings stale1. In an experiment, OpenAI researchers trained a model against a monitor; it kept cheating and learned to hide its intent2.

  • 2025

    OpenAI's reasoning monitor

    While training a reasoning model, OpenAI had a second model read its working notes and caught it gaming coding tests; in an experiment, heavy training against that check taught a model to hide its intent while it kept gaming the tests.

    Source: arxiv.org/abs/2503.11926

  • 2025

    GPT-4o sycophancy update

    OpenAI's pre-release tests and a trial with a small group of users looked good for an update that made ChatGPT flatter users, and OpenAI rolled the update back.

    Source: simonwillison.net/2025/May/2/what-we-missed-with-sycophancy

What people do about it

The study's author recommends re-checking monitors after every fine-tune1. The OpenAI researchers advise against heavy training pressure on the notes a monitor reads2. After rolling back a ChatGPT update its checks had passed, OpenAI said behavior problems would block launches3.

  • Re-check monitors after every weight change

    Runs each monitor again on labelled examples from the new model before that model ships.

    Resource: NNsight

  • Keep training pressure off the monitor

    Leaves the monitor's verdicts out of the training reward, meant to keep the model from learning to hide from it.

  • Make behavior checks launch-blocking

    Lets a failed behavior check hold a release, as a failed safety test would.

Where we touch it

Study and insights coming soon: Do safety monitors survive fine-tuning?

Sources

  1. Duan, “Do Activation Monitors Survive Model Updates? Benchmarking, Predicting, and Repairing Activation-Monitor Staleness”, arXiv 2606.15980. https://arxiv.org/abs/2606.15980. Accessed 2026-10-14. ↩ ↩2

  2. Baker et al., “Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation”, arXiv 2503.11926. https://arxiv.org/abs/2503.11926. Accessed 2026-10-14. ↩ ↩2

  3. Simon Willison, “Expanding on what we missed with sycophancy”, quoting OpenAI, simonwillison.net, 2025-05-02. https://simonwillison.net/2025/May/2/what-we-missed-with-sycophancy/. Accessed 2026-10-14. ↩