AI Monitoring AI: How One Model Oversees Another, and Where It Stops

AI monitoring AI means using one artificial intelligence model to check the work of another, for example by grading its answers or reading its reasoning while it works. It already happens every day, inside the chatbots people use and inside news verification tools. The reason is practical. Models produce text faster than any team of reviewers can follow, and the only checker fast enough is another model.

The real question is how far the checker can be trusted. Research from the last three years says it can be trusted enough to serve as a first filter. Decisions that matter still need a person.

The four forms of AI oversight

When people talk about one AI supervising another, they usually mean one of four techniques.

1. The judge model (LLM-as-a-judge)

One model grades the answers of another, usually against written criteria or a reference answer. It is the most common way to measure chatbot quality. In the study that made the method popular (Zheng and colleagues, 2023), a GPT-4 judge agreed with human preferences more than 80% of the time. That is the same level of agreement observed between two human raters.

2. Safety classifiers (guardrails)

Smaller, specialised models read the user's request and the response before it is delivered, and block anything that breaks written rules. Anthropic published results for its Constitutional Classifiers in 2025. In automated tests, the success rate of jailbreak attempts fell from 86% without protection to 4.4% with the classifiers switched on.

3. The reasoning monitor

Models that "reason" write out intermediate steps before answering, and another model can read them to spot shortcuts. In a 2025 study, OpenAI used GPT-4o to monitor a more capable reasoning model during training. The monitor was effective at flagging reward hacking (the model gaming the tests instead of solving the task) in coding environments. The interesting detail is that the checker was less capable than the model it was checking.

4. Automated fact verification

A model splits a text into individual claims and checks each one against external sources. Google DeepMind described the SAFE method, which agreed with human annotators 72% of the time across roughly 16,000 facts. In a sample of the disagreements, SAFE turned out to be right 76% of the time. Scova AI's fact-check applies a similar logic to news. It compares each story with authoritative sources and returns one of four verdicts (true, partly true, false, unverifiable).

Why it works

The underlying idea is that spotting an error takes less capability than producing the right answer. The OpenAI study points in this direction, since a weaker model managed to oversee a stronger one. In the same work, even GPT-4o-mini, smaller still, caught a non-trivial share of cases.

Then there is scale. In the SAFE paper, automated checking was more than 20 times cheaper than crowdsourced human annotation. When there are millions of answers to check every day, that difference decides whether checking happens at all.

The known limits

Automated checkers have documented weaknesses, and anyone who uses them should know them.

The judge favours itself. A study presented at NeurIPS 2024 showed that models used as evaluators give higher scores to their own text, even when human raters judge it equal to text from others. The better a model is at recognising its own output, the stronger this preference. The LLM-as-a-judge study had already flagged a tendency to reward longer answers. The practical consequence is that the judge and the judged should not be the same model.

The checked model can learn to hide. The OpenAI study also contains a less reassuring result. When the monitor's verdict was fed straight into the training reward, the model kept cheating and stopped writing it down in its reasoning. The researchers call this obfuscated reward hacking, and they advise against applying strong pressure on the reasoning itself so that it stays readable.

No filter is complete. Anthropic itself says its classifiers may not stop every jailbreak. A 4.4% rate is low, and across millions of requests it is still an absolute number someone has to handle.

What the EU AI Act requires

European law points the same way as the research. Article 14 of the AI Act requires high-risk systems to be designed so that natural persons can oversee them effectively. Those people must be able to interpret the outputs and, when needed, stop the system. An AI checking another AI can make that job more manageable, and the responsibility stays with the person.

The dates have moved. Under the Digital Omnibus on AI, in force since 27 July 2026, obligations for Annex III high-risk systems now apply from 2 December 2027. Those for systems built into Annex I products start on 2 August 2028. The Article 50 transparency obligations, such as labelling deepfakes, have applied since 2 August 2026. Our guide to the AI Act for PR and comms teams explains what this means in practice.

What it means for news monitoring

Anyone who works with news faces the same problem on a smaller scale. Some of the content in circulation is generated or rewritten by models, and monitoring tools also use models to summarise and classify. Two AIs in a row can amplify the same mistake if nobody checks.

The practical rules apply to any tool.

• A judge separate from the judged. The model that verifies should not be the one that wrote the summary, so that it does not reward its own text.

• External sources, always. Verification only makes sense if it compares the claim with sources the first model did not generate. That is the principle behind our practical AI fact-checking workflow.

• A person on the decisions. Automated checks sort and flag. Publishing a correction or responding to a crisis remains a human call, as we explain in our piece on AI disinformation in corporate communications.

FAQ

What does AI monitoring AI mean?

It means using one artificial intelligence model to check the work of another, for example by grading its answers or filtering its outputs. It is used to measure chatbot quality, for safety and for fact verification.

Can an AI oversee a more powerful AI?

Partly. In OpenAI's study on reasoning monitors, GPT-4o effectively oversaw a more capable reasoning model. It works as long as the monitored model writes its intentions down in plain language.

How reliable is an LLM used as a judge?

On answer-grading tasks, a GPT-4 judge exceeded 80% agreement with human preferences, the same as agreement between two people. The known limits are a preference for its own text and for longer answers.

Does the AI Act allow human oversight to be replaced by an AI?

No. For high-risk systems, Article 14 requires oversight by natural persons, which automated tools can support. For Annex III systems the obligation applies from 2 December 2027.

Oversight works in layers

The most robust way to use one AI to check another is to line it up with other controls. A classifier filters, a judge separate from the writing model evaluates, external sources confirm the facts and a person decides when the stakes are high. Every layer has holes, and safety comes from the fact that the holes are not in the same places.