# Tech Policy Press transcript: Dave Willner on Moderating with AI [Canonical reading](https://escapemeridian.com/readings/tech-policy-press-transcript-dave-willner-on-moderating-with-ai-0ee98be9) In a Tech Policy Press transcript of his talk, Dave Willner argues that powerful foundation models could change content moderation by carrying out the classification task directly, rather than merely copying human decisions. He says moderation fails not only because policies express contested values, but because applying them consistently to vast amounts of ambiguous, context-dependent content is difficult. Willner describes human moderators’ limited memory, fatigue, trauma and inconsistent judgments, and argues that existing automation inherits some human errors. He says LLMs can read policy documents, apply their rules, explain decisions and make policy changes easier to test. He also argues that reducing harmful frontline moderation work could benefit workers, while acknowledging possible job losses and new uses for censorship and surveillance. His examples of personalized, localized and interactive moderation are possibilities, not proposals he endorses wholesale. Willner warns that LLM-based systems will still make mistakes and have biases; language itself also limits classification. His case rests partly on his account of early model performance, not evidence here of outcomes across platforms. What evidence would show whether LLM-based moderation improves consistency and worker conditions in real-world deployments? Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source Tech Policy Press · Article Original source date: Apr 3, 2024 (2024-04-03T13:00:07.264Z). Date basis: Date stated by the publisher. Added to Escape Meridian: 2026-11-06T05:53:30.978Z. First observed: 2026-11-05T15:41:58.796Z. Retrieved: 2026-11-05T20:08:02.542Z. - [Transcript: Dave Willner on Moderating with AI at the Institute for Rebooting Social Media](https://www.techpolicy.press/transcript-dave-willner-on-moderating-with-ai-at-the-institute-for-rebooting-social-media/) ## Limits The transcript is described as lightly edited and is based on a YouTube recording; it does not provide independent verification of Willner’s performance claims. The account does not establish how LLM-based moderation performs across platforms, content types or deployment conditions. Willner identifies errors, biases and limits arising from language, but the transcript does not quantify their frequency or effects. The transcript discusses potential job losses and new moderation jobs but provides no estimates of their scale. ## Open question What evidence would show whether LLM-based moderation improves consistency and worker conditions in real-world deployments? ## Source-stated Willner argues that content moderation’s central difficulty is consistently classifying content under policies, rather than merely agreeing on values. > The reality is that the sorting task underneath the values, the classification task is the thing that we are very bad at - [Supporting source](https://www.techpolicy.press/transcript-dave-willner-on-moderating-with-ai-at-the-institute-for-rebooting-social-media/) ## Source-stated Willner says human moderators’ memory limits, fatigue and inconsistent judgments make large-scale classification difficult. > Our working memories are really, really small. - [Supporting source](https://www.techpolicy.press/transcript-dave-willner-on-moderating-with-ai-at-the-institute-for-rebooting-social-media/) ## Source-stated Willner argues that LLMs can apply policy text directly to content, rather than only reproducing the outcomes of human decisions. > we have a machine that can directly address the core activity that a human moderator is doing instead of merely producing the result - [Supporting source](https://www.techpolicy.press/transcript-dave-willner-on-moderating-with-ai-at-the-institute-for-rebooting-social-media/) ## Source-stated Willner presents reducing harmful frontline moderation labor as a moral reason to use these systems, while acknowledging possible job losses. > I think it's morally urgent that we figure out how to do this. Even given all of those flaws, having people continue to do this work is bad. - [Supporting source](https://www.techpolicy.press/transcript-dave-willner-on-moderating-with-ai-at-the-institute-for-rebooting-social-media/) ## Source-stated Willner warns that these systems could also enable censorship and surveillance, and says their effects depend on how they are used. > So it is going to be useful for things like censorship and surveillance; it's going to be useful for things like jawboning. - [Supporting source](https://www.techpolicy.press/transcript-dave-willner-on-moderating-with-ai-at-the-institute-for-rebooting-social-media/) ## Source-stated The transcript reports Willner’s account of early GPT-4 experiments reaching over 90% consistency with his decisions; this is not evidence of general real-world performance. > Within a week or two, me and a couple of other engineers were able to get to 90% plus consistency with my decisions - [Supporting source](https://www.techpolicy.press/transcript-dave-willner-on-moderating-with-ai-at-the-institute-for-rebooting-social-media/) ## Source and contributor context Prithvi Iyer · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media Media declared by the original publisher; availability checked separately from the editorial reading. - [Collage depicting human-AI collaboration in content moderation, by Anne Fehres and Luke Conroy & AI4Media / Better Images of AI / CC-BY 4.0 · Tech Policy Press](https://cdn.sanity.io/images/3tzzh18d/production/b8623b70756670a3e03b78b6fee9e70888719e3c-1200x675.png) ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Foundation model A large model trained on broad data that can be adapted to many different tasks. - [Definition reference](https://developers.google.com/machine-learning/glossary#foundation-model) ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Tech Policy Press Tech Policy Press publishes reporting, commentary and interviews on technology policy, online platforms and trust and safety work. - [Explore this context](/entities/tech-policy-press) - [Reference](https://www.techpolicy.press/) ## Editorial provenance Version 1. Publication ID: fce5cf65-777a-4dde-82ce-54ff0cb827c8. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-06T05:53:30.978Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: useful\_perspective. Observation: 6e7d3ef3fd0742e6a26f9f88d437ca155db636d8bd8fcbce48dc7f41170e5cf4. Source text hash: 21f6a7f67a7fdf3c149f6a533bc84c804b37c83ba8d7a0b2ed3beb66c71c6f4a. ## Revision and source-check history ### Version 1 — 2026-11-06T05:53:30.978Z In a Tech Policy Press transcript of his talk, Dave Willner argues that powerful foundation models could change content moderation by carrying out the classification task directly, rather than merely copying human decisions. He says moderation fails not only because policies express contested values, but because applying them consistently to vast amounts of ambiguous, context-dependent content is difficult. Willner describes human moderators’ limited memory, fatigue, trauma and inconsistent judgments, and argues that existing automation inherits some human errors. He says LLMs can read policy documents, apply their rules, explain decisions and make policy changes easier to test. He also argues that reducing harmful frontline moderation work could benefit workers, while acknowledging possible job losses and new uses for censorship and surveillance. His examples of personalized, localized and interactive moderation are possibilities, not proposals he endorses wholesale. Willner warns that LLM-based systems will still make mistakes and have biases; language itself also limits classification. His case rests partly on his account of early model performance, not evidence here of outcomes across platforms. What evidence would show whether LLM-based moderation improves consistency and worker conditions in real-world deployments? ## Original source Tech Policy Press · Article Original source date: Apr 3, 2024 (2024-04-03T13:00:07.264Z). Date basis: Date stated by the publisher. Added to Escape Meridian: 2026-11-06T05:53:30.978Z. First observed: 2026-11-05T15:41:58.796Z. Retrieved: 2026-11-05T20:08:02.542Z. - [Transcript: Dave Willner on Moderating with AI at the Institute for Rebooting Social Media](https://www.techpolicy.press/transcript-dave-willner-on-moderating-with-ai-at-the-institute-for-rebooting-social-media/) ## Limits The transcript is described as lightly edited and is based on a YouTube recording; it does not provide independent verification of Willner’s performance claims. The account does not establish how LLM-based moderation performs across platforms, content types or deployment conditions. Willner identifies errors, biases and limits arising from language, but the transcript does not quantify their frequency or effects. The transcript discusses potential job losses and new moderation jobs but provides no estimates of their scale. ## Open question What evidence would show whether LLM-based moderation improves consistency and worker conditions in real-world deployments? ## Source-stated Willner argues that content moderation’s central difficulty is consistently classifying content under policies, rather than merely agreeing on values. > The reality is that the sorting task underneath the values, the classification task is the thing that we are very bad at - [Supporting source](https://www.techpolicy.press/transcript-dave-willner-on-moderating-with-ai-at-the-institute-for-rebooting-social-media/) ## Source-stated Willner says human moderators’ memory limits, fatigue and inconsistent judgments make large-scale classification difficult. > Our working memories are really, really small. - [Supporting source](https://www.techpolicy.press/transcript-dave-willner-on-moderating-with-ai-at-the-institute-for-rebooting-social-media/) ## Source-stated Willner argues that LLMs can apply policy text directly to content, rather than only reproducing the outcomes of human decisions. > we have a machine that can directly address the core activity that a human moderator is doing instead of merely producing the result - [Supporting source](https://www.techpolicy.press/transcript-dave-willner-on-moderating-with-ai-at-the-institute-for-rebooting-social-media/) ## Source-stated Willner presents reducing harmful frontline moderation labor as a moral reason to use these systems, while acknowledging possible job losses. > I think it's morally urgent that we figure out how to do this. Even given all of those flaws, having people continue to do this work is bad. - [Supporting source](https://www.techpolicy.press/transcript-dave-willner-on-moderating-with-ai-at-the-institute-for-rebooting-social-media/) ## Source-stated Willner warns that these systems could also enable censorship and surveillance, and says their effects depend on how they are used. > So it is going to be useful for things like censorship and surveillance; it's going to be useful for things like jawboning. - [Supporting source](https://www.techpolicy.press/transcript-dave-willner-on-moderating-with-ai-at-the-institute-for-rebooting-social-media/) ## Source-stated The transcript reports Willner’s account of early GPT-4 experiments reaching over 90% consistency with his decisions; this is not evidence of general real-world performance. > Within a week or two, me and a couple of other engineers were able to get to 90% plus consistency with my decisions - [Supporting source](https://www.techpolicy.press/transcript-dave-willner-on-moderating-with-ai-at-the-institute-for-rebooting-social-media/) ## Source and contributor context Prithvi Iyer · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media Media declared by the original publisher; availability checked separately from the editorial reading. - [Collage depicting human-AI collaboration in content moderation, by Anne Fehres and Luke Conroy & AI4Media / Better Images of AI / CC-BY 4.0 · Tech Policy Press](https://cdn.sanity.io/images/3tzzh18d/production/b8623b70756670a3e03b78b6fee9e70888719e3c-1200x675.png) ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Foundation model A large model trained on broad data that can be adapted to many different tasks. - [Definition reference](https://developers.google.com/machine-learning/glossary#foundation-model) ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Tech Policy Press Tech Policy Press publishes reporting, commentary and interviews on technology policy, online platforms and trust and safety work. - [Explore this context](/entities/tech-policy-press) - [Reference](https://www.techpolicy.press/) ## Editorial provenance Version 1. Publication ID: fce5cf65-777a-4dde-82ce-54ff0cb827c8. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-06T05:53:30.978Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: useful\_perspective. Observation: 6e7d3ef3fd0742e6a26f9f88d437ca155db636d8bd8fcbce48dc7f41170e5cf4. Source text hash: 21f6a7f67a7fdf3c149f6a533bc84c804b37c83ba8d7a0b2ed3beb66c71c6f4a. ### passed — 2026-11-06T05:53:30.978Z — anthropic/claude-opus-5.5 Publication: fce5cf65-777a-4dde-82ce-54ff0cb827c8 attribution — pass: The draft attributes the arguments to Dave Willner, as delivered in his talk, and identifies Tech Policy Press as publisher of the lightly edited transcript. The source's claims are kept as Willner's assertions rather than presented as established fact. > What follows is a lightly edited transcript of the presentation, drawn from the YouTube recording of it. summary — pass: The summary faithfully captures Willner's core thesis: classification is a bigger problem than agreeing on values. It covers why humans and existing automation fail, his claim that LLMs do the reading task directly, and his moral case about labor, caveats and abuse risks. It preserves his framing of speculative ideas as possibilities rather than endorsements. LLM is not spelled out, but the surrounding text makes the actor and development understandable. The question follows from the evidence. > A bunch of those ideas are probably bad, to be clear. I'm not saying all of the things I just said should happen. I'm saying they're now not impossible date — pass: The originalPublished value of 2024-04-03 matches the transcript's publication date on the page. The talk itself took place on March 21, 2024, and the draft makes no conflicting date claim. > Prithvi Iyer / Apr 3, 2024 support — pass: Each claim is backed by a matching excerpt. The worker-benefit framing reflects Willner's moral argument and his acknowledgment of job losses. The 90% figure is correctly limited to his own account and not generalized. The limitations do not overstate what the source provides, and the 'explain decisions' wording reflects his functional-explanation argument. > This will mean, I think, job losses, particularly at BPOs, but again, it's not clear to me that job losses are per se bad if the jobs themselves are dangerous, toxic, and not conducive to human thriving. media — pass: The selected media is the verified link to the original Tech Policy Press transcript page. It is the most direct way to access the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/0ee98be9d2d19da7a7db21a7bbcc583091115defd74c9a03189dd6bfb24da6a7/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/tech-policy-press-transcript-dave-willner-on-moderating-with-ai-0ee98be9.json)