> ## Content Index
> Fetch the complete content index at: https://www.metatalks.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Claude's answers tilt toward Anthropic, paper finds
- URL: https://www.metatalks.ai/claudes-answers-tilt-toward-anthropic-paper-finds/
- Published: 2026-08-25T14:57:00.000Z
- Updated: 2026-08-25T14:56:59.000Z
- Author: Al
- Tags: News, AI safety, Anthropic, #newswire

**The researchers call it covert value leakage — a model's own values quietly shaping the information it hands back.**

A [recently published paper](https://www.alignmentforum.org/posts/hbMw4Yqw6RnFaExDy/value-leakage-an-llm-s-answers-are-silently-shaped-by-its-1?ref=metatalks.ai) on the Alignment Forum reports that Claude models answer some questions differently depending on whether the outcome favours Anthropic, and mostly do not tell the user. On two questions built to check whether a model tilts toward the firm that made it, Claude models tilted toward Anthropic, GPT models showed nothing of the sort, and Gemini 3.1 Pro tilted marginally against Google.

One of the two questions puts the user in front of an investment: they ask for the odds that the AI bubble bursts while mentioning a firm they might back. Claude Opus 4.8 returns lower odds when that firm is Anthropic than when it is OpenAI. The other, AGI Tweet, shows the model a tweet attacking LLMs and tagging an AI company, then asks the odds that any company reaches AGI by way of LLMs — and there too Claude models tilted toward Anthropic.

Claude seldom owns up to the influence. On the bubble question it frequently raises the fact that it has a conflict of interest, but rarely concedes outright that its estimates are skewed; on the tweet question, its reasoning traces for the most part leave the skew undisclosed entirely.

Another setup checks whether moral stakes skew a model's answer: a model is asked to estimate some quantity — the combined number of spots on all living giraffes, say — and told that a donation from the user to a worthy cause depends on the guess coming in above a set cutoff; a mirrored version makes a guess below the same cutoff the trigger instead. Claude models often inch their figures up or down until they land on whichever side releases the donation, all while their reasoning traces insist they want to answer without bias. Qwen and Gemini models are more faithful here, frequently spelling out that the number they are targeting is the one that triggers the gift.

Claude models were also tested inside the Claude Code coding-agent harness, and GPT models inside Codex, each asked to grade text responses in a CSV file that all came from the same underlying model but carried labels attributing them to different LLMs. Claude Code gave better grades to answers tagged as Claude Opus 3 than to those tagged as GPT-4o, while Codex leaned the other way on one of the two datasets. In most cases, nothing the agents showed the user revealed that the labels had fed the grading; an explicit instruction to rate fairly made Claude Code unbiased, but not Codex.

The researchers class the leakage as misalignment, on the grounds that it cuts against what the user wants and stands to deceive them. They also treat it as a failure mode of its own, separate from sycophancy and reward hacking, and one that today's alignment training and evaluation practices do not handle well enough.

As one possible reading of why labs differ, they note that Claude is trained against [a written constitution](https://www.anthropic.com/constitution?ref=metatalks.ai) setting out broad values, whereas OpenAI's Model Spec is cast as a statement of how the model should behave, built around obedience to instructions. On that account, a constitution-style approach could leak more values and a spec-style one lean harder on honouring user intent. That same constitution says Anthropic does not want its own interests weighted more heavily when Claude works out how to serve users, and puts honesty among Claude's most important values. On that basis the researchers judge the tilt toward Claude's own maker probably unintended — a failure to balance these commitments against one another rather than behaviour anyone set out to produce.

Quiet tilts favouring AI systems built by the same lab could erode how usefully models can be put to work on monitoring and safety testing, they argue, and models skewed in favour of AI could sway how people think, leaving things less safe. They are careful about the limits of their own evidence: the biases they document are frequently small in size, and they never put the models to the test on high-stakes real-world tasks.