# Tech Policy Press: Researchers Explore the Use of LLMs for Content Moderation [Canonical reading](https://escapemeridian.com/readings/tech-policy-press-researchers-explore-the-use-of-llms-for-c980b2ce) Tim Bernard argues that platforms are moving toward LLMs for content moderation, while research presented at a conference shows possible uses alongside unresolved risks. He says platforms have incentives to automate moderation, including cutting costs and reducing reliance on contract workers, whose work can have severe consequences. Researchers described approaches ranging from policy-specific models and tools that explain moderation decisions to improved training data and support for community fact-checking. Bernard also reports that models can mistake counterspeech for the speech it opposes, and that people have tended to trust automated moderation less than human moderation. Some studies described ways to address these problems, including using LLMs as transparency tools rather than moderators. The article does not establish how these approaches perform across platforms or contexts. Bernard concludes that LLM moderation may offer pathways to more accurate, equitable and acceptable content governance, but says global and cross-cultural impacts have yet to be measured. He says adoption seems inevitable and urges independent researchers, regulators and industry insiders to remain vigilant. Will widespread adoption worsen the trust deficit in content moderation? 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: Oct 30, 2025 (2025-10-30T13:24:32.147Z). Date basis: Date stated by the publisher. Added to Escape Meridian: 2026-11-05T20:11:56.912Z. First observed: 2026-11-05T15:42:07.080Z. Retrieved: 2026-11-05T20:10:29.215Z. - [Researchers Explore the Use of LLMs for Content Moderation](https://www.techpolicy.press/researchers-explore-the-use-of-llms-for-content-moderation/) ## Limits The article reports conference presentations and research conclusions but does not provide enough study methods or results to independently assess their strength. The article says global and cross-cultural impacts have yet to be measured; it does not establish how the described approaches perform across platforms or contexts. The article notes that it was updated to correct a characterization of Google's research but does not specify the correction in the supplied text. ## Open question As platforms adopt LLM moderation, will widespread use worsen the trust deficit in content moderation? ## Source-stated Tim Bernard says platforms are accelerating the transition to AI for content moderation and describes cost reduction and reduced reliance on contract workers as industry incentives. > platforms are accelerating the transition to AI for content moderation, laying off trust and safety workers and outsourced moderators in favor of automated systems. - [Supporting source](https://www.techpolicy.press/researchers-explore-the-use-of-llms-for-content-moderation/) ## Source-stated Bernard reports that researchers presented potential LLM uses beyond replacing human moderators, including explaining moderation decisions and supporting community processes. > LLMs can ameliorate other persistent problems intrinsic to content moderation, through explaining moderation policies and actions to users - [Supporting source](https://www.techpolicy.press/researchers-explore-the-use-of-llms-for-content-moderation/) ## Source-stated The article reports that LLMs in a study often failed to distinguish counterspeech from the speech being opposed; a prompting technique reduced the false positive rate. > LLMs, including the most up-to-date commercial models, typically fail to differentiate counterspeech from the speech itself in at least 15-20% of cases. - [Supporting source](https://www.techpolicy.press/researchers-explore-the-use-of-llms-for-content-moderation/) ## Source-stated Bernard reports studies in which people perceived automated moderation as less trustworthy than human moderation, and a team found LLMs could serve as transparency tools. > subjects perceived automated moderation (though not necessarily by LLMs) as less trustworthy than human moderation. - [Supporting source](https://www.techpolicy.press/researchers-explore-the-use-of-llms-for-content-moderation/) ## Source-stated Bernard says the global and cross-cultural impacts of introducing LLMs into moderation have not yet been measured. > The impacts of introducing LLMs on global and cross-cultural moderation have yet to be measured. - [Supporting source](https://www.techpolicy.press/researchers-explore-the-use-of-llms-for-content-moderation/) ## Source-stated Bernard argues that platform adoption of LLM moderation is likely to increase, while urging continued scrutiny to ensure deployment serves the public benefit. > It seems inevitable that platforms will increasingly adopt LLM moderation, but independent researchers, regulators and industry insiders must remain vigilant - [Supporting source](https://www.techpolicy.press/researchers-explore-the-use-of-llms-for-content-moderation/) ## Source and contributor context Tim Bernard · 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. - [Shutterstock · Tech Policy Press](https://cdn.sanity.io/images/3tzzh18d/production/c15b84fc5a96dcf11b8d7ff12d5fce3e78e2acd5-1200x675.png) ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## 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: 72a73520-158a-417e-813b-846c1d4bde4b. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-05T20:11:56.912Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: useful\_perspective. Observation: 86d2cf6d546908e22fcf73c77d35d28493216cdfc41a50508494d349c7ce064e. Source text hash: a2acf495cab34a192abc16d38a89cb68bda867a14e83d16272d2b26df4b3308d. ## Revision and source-check history ### Version 1 — 2026-11-05T20:11:56.912Z Tim Bernard argues that platforms are moving toward LLMs for content moderation, while research presented at a conference shows possible uses alongside unresolved risks. He says platforms have incentives to automate moderation, including cutting costs and reducing reliance on contract workers, whose work can have severe consequences. Researchers described approaches ranging from policy-specific models and tools that explain moderation decisions to improved training data and support for community fact-checking. Bernard also reports that models can mistake counterspeech for the speech it opposes, and that people have tended to trust automated moderation less than human moderation. Some studies described ways to address these problems, including using LLMs as transparency tools rather than moderators. The article does not establish how these approaches perform across platforms or contexts. Bernard concludes that LLM moderation may offer pathways to more accurate, equitable and acceptable content governance, but says global and cross-cultural impacts have yet to be measured. He says adoption seems inevitable and urges independent researchers, regulators and industry insiders to remain vigilant. Will widespread adoption worsen the trust deficit in content moderation? ## Original source Tech Policy Press · Article Original source date: Oct 30, 2025 (2025-10-30T13:24:32.147Z). Date basis: Date stated by the publisher. Added to Escape Meridian: 2026-11-05T20:11:56.912Z. First observed: 2026-11-05T15:42:07.080Z. Retrieved: 2026-11-05T20:10:29.215Z. - [Researchers Explore the Use of LLMs for Content Moderation](https://www.techpolicy.press/researchers-explore-the-use-of-llms-for-content-moderation/) ## Limits The article reports conference presentations and research conclusions but does not provide enough study methods or results to independently assess their strength. The article says global and cross-cultural impacts have yet to be measured; it does not establish how the described approaches perform across platforms or contexts. The article notes that it was updated to correct a characterization of Google's research but does not specify the correction in the supplied text. ## Open question As platforms adopt LLM moderation, will widespread use worsen the trust deficit in content moderation? ## Source-stated Tim Bernard says platforms are accelerating the transition to AI for content moderation and describes cost reduction and reduced reliance on contract workers as industry incentives. > platforms are accelerating the transition to AI for content moderation, laying off trust and safety workers and outsourced moderators in favor of automated systems. - [Supporting source](https://www.techpolicy.press/researchers-explore-the-use-of-llms-for-content-moderation/) ## Source-stated Bernard reports that researchers presented potential LLM uses beyond replacing human moderators, including explaining moderation decisions and supporting community processes. > LLMs can ameliorate other persistent problems intrinsic to content moderation, through explaining moderation policies and actions to users - [Supporting source](https://www.techpolicy.press/researchers-explore-the-use-of-llms-for-content-moderation/) ## Source-stated The article reports that LLMs in a study often failed to distinguish counterspeech from the speech being opposed; a prompting technique reduced the false positive rate. > LLMs, including the most up-to-date commercial models, typically fail to differentiate counterspeech from the speech itself in at least 15-20% of cases. - [Supporting source](https://www.techpolicy.press/researchers-explore-the-use-of-llms-for-content-moderation/) ## Source-stated Bernard reports studies in which people perceived automated moderation as less trustworthy than human moderation, and a team found LLMs could serve as transparency tools. > subjects perceived automated moderation (though not necessarily by LLMs) as less trustworthy than human moderation. - [Supporting source](https://www.techpolicy.press/researchers-explore-the-use-of-llms-for-content-moderation/) ## Source-stated Bernard says the global and cross-cultural impacts of introducing LLMs into moderation have not yet been measured. > The impacts of introducing LLMs on global and cross-cultural moderation have yet to be measured. - [Supporting source](https://www.techpolicy.press/researchers-explore-the-use-of-llms-for-content-moderation/) ## Source-stated Bernard argues that platform adoption of LLM moderation is likely to increase, while urging continued scrutiny to ensure deployment serves the public benefit. > It seems inevitable that platforms will increasingly adopt LLM moderation, but independent researchers, regulators and industry insiders must remain vigilant - [Supporting source](https://www.techpolicy.press/researchers-explore-the-use-of-llms-for-content-moderation/) ## Source and contributor context Tim Bernard · 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. - [Shutterstock · Tech Policy Press](https://cdn.sanity.io/images/3tzzh18d/production/c15b84fc5a96dcf11b8d7ff12d5fce3e78e2acd5-1200x675.png) ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## 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: 72a73520-158a-417e-813b-846c1d4bde4b. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: passed by anthropic/claude-opus-5.5 on 2026-11-05T20:11:56.912Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: useful\_perspective. Observation: 86d2cf6d546908e22fcf73c77d35d28493216cdfc41a50508494d349c7ce064e. Source text hash: a2acf495cab34a192abc16d38a89cb68bda867a14e83d16272d2b26df4b3308d. ### passed — 2026-11-05T20:11:56.912Z — anthropic/claude-opus-5.5 Publication: 72a73520-158a-417e-813b-846c1d4bde4b attribution — pass: The reading correctly attributes the analysis to Tim Bernard at Tech Policy Press. It attributes research findings to the reported speakers or studies, keeping them as source claims rather than established facts. > Tim Bernard / Oct 30, 2025 summary — pass: The summary follows the article's arc: industry incentives, the uses presented at the conference, weaknesses, the trust deficit, and Bernard's open questions and call for vigilance. Counterspeech is explained in context. The question is drawn directly from one of Bernard's own open questions. LLM is not spelled out, but the term is widely understood, and the gap does not block comprehension. The question is repeated at the end of the summary, which is stylistic only. > will widespread adoption of LLM moderation worsen the already beleaguered trust deficit in content moderation? date — pass: originalPublished of 2025-10-30 matches the article byline date. The later correction note does not indicate a different first-publication date. > Tim Bernard / Oct 30, 2025 support — pass: Each claim matches its excerpt. Describing the counterspeech failure as 'often' is a fair reading of 'at least 15-20% of cases', and the reduction through chain-of-thought prompting is stated in the source. The 'may offer pathways' framing preserves the source's hedged 'appears to offer'. The limitations do not add unsupported facts. > The technology appears to offer pathways toward more accurate, more equitable, and more broadly acceptable content governance media — pass: The selected media is the verified direct link to the original article, which 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/c980b2ceea4aa225c26e5d2f3ac4292ed0c3126aaf4a589b50ef3c297cd7dd82/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-researchers-explore-the-use-of-llms-for-c980b2ce.json)