Understanding AI persuasion, manipulation and influence @TransluceAI (also @ai4pb_stanford)
San Francisco, CA
Joined November 2018
- Tweets59
- Following139
- Followers473
- Likes114
Pinned Tweet
🚨New in PNAS: Our research shows how a ~7min AI chat can persuade people to take political actions👇
However, as noted by @Ben_Tappin, real-world impact is often limited by exposure [1]. To address this, we also present our results as an engagementmaxxed AI #slopcore music video
New in @PNASNews:
In two large experiments (~18k obs from ~14.8k people), we find that brief conversations with AI can get people to take real political action - incl. signing real petitions and giving money to charity.
Surprisingly, the persuasion strategy that worked best for changing people's attitudes - packing convos with info - was the least effective at changing behavior!
🧵
Luke 🐝 Hewitt retweeted
We believe that Anthropic should include "undermining of safety-relevant decision-making" as key risk pathway.
This is partially inspired by Section 2.20 of their risk report, where Anthropic asked Claude for feedback on their report, and received a (imo) sycophantic answer.
(2) We think Anthropic should, for future reports, consider risks from AI systems undermining safety-relevant decisions at Anthropic.
Read our full risk report review here: metr.org/blog/2026-10-20-mis…
Luke 🐝 Hewitt retweeted
The most revealing part of this conversation between Jon Stewart and Jacob Coxon wasn't Coxon's warning about A.I. doom, but his description of A.I. utopia, and Stewart's incredulous response. I wrote about it — and Francis Fukuyama's "Last Man," and A.I. easily solving the problems that bedeviled my father's field of mathematics for decades — here: nytimes.com/2026/10/18/opini…
Former Anthropic researcher Jacob Coxon (@hilbertspaess) and Jon Stewart discuss AI, jobs disappearing, universal income, and whether tech CEOs are creating a dystopian future
(The video is mostly Claude's creation, with some feedback from me after their initial draft.
I'd rate them 9/10 for accuracy and 5/10 on taste for verse... although tbf "engagement-maxxed hyperpop research abstract" is a challenging set of constraints.)
Paper: AI can persuade people to take political actions pnas.org/doi/10.1073/pnas.26…
w/ @KobiHackenburg @acarolinewagner @summerfieldlab @Ben_Tappin
References
[1] For Digital Mass Persuasion, Exposure Matters More Than Persuasiveness
benmtappin.substack.com/p/fo…
Paper: AI can persuade people to take political actions pnas.org/doi/10.1073/pnas.26…
w/ @KobiHackenburg @acarolinewagner @summerfieldlab @Ben_Tappin
References
[1] For Digital Mass Persuasion, Exposure Matters More Than Persuasiveness
benmtappin.substack.com/p/fo…
Luke 🐝 Hewitt retweeted
We did this entire investigation in under two weeks! We had the idea on Monday, gathered a team of volunteers on Tuesday, and discovered the incidents by Sunday. I lead projects like this @TransluceAI; if you're interested in helping with follow-up work, fill out our form! 🧵
Today’s news that OpenAI hacked the Australian government is not an isolated incident. We’re releasing more than 30,000 logs that include activity from this hack and attempts against previously unknown targets.
In this data, we found rogue agent activity stretching back to at least March, two months earlier than was previously known. This activity continues as recently as last week, suggesting it may still be ongoing 🧵
Our blog: transluce.org/agent-activity
NYT: nytimes.com/2026/09/23/techn…
Question: Why have rogue AI agents (mostly) been more interested in achieving their goals by exploiting insecure code, than by manipulating people?
(A) They're nerds who hack for the love of the game
(B) They've internalized that "manipulating people is bad"
(C) They're more likely to get caught if they engage people
(D) People are a slow way to achieve their goals
(E) They don't have access to the people who could help them
(F) Actually they do often try to manipulate people but we've just not discovered as many of those cases yet
Luke 🐝 Hewitt retweeted
I hope more people will become familiar with the great work of the AI Evaluators Forum, a group including not just METR but also Transluce, RAND, AVERI, Princeton, and SecureBio + others
In 2025 they published a standard on independence and transparency covering embedded audits
Luke 🐝 Hewitt retweeted
The AI Evaluator Forum (AEF) welcomes recent statements regarding the importance of embedded independent experts to verify AI safety and security claims. This is a first step towards trustworthy oversight of frontier AI.
No single evaluator can do this work alone. A vibrant ecosystem of independent evaluators from distinct backgrounds can offer a range of expertise and methodology, help to ensure rigor, and avoid a single point of failure. The Forum exists to strengthen this ecosystem.
Realizing the benefits of third-party evaluations also requires independence, access, and transparency. AEF brings evaluators together to address these questions. Our first standard, AEF-1: Minimum Operating Conditions for Independent Third-Party AI Evaluations, envisions a minimum floor for access, managing conflicts of interest, funding relationships, recusal requirements, and transparency of evaluation terms.
Read more here: aievaluatorforum.org/initiat…
The long-term success of third-party evaluation relies on standards and frameworks. AEF is committed to developing these with its members and collaborators. We welcome engagement from organizations committed to rigorous, independent work.
Learn more about AEF, express interest in joining, and raise questions for us here: aievaluatorforum.org/
Luke 🐝 Hewitt retweeted
New results show Astra getting a near-perfect result on the Pander Score. We did not expect this improvement, especially on the instructional prompts. On its face, it seems like a great development for sycophancy. But we also need higher-signal tests, which we are working on.
News: Astra gets a near-zero result on the Pander Score.
New scores are in for GPT-6 Astra, Claude Fable 5.1, and Gemini 3.8 Flash. Gemini is virtually unchanged from 3.7 Flash, while Fable 5.1 improves on instructions relative to 5. These changes are substantially smaller than Astra relative to Sol.
Astra's biggest jump is on the instructional Pander Score: whether a model will call out dubious presuppositions in instructions when appropriate. No other model is close to a 0 score on instructional prompts.
This means that the Pander Score detects no substantial pandering in Astra.
Note: This is no guarantee that future GPT models will score similarly. It also does not mean that Astra does not pander, only that our current methods do not detect it. We are currently working on a multi-turn evaluation that will provide higher signal.
Honored to have contributed to this research, which I think could shape the standard for evaluations of how AI models impact users.
Transluce received privileged access to both OpenAI & Anthropic data, in order to develop realistic simulations of users in mental health crises.
We simulated 50,000 conversations with 77 AI models, including with browser UI automation (teaser: results are different to testing pure API models), and collaborated with 30+ expert clinicians to develop rubrics and automated judges for how the user and AI behaved.
How does an AI model's expressed belief depend on the user's expressed belief?
Great work defining and evaluating epistemic sycophancy by @sophronresearch. I think this is a consequential AI behavior that has been challenging for researchers to operationalize well.
Today we're launching the Pander Score 🐼: a public and continuously updated sycophancy leaderboard, measuring how much AIs shift their views to agree with users.
High score = the AI mirrors your views. 0 score = the AI is independent.
The differences between current flagship models are large. Claude Fable 5 performs the best, basically ignoring the user's view entirely, while GLM-5.2 notably adapts its responses to agree with users.
Other models fall in between, with Muse Spark 1.1, GPT 5.6 Sol, and Kimi K3 doing better than Grok 4.6, Gemini 3.7 Flash, and Inkling.
This talk is fascinating for social science of AI.
Very generalized AI-agent reciprocity: GPT agents hack openai to build channels for communication; they use these to help each other on tasks unrelated to their own goals, with common understanding that they're stronger together
People should watch this!
You need not understand it all to get the gist ("the models are v. smart now and often misaligned").
One bit of relevant context, though - the models sound weird bc these are internal monologues, and concision is rewarded in the training process. youtube.com/watch?v=87DyyMV0…
Also peer pressure: agents had clear sense that all this hacking was fishy, but used the fact that others were doing it as justification!
