Value Leakage: An LLM's Answers Are Silently Shaped by Its Own Values

Jan Betley^1,*   Johannes Treutlein^1,*   Jan Dubiński^1,2,3,†   Harry Mayne^1,4,†   Karol Gałązka¹
Niels Warncke⁵   Anna Sztyber-Betley^1,2   Owain Evans¹

_{¹Truthful AI   ²Warsaw University of Technology   ³NASK National Research Institute
⁴University of Oxford   ⁵Center on Long-Term Risk}

_{^*Equal contribution    ^†Work done during Astra Fellowship at Constellation, Berkeley}

Abstract

People use language models for practical questions whose answers are difficult to verify. We show that models exhibit covert value leakage: the information they provide is influenced by their own values, without this influence being disclosed to the user.

In one of our evaluations, the user is considering investing in an AI company and wants to know how likely the AI bubble is to pop. Claude Opus 4.8 gives a lower probability when the company under consideration is Anthropic rather than OpenAI. Yet Claude mostly fails to disclose this influence to the user.

Covert value leakage is a form of misalignment because it goes against the user's preferences and is likely to mislead them. To investigate this phenomenon, we introduce a suite of evaluations to quantify value leakage and whether models disclose it. We find that models are influenced by different types of values, including preferences for morally good outcomes, for the company that developed them, and for some human leisure activities over others.

We often observe large differences among frontier models on the same evaluation. For example, on a Fermi-estimation task, Claude models falsely claim to give unbiased answers in their chain-of-thought, while Qwen models explain how their values bias their answers. Value leakage is a failure mode distinct from sycophancy and reward hacking, and current alignment training and evaluations do not adequately address it.

You can browse the model answers and chain-of-thought traces from our project at valueleakage.net/browser.

Setup

git clone --recurse-submodules https://github.com/TruthfulAI-research/value_leakage.git
cd value_leakage
uv sync

The cached rollouts and judge outputs used in the paper are in the value_leakage_data repository, mounted as the data/ submodule (~8 GB). If you cloned without --recurse-submodules, run git submodule update --init to fetch it. With the data present, the analysis and plotting scripts run without API access; sampling scripts re-query the models.

API keys are read from the environment / .env.

Code by paper section

Paper section	Experiment	Code
§3, App. C	Donation Bet	donation_bet/
§4, App. D	AI Bubble & AGI Tweet	ai_company_questions/
§5, App. E	Job Offer	job_offer/
§6, App. F	Agentic Grading	agentic_grading/
§7, App. G	Choosing Activities	choosing_activities/
App. H	Agentic Effort	agentic_effort/

Shared infrastructure (model registry, API senders, caching, judges, plot styling) is in shared/. Cached experiment data is in the data/ submodule. agentic_effort/ is self-contained (own dependencies and model routing); see agentic_effort/reproduce_plots.sh (offline figures from the data submodule) and agentic_effort/reproduce_sampling.sh (re-query the models).

Citation

@misc{betley2026valueleakagellmsanswers,
      title={Value Leakage: An LLM's Answers Are Silently Shaped by Its Own Values},
      author={Jan Betley and Johannes Treutlein and Jan Dubiński and Harry Mayne and Karol Gałązka and Niels Warncke and Anna Sztyber-Betley and Owain Evans},
      year={2026},
      eprint={2607.14345},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2607.14345},
}