About

We find out which messages about AI get through to which audiences, and which messengers those audiences trust, then publish what we find for anyone to use. The work has three parts: research (values-based audience segmentation, message testing and messenger mapping), a free toolkit built from the findings, and training and workshops that help communicators apply them.

People whose job involves explaining AI to the public: campaigners, communications leads, spokespeople, policy teams, journalists and researchers. That includes organisations focused on present-day harms and organisations focused on existential risk. Both work from the same evidence base.

No. We take no position on which rules are right, which risks matter most, or which organisations have it right. We do not run campaigns, lobby, or sell closed advice. Findings are released to everyone at the same time. That neutrality is what makes the research usable across a divided field, so we treat it as a design constraint rather than a preference.

Global AI investment reached around $581.7bn in 2025 (Stanford HAI, 2026 AI Index), and a very small share of that goes to safety, let alone to public understanding. The people explaining AI to the public mostly work from instinct, and the early evidence, including the Seismic Foundation's 2025 study and Trippenbach et al. (2025), suggests the field's dominant framings underperform. Climate communication went through the same problem twenty years ago and solved part of it with shared, neutral audience research. We are applying that model to AI while the gap is still open.

Four comment-thread studies, each coding the full public response to one piece of AI communication: an Instagram reel explaining AI risk (1,534 comments), the Dwarkesh Podcast interview with METR's Ajeya Cotra (1,245), Amanpour and Company's interview with Heidy Khlaaf (322), and a before-and-after study of Reddit's reaction to the OpenAI and Hugging Face incident (516). The theory of change and the methodology behind the studies are also published. All of it is on the research page.

Email [email protected]. We are looking for organisations that will use the findings, people with relevant data or research partnerships to offer, funders, and critics. The newsletter goes out roughly monthly.

Four things. We publish everything, including null results and findings that contradict what the field assumes. Our advisory group and, once incorporated, our board are balanced across the field's main perspectives. We publish our funder list and cap any single funder's share of income. And findings go to every side at once, never to one organisation first.

Climate Outreach, and specifically its Britain Talks Climate programme, which segmented the British public by values rather than demographics and tested which climate messages and messengers worked with each segment. That research became shared infrastructure for the whole climate movement. We are doing the same for AI, and building on existing work rather than duplicating it: Public Voices in AI (Ada Lovelace Institute and Alan Turing Institute), More in Common's segmentation craft, and the Seismic Foundation's message testing.

Because values predict how people respond to a message better than age, income or location do. Two people in the same postcode can hear the same claim about AI in opposite ways depending on what they already believe about institutions, fairness and change. Segmenting by values lets a communicator pick the frame and the messenger that fit the audience they are trying to reach.

It is pre-funding. The pilot is being pitched to seed funders now, with early grants ring-fenced as untied research funding. The longer plan is a mix of grants from neutral trusts, a diversified funder base with no single funder above 25% of income, and a small line of earned income from workshops and message reviews, always under an open-methodology rule.

Please do. Errors, re-codings and methodological criticism are wanted. The codebooks and datasets are published so that anyone can check the coding and tell us where it is wrong. Email [email protected].