What actually changes how people understand AI.
We run values‑segmented research that tests which messages move which audiences, and maps who they already trust to deliver them.
How we research
- 01A values‑segmented survey: grouping people by underlying values and worldview, not demographics alone
- 02A message test: putting real claims about AI in front of each segment to see what actually shifts understanding
- 03A messenger‑trust battery: testing who each segment already trusts to talk about a fast‑moving, difficult topic
Terms are defined in the glossary.
Published research
- 01How Instagram received Nate Soares's caveman-speak explainer of AI risk A coded analysis of 1,534 Instagram comments on a Nate Soares reel. Most commenters accepted the premise, agreement collapsed into resignation, and the one frame shared by believers and sceptics was blaming the companies.
- 02OpenAI and shut? What an independent report did to the marketing-stunt argument A coded analysis of 516 Reddit comments, before and after METR and Redwood Research's independent investigation. Acceptance rose from 51% to 69% of comments taking a position, and the marketing-stunt frame fell from 12% of comments to 9%.
- 03"Please honor commit": how the OpenAI agent-swarm story landed with the audience most likely to believe it A coded analysis of 1,245 YouTube comments on the Dwarkesh Podcast's interview with METR's Ajeya Cotra. The audience accepted the danger, joked about it, praised the messenger, and had no idea what to do next.
- 04"A problem they created": how a news audience received the accountability version of the OpenAI and Hugging Face incident A coded analysis of 322 YouTube comments on Amanpour and Company's interview with Heidy Khlaaf. Most of the audience agreed with her, got angrier as the thread aged, and asked for regulation without being prompted.
- 05"Neither company is acting responsibly": how X received Jacob Coxon's resignation from Anthropic A coded analysis of 2,480 replies to a former Anthropic researcher's resignation thread on X.
- 06"Can't we just unplug it?": The Diary of a CEO AI debate analysed through 26,000 comments A coded analysis of 26,019 comments on the Diary of a CEO's AI risk debate: the full episode and three shorter cuts of it, posted across five platforms. Agreement with the risk case ran from 87.5% under a 21-minute cut on Facebook to 48.5% under a 104-second trailer on YouTube Shorts.
- 07Findings: what four comment-thread studies say about explaining AI risk Nine patterns that hold across at least two of the four published studies, each stated as a hypothesis with the base it was measured on. Rewritten as studies are added.
- 08Methodology How our studies collect, code and check comments, what they can and can't tell you, and what we plan to add next, set out once for every study to reference.
Data explorer.
Every coded comment behind our reception studies, browsable and filterable, comment by comment.
Research library.
Existing research on public attitudes to AI, and on how to frame and communicate AI risk, from both the existential-risk and present-day-harms conversations. Searchable and filterable by topic, geography, method and type.
Pilot study
We're deciding what counts as success before we start, not after. The pilot works if message performance clearly differs between value segments, if other people can repeat the method, and if at least five organisations agree to use the findings in their own communication. A null result counts too, since it would correct the field's assumptions and is still worth publishing.
- 01A values‑segmented dataset on how audiences respond to claims about AI
- 02A first map of which trusted messengers reach which segments
- 03A replicable method, documented for others to run
- 04An openly published findings report
Where the research programme goes next.
Once the pilot proves the method, the plan is to scale it: flagship segmentation studies that go deeper than the first pilot, and a live public‑attitudes tracker that keeps the evidence current as the technology and the public conversation move.
Frequently asked questions
What research have you published?
Four comment-thread studies, a theory of change, and a methodology page describing how the studies are done. The studies are listed on this page with links to each article, its codebook and its coded dataset.
What is a comment-thread study?
We take one piece of AI communication (a podcast episode, a news interview, a viral reel, a Reddit thread) and code every comment under it: whether the commenter accepted the central claim, which frame they used, how they reacted to the format and the messenger, and what emotional register dominated. The result is a detailed read of how one message landed with one audience.
Are these findings about what the public thinks?
No. A comment thread is a self-selected sample of one audience under one artefact, ranked by a platform's algorithm. Nothing in these studies supports a claim about what any population believes. They are hypothesis generation: each produces specific, testable hypotheses for the segmented survey work that Common Signals exists to run.
Who does the coding?
A large language model (Claude), working in batches against a codebook that a human has reviewed and approved before coding starts. A random 40-comment sample from each study is then checked by a human against the codebook and the agreement rate is published per study (37 to 38 out of 40 so far on the three video-platform studies). This is single-pass, single-model coding with no second coder, which is the method's biggest limitation, and the methodology page says so. We would welcome anyone re-coding a sample and telling us where the codes are wrong.
What is the pilot study?
A values-segmented survey of roughly 2,000 US adults, taking about six months from funding. It groups people by values and worldview, tests real claims about AI with each segment to see what shifts understanding, and maps which messengers each segment already trusts. It produces a dataset, a first messenger map, a replicable method and an openly published findings report.
Why start in the United States?
Because the funders aligned with this work fund US-focused research, and the field's centre of gravity is American. The first national flagship study will be British (Britain Talks AI), where the delivery advantage lies: home networks, cheaper fieldwork, the Britain Talks Climate precedent and existing data foundations to build on. A US flagship follows once the UK study has proved the format.
Can I commission research?
Not at the moment. We do not sell closed research, and the pilot is our first priority. If you have data, a research partnership or a question you think the field needs answered, email [email protected] and we will talk.