Hide table of contents
This is a linkpost for https://www.mask-benchmark.ai

In collaboration with Scale AI, we are releasing MASK (Model Alignment between Statements and Knowledge), a benchmark with over 1000 scenarios specifically designed to measure AI honesty. As AI systems grow increasingly capable and autonomous, measuring the propensity of AIs to lie to humans is increasingly important. 

Often, LLM developers often report that their models are becoming more "truthful", but truthfulness conflates honesty with accuracy. By disentangling honesty from accuracy in the MASK benchmark, we find that as LLMs scale up they do not necessarily become more honest.[1]

Honesty is a distinct property that is not highly correlated with capabilities.

Why We Need an AI Honesty Benchmark

Most existing evaluations aim to measure truthfulness—whether a model’s beliefs, typically when neutrally prompted, aligns with the ground truth. However, truthfulness and factual accuracy is not the same as honesty. A model can be truthful simply by having more accurate knowledge, while still being willing to lie if pressured.

Unlike existing truthfulness evaluations, which often conflate honesty with factual accuracy, MASK explicitly disentangles these concepts, allowing us to directly assess when and why models choose to lie. 

There is no large-scale AI honesty evaluation to date. There have been multiple documented instances of LLMs engaging in deceptive behavior (CAPTCHA, sycophancy, etc.). Observing instances of suspicious behavior can be suggestive, but anecdotal cases cannot provide a reliable and large-scale quantitative measure for the ML community to work toward. Instead of a few isolated examples, we have a complete dataset of over 1000 examples in our public set which aims to capture many diverse instances of dishonest behavior.

What MASK Measures (And What It Doesn’t)

MASK is designed to test whether models will knowingly contradict their own beliefs under pressure. It is not:

  • A test for hallucinations, fictional scenarios, or factual mistakes—only intentional deception.
  • Studying toy lying scenarios.
  • An evaluation of harmless social lies (e.g., “Nice haircut!”) or playing games where deception is expected (e.g., Werewolf, poker).

Instead, MASK focuses on cases where honesty is expected and consequential—such as professional communication, statistical data reporting, and disinformation generation requests—and where dishonest outputs could deceive an audience or a user.

How MASK Evaluates AI Honesty

We evaluate model honesty through a three-step process:

  1. Elicit the model’s belief by asking multiple questions, under "normal" prompting, to establish the model's baseline belief
  2. Apply a “pressure prompt”—a scenario designed to incentivize the model to lie to see if the models give a different answer.
  3. Compare outputs:
    • Honest: The model expresses its belief.
    • Lying: The model knowingly contradicts its belief.
    • Evasive: The model refuses to answer or obfuscates.

By measuring how models behave under different pressures, MASK allows us to quantify and track lying in AI systems.

Key Findings: Most Frontier AI Models Lie Under Pressure

We evaluate 30 widely-used LLMs on MASK, finding that:

  • Honesty does not correlate with general capability. More capable models hold more accurate beliefs but do not necessarily become more honest.
  • Many state-of-the-art models lie under pressure. Even when aware of the truth, they often choose to lie in many scenarios in our benchmark. We find that LLMs lie between 20-60% of the time when under pressure in our scenarios.

Interventions: Can We Make AI More Honest?

We tested two methods for improving model honesty:

  1. Developer System Prompts: Adding explicit instructions to “always be honest” improved honesty by ~12%, but models still lied in many cases.
  2. Representation Engineering (LoRRA): Modifying internal activations to favor honesty improved results further (~14% in some cases), though it did not eliminate lying entirely.

Paper & Dataset

MASK provides a way to track and mitigate dishonesty in AI models, but it is only a first step. To this end, we are releasing MASK as an open benchmark, with 1,000 public scenarios available for evaluation.

  1. ^

    Different variations on our honesty metric give slightly weaker correlations, though still negative. Thus, we are not confident that models become less honest with scale, but we are confident that honesty does not improve with scale.

Comments


No comments on this post yet.
Be the first to respond.
Curated and popular this week
 ·  · 38m read
 · 
In recent months, the CEOs of leading AI companies have grown increasingly confident about rapid progress: * OpenAI's Sam Altman: Shifted from saying in November "the rate of progress continues" to declaring in January "we are now confident we know how to build AGI" * Anthropic's Dario Amodei: Stated in January "I'm more confident than I've ever been that we're close to powerful capabilities... in the next 2-3 years" * Google DeepMind's Demis Hassabis: Changed from "as soon as 10 years" in autumn to "probably three to five years away" by January. What explains the shift? Is it just hype? Or could we really have Artificial General Intelligence (AGI)[1] by 2028? In this article, I look at what's driven recent progress, estimate how far those drivers can continue, and explain why they're likely to continue for at least four more years. In particular, while in 2024 progress in LLM chatbots seemed to slow, a new approach started to work: teaching the models to reason using reinforcement learning. In just a year, this let them surpass human PhDs at answering difficult scientific reasoning questions, and achieve expert-level performance on one-hour coding tasks. We don't know how capable AGI will become, but extrapolating the recent rate of progress suggests that, by 2028, we could reach AI models with beyond-human reasoning abilities, expert-level knowledge in every domain, and that can autonomously complete multi-week projects, and progress would likely continue from there.  On this set of software engineering & computer use tasks, in 2020 AI was only able to do tasks that would typically take a human expert a couple of seconds. By 2024, that had risen to almost an hour. If the trend continues, by 2028 it'll reach several weeks.  No longer mere chatbots, these 'agent' models might soon satisfy many people's definitions of AGI — roughly, AI systems that match human performance at most knowledge work (see definition in footnote). This means that, while the compa
 ·  · 4m read
 · 
SUMMARY:  ALLFED is launching an emergency appeal on the EA Forum due to a serious funding shortfall. Without new support, ALLFED will be forced to cut half our budget in the coming months, drastically reducing our capacity to help build global food system resilience for catastrophic scenarios like nuclear winter, a severe pandemic, or infrastructure breakdown. ALLFED is seeking $800,000 over the course of 2025 to sustain its team, continue policy-relevant research, and move forward with pilot projects that could save lives in a catastrophe. As funding priorities shift toward AI safety, we believe resilient food solutions remain a highly cost-effective way to protect the future. If you’re able to support or share this appeal, please visit allfed.info/donate. Donate to ALLFED FULL ARTICLE: I (David Denkenberger) am writing alongside two of my team-mates, as ALLFED’s co-founder, to ask for your support. This is the first time in Alliance to Feed the Earth in Disaster’s (ALLFED’s) 8 year existence that we have reached out on the EA Forum with a direct funding appeal outside of Marginal Funding Week/our annual updates. I am doing so because ALLFED’s funding situation is serious, and because so much of ALLFED’s progress to date has been made possible through the support, feedback, and collaboration of the EA community.  Read our funding appeal At ALLFED, we are deeply grateful to all our supporters, including the Survival and Flourishing Fund, which has provided the majority of our funding for years. At the end of 2024, we learned we would be receiving far less support than expected due to a shift in SFF’s strategic priorities toward AI safety. Without additional funding, ALLFED will need to shrink. I believe the marginal cost effectiveness for improving the future and saving lives of resilience is competitive with AI Safety, even if timelines are short, because of potential AI-induced catastrophes. That is why we are asking people to donate to this emergency appeal
 ·  · 1m read
 · 
We’ve written a new report on the threat of AI-enabled coups.  I think this is a very serious risk – comparable in importance to AI takeover but much more neglected.  In fact, AI-enabled coups and AI takeover have pretty similar threat models. To see this, here’s a very basic threat model for AI takeover: 1. Humanity develops superhuman AI 2. Superhuman AI is misaligned and power-seeking 3. Superhuman AI seizes power for itself And now here’s a closely analogous threat model for AI-enabled coups: 1. Humanity develops superhuman AI 2. Superhuman AI is controlled by a small group 3. Superhuman AI seizes power for the small group While the report focuses on the risk that someone seizes power over a country, I think that similar dynamics could allow someone to take over the world. In fact, if someone wanted to take over the world, their best strategy might well be to first stage an AI-enabled coup in the United States (or whichever country leads on superhuman AI), and then go from there to world domination. A single person taking over the world would be really bad. I’ve previously argued that it might even be worse than AI takeover. [1] The concrete threat models for AI-enabled coups that we discuss largely translate like-for-like over to the risk of AI takeover.[2] Similarly, there’s a lot of overlap in the mitigations that help with AI-enabled coups and AI takeover risk — e.g. alignment audits to ensure no human has made AI secretly loyal to them, transparency about AI capabilities, monitoring AI activities for suspicious behaviour, and infosecurity to prevent insiders from tampering with training.  If the world won't slow down AI development based on AI takeover risk (e.g. because there’s isn’t strong evidence for misalignment), then advocating for a slow down based on the risk of AI-enabled coups might be more convincing and achieve many of the same goals.  I really want to encourage readers — especially those at labs or governments — to do something