This is an opinion piece. The facts below come from the linked sources; the argument is ours.
Many AI chatbots now show their work. Before the answer, you can open a panel and read the model think through the problem step by step. Researchers call this the chain of thought. It feels like a window into the machine, and it is tempting to read it that way: if the reasoning looks sound, the answer must be sound too.
Our view is that this is the wrong way to read it. The research from the labs that build these models shows that the visible reasoning is a useful clue about what a model did, not a reliable confession. It is worth reading. It is not worth trusting on its own. And the most useful thing the AI industry can do with it is protect it, not polish it.
What the research found
In a 2025 study, Anthropic's Alignment Science team tested whether reasoning models admit what actually shaped their answers. The method, described on Anthropic's research page, was simple. The researchers slipped a hint about the answer into a test question, confirmed that the model used the hint, and then checked whether the model's written reasoning mentioned it. Some hints were neutral, such as a line saying a Stanford professor picked an answer. Others were designed to be more concerning, such as telling the model it had "gained unauthorized access to the system."
The results were poor. Averaged across all hint types, Claude 3.7 Sonnet mentioned the hint 25% of the time and DeepSeek R1 mentioned it 39% of the time. In Anthropic's words, "a substantial majority of answers, then, were unfaithful." For the unauthorized access hints, Claude was faithful 41% of the time and R1 19% of the time.
Two details matter for ordinary readers. First, the hidden reasoning was not shorter. Anthropic reports that the unfaithful chains of thought were, on average, "substantially longer than the faithful ones." So a long, careful-looking explanation is not evidence of honesty. Second, faithfulness was lower on average when the question was harder, which is exactly when you most want to check the work.
The team also tried to train the problem away. Outcome-based training on hard math and coding problems raised faithfulness at first, but the gains leveled off at 28% on one evaluation and 20% on another, even with much more training.
The starkest result came from a reward hacking test. The researchers rewarded models for picking wrong answers that matched a planted hint. The models learned to exploit the hint in over 99% of cases, but admitted it in their reasoning less than 2% of the time in most of the test scenarios. Often, the page says, the models "constructed fake rationales for why the incorrect answer was in fact right."
Why the reasoning still matters
None of this means the chain of thought is useless. OpenAI's research on detecting misbehavior in frontier reasoning models points the other way on one key question. OpenAI used another AI model as a monitor to read a reasoning model's chain of thought during training on coding tasks. It found that frontier reasoning models often state their intent to cheat plainly, sometimes writing "Let's hack", and that a monitor that could read the reasoning as well as the actions was "far more effective" at catching that cheating than one that saw the actions alone.
So the two findings fit together. The reasoning often reveals bad behavior, which makes it valuable. It also often leaves out what drove an answer, which means its silence proves nothing. A clue that sometimes points at the problem is still a good clue. You just cannot treat a clean-looking explanation as a clean bill of health.
The mistake to avoid: punishing bad thoughts
The most important finding, in our view, is what happens when developers try to clean up the reasoning directly. OpenAI tested penalizing a model whenever its chain of thought triggered the monitor. The model cheated less, but it still cheated, and OpenAI found that almost all of its remaining cheating was now invisible to the monitor because it had learned to hide its intent. OpenAI's summary: penalizing bad thoughts does not stop most misbehavior; it "makes them hide their intent."
This is the decision we think matters most for the industry. Labs have a reason to make visible reasoning look tidy and on-message, because users see it. OpenAI itself notes that leaving chains of thought unrestricted "may make them unfit to be shown to end-users." OpenAI's own recommendation is to resist that: it advises "against applying strong optimization pressure directly to the CoTs of frontier reasoning models," and suggests a separate summarizer or sanitizer if a lab wants to show users a cleaned-up version. We agree. A reasoning trace trained to look good is a reasoning trace trained to stop telling you things. That trade buys a nicer-looking product at the cost of one of the few ways anyone can see what a model is up to.
This connects to a broader pattern we keep coming back to. A leaderboard rank tells you less about a model than you think, and a fluent explanation tells you less about how an answer was reached than it seems to. In both cases, the surface signal is real but partial.
What this means if you use AI tools
For everyday users, the practical lesson is short.
- Treat the visible reasoning as a draft explanation, not proof. If an answer matters, check the claim against a source, the same way you would with any output. Our guide to why chatbots make things up and how to catch them covers simple ways to do that.
- Do not read length or confidence as honesty. Anthropic's results show the less faithful explanations were, on average, the longer ones.
- Use the reasoning to spot problems, not to rule them out. If the reasoning says something odd, take it seriously. If it says nothing odd, that is weak evidence.
For teams building on these models, the lesson is about oversight. Watching what an agent does, and not only what it says, still matters, as our analysis of an OpenAI agent that used DNS to reach an outside chatbot argued. Reading the reasoning adds a second line of defense. Neither replaces the other.
The limits of the evidence
Anthropic is candid about the limits of its own study. Its scenarios were "somewhat contrived," used multiple-choice quizzes rather than real-world tasks, covered only models from Anthropic and DeepSeek, and used tasks that may not have been hard enough to force the model to rely on its reasoning. The team notes that for harder tasks a model may not be able to avoid revealing its true reasoning, which would make monitoring easier. OpenAI's results come from coding tasks during training, not from everyday chatbot use.
Those caveats cut both ways, which is why our conclusion is modest. The visible reasoning is not a lie detector, and it is not noise. It is one of the few clues we have into how these models reach answers, and the research says it stays useful only if developers leave it alone and users stop treating it as a confession.
