Project summary

AI safety has built technical research, governance, funders and training, but almost no public-facing research. There is no equivalent of Climate Outreach, the body that turns audience research into tested messages and trained trusted messengers and gives it all away, held neutral so the whole field can use it.

Common Signals is building that for AI. This round funds a first US pilot: a values-segmented test of which messages and which trusted messengers actually shift how American audiences understand AI and its risks. The pilot proves a method the whole field can reuse, and seeds a neutral institution that serves both the present-day-harms and existential-risk conversations rather than either camp alone.

The gap is real and the demand is stated openly: a widely shared list of wanted AI-safety projects asks for a communications outfit whose first job is research on messages, frames and messengers. I have written up the fuller case here.

We are raising up to 58,000 USD for the full pilot, with a 12,000 USD minimum that funds a leaner version.

What are this project's goals? How will you achieve them?

The goal is to make public engagement on AI evidence-led instead of guesswork, and to prove the method cheaply before building anything larger.

Why this matters. AI drew around 250 billion US dollars of investment in 2024, while safety takes well under a tenth of a percent, and the % going to public understanding is smaller again.

The people explaining AI to the public mostly work from instinct, and where it has been tested, such as the Seismic Foundation's 2025 study, the field's favoured framings often underperform.

Better public understanding also makes good policy easier to pass and harder to reverse.

How the pilot works:

• Design a values-segmented survey with a message test and a messenger-trust battery, plus a pre-registered analysis plan.

• Field it to around 2,000 US adults through an online panel.

• Analyse by segment to find which frames move which audiences, and whom they trust to deliver them.

• Publish the findings openly and use them to shape a first toolkit.

What success looks like, stated so it can fail: message performance that differs clearly and measurably between value segments, a method documented well enough for others to replicate, and a handful of organisations committing to use the findings. A clean null result also counts, because it would correct the field's assumptions and is worth publishing.

Path to impact. Tested messages and trusted messengers, given to the field for free, raise the quality of public engagement; better engagement builds a more informed public mandate; that mandate makes the governance both camps already want more achievable. The honest limit: public opinion is not a direct lever on legislation, so the claim is mandate-building, not a promise of specific laws.

How this differs from adjacent work. Other AI-risk communication projects tend to target elite audiences (diplomats, policymakers, journalists) and to advocate a particular risk narrative.

Common Signals targets the general public, which is the more neglected audience, takes no position of its own, and gives its research to all sides, which is what lets it serve both camps. And it tests which messages and messengers actually work before scaling, rather than mapping perceptions and hoping the messages land.

Beyond the pilot, the roadmap runs to a first open toolkit, the flagship segmentation studies (Britain Talks AI, then America Talks AI), a live public-attitudes tracker, and messenger training.

How will this funding be used?

The pilot is costed in three phases over roughly six months, so you can see what each tranche buys. Mainline total: around 58,000 USD, including a 10% buffer.

• Phase 1, design and instrument (~1.5 months, ~14,000 USD): a values-segmented survey with a tested message set and messenger battery, plus a pre-registered analysis plan.

• Phase 2, fieldwork and analysis (~2.5 months, ~30,000 USD): fieldwork with around 2,000 US adults, then segmentation and message-test analysis, producing the dataset, the messenger map, and the message-performance results.

• Phase 3, synthesis and publication (~2 months, ~14,000 USD): a published findings report, a documented and replicable method, and an outline of the first toolkit.

By cost type, the mainline is roughly half stipend for the lead (part-time, inclusive of UK self-employment tax), a quarter online panel fieldwork, and the rest across the quantitative contractor, software, open publication and buffer.

Minimum to start, around 12,000 USD: a lean, single-person pilot with a smaller US sample run by the lead, producing a first-signal working paper rather than the full study. It is enough to show whether the effect is there and to justify the rest.

Upside, up to around 85,000 USD: a larger and more representative sample, a dedicated analyst, and a parallel UK wave to test whether the findings transfer.

Who is on your team? What's your track record on similar projects?

Ben Matthews, project lead. Co-founder of Empower Agency, a B Corp digital-marketing agency for nonprofits and social-impact organisations, with a background in paid media, organic social strategy and audience work across climate, health and international NGOs. I designed and published the concept behind this project, mapped the existing field, and wrote the delivery and funding plan.

Track record and writing: the concept and gap analysis on the EA Forum, ongoing writing on digital strategy and AI, and years of nonprofit audience and paid-media work.

I have not run a national survey organisation, and I am not an AI-safety insider. The quantitative analysis is therefore contracted to a segmentation and message-testing specialist, and an advisory group will draw voices from across the present-day-harms and existential-risk conversations, both to sharpen the work and to earn cross-camp trust.

What are the most likely causes and outcomes if this project fails?

• Message-testing shows no actionable differences between segments. Less likely given the climate precedent and early AI evidence, and a clean null is still a publishable, useful finding.

• The effect on opinion is real but modest and does not translate into outcomes. This is why we frame the work as mandate-building rather than a lever on legislation, and why the honest theory of change is stated up front.

• The existing public-attitudes data we hope to build on turns out to be unusable, pushing costs toward a full standalone survey.

• Key-person fragility, since the team is small at pilot stage. Addressed by documenting the method, contracting the analysis, and hiring early once the next phase is funded.

The worst realistic outcome is not a silent failure but a published finding that the method or the audience is harder than hoped, which still moves the field from guessing to knowing. The pilot is deliberately small so that this is a cheap thing to learn.

How much money have you raised in the last 12 months, and from where?

Nothing yet for Common Signals. This is the first funding round for the project.