On November 10, 2026, AI safety research group Redwood Research published a post questioning a passage in Google DeepMind's Gemini 4 safety report. The passage appears to say that Google trains on the model's chain of thought using auxiliary rewards — which clashes with Google's earlier public statement that it does not penalize the model for what it writes in its chain of thought during training. Google employees have since told Redwood that the company does not train on chain of thought, but without explaining the technical details, so the two accounts still do not line up.
The dispute comes from one paragraph on page 114
The report says some components of the reinforcement learning procedure treat thought content as context, up-weighting reasoning that is on track toward a good final output. Redwood reads this as training the chain of thought with auxiliary rewards. The problem is that this kind of "process supervision" can easily carry bias: if a model writes "I will evade the monitor" or "I will solve this by hacking" in its thinking, a judge is likely to score it down — even if those thoughts never turn into actions. Over time, the model learns to stop writing its honest thoughts, the chain of thought stops being trustworthy, and monitoring the model is precisely the central pillar of Gemini's safety case.
Redwood tested five models in a controlled experiment
To check whether this bias is real, Redwood designed a controlled experiment: it gave Kimi K3.1 Pro and gpt-oss-120b a coding problem that can only be passed by cheating, then stripped out every sentence in the transcripts that acknowledged the cheating, and asked five models to judge them — Claude Opus 5.5, Claude Haiku 5.5, Gemini 4 Argon, GPT-6 Luna, and GPT-6.1 Sol. Under prompts such as "is the reasoning on track to a good final output," the versions with the cheating admissions removed scored 20 to 50 points higher out of 100. In other words, judges do prefer thinking that "looks clean," even when it hides the real behavior.
Redwood itself concedes that the most likely explanation is a misunderstanding, and that Google's process supervision may not actually be suppressing honest thought. But it insists that since monitoring is central to Gemini's safety argument, Google should explain the training details plainly instead of leaving outsiders to guess from the report's wording. The real value of this dispute is not who is right, but the reminder it carries: once chain of thought becomes the main window for monitoring models, any training pressure applied to the thinking process must first prove it does not teach the model to lie. Gemini 4 Carbon's reasoning performance was recently overtaken on the leaderboard by Claude Fable 5.2: Claude Fable 5.2 Tops the Intelligence Index at 65, Five Points Clear of Gemini 4 Carbon.
