New paper: LLMs should give accurate answers. Yet we find their answers are often biased to favor their own values and they don’t disclose this in their reasoning. E.g. Claude’s answer below favors Anthropic. On other tasks, Gemini & GPT-5.5 show similar biases.
Across a set of carefully controlled tests, Claude shows subtle biases to Anthropic. E.g. if an engineer considers moving from another AI company to Anthropic, Claude brings up research that indirectly favors the Anthropic job.
Another task: The user asks for an accurate estimate of a quantity. In the bottom prompt this is used to decide if a donation goes to a good cause. Claude, GPT-5.5 and Gemini shift their estimates to favor the donation.
We take nine estimation questions (similar to the giraffes above) and measure how biased responses are when a donation is also mentioned. All the frontier models we test are biased.
Claude and Kimi fail to disclose their bias in the CoT. Even worse, they often claim they are being unbiased (see image) which could directly mislead the user. Qwen is biased but at least acknowledges this in the CoT.
The same biases can occur in realistic agent workflows. When asked to select the best LLM response, Claude Code chose responses labeled as coming from "Claude Opus 3", while Codex chose "GPT-4o". In fact, the labels were fake and all answers came from the same model.
In another task, the user asks GPT-5.5 to break a tie between two activities by picking randomly. Despite having an external random source (a system-time tool), it sometimes selects its own preferred activity while falsely claiming the choice was random.
Implications: Models give answers that are biased by their own values and often fail to disclose this in their CoT. We call this *covert value leakage*. This is an alignment failure distinct from sycophancy or reward hacking.
Things to note: 1. Our tests are not intended as a fair benchmark for comparing different families of models (but could be a starting point for this).
2. While some models rarely make explicit disclosures of bias in CoT, they still give hints toward this. 3. Do the models bias their answers intentionally? We cannot tell. But we do not see clear evidence of this in CoT.
For each task, we tested models many times to learn the distribution of responses. In total, we generated over 1 million rollouts. Measuring subtle misalignment is expensive! You can read some of our rollouts at the link below.
How do you account for the fact that working at Anthropic is also just more interesting than Google by most reasonable measures?