Problem: Public understanding of the longterm risks and benefits of AI is the current bottleneck for AI Safety, and is limited to interest and exposure. The broader impact of AI and automation on society is seldom discussed in mainstream public spaces, apart from tech-utopian or anti-tech circles. Insights and concerns identified through the AI safety and AI governance communities tend to stay within a bubble of intellectuals, with public outreach often being an afterthought.

An informed and engaged public is upstream of meaningful policy shifts. This is especially important in the area of technology, where people tend to form their views within an information landscape that updates quickly and tends toward polarization. Public outreach is needed to ensure that a range of perspectives are available to the public, and that the important discussions about what kind of future we want to build are not limited to experts removed from the concerns and hopes of ordinary people.

Good educational material exists online (AI Digest's interactive demos, Rational Animations, BlueDot Impact courses), but there is no in-person, social experience where the general public and important visitors like policymakers, journalists, and teachers can encounter this material together.

To increase public awareness and deep engagement with these topics, we need a model for a scalable, in-person learning environment designed to inform, inspire, and connect.

Proposal: We are building the AI Safety Museum in two tracks:

1. In-person track (primary): Design a professional, replicable popup exhibition, piloting in Berlin in November 2026, that can travel to conferences, existing museums (as popups or residencies), and civic spaces, building toward a permanent institution.

2. Digital track (supporting): Curate the best existing online interactive AI safety education into coherent online exhibitions, and build new interactive exhibits. This becomes the content pipeline and a standalone public resources.

Delivery of the in-person track is the main goal, with the online community contributing to promotion and the quality, range, and depth of material.

We will achieve this in phases:

1. Scoping a professionally designed exhibition template with interactive exhibits, for example:

   • "Can I pause the AI?": a live system visitors try to interrupt, so they experience corrigibility questions firsthand

   • Trajectory room: where we've been (2012 to today) and forecast ranges for where we're going

   • Mech interp and agent observability: What we know and don't know about how AI works, and effects of various policy proposals on issues like privacy.

   • Automation timelines: visitors estimate "years until X is automated" against expert forecasts

   • Expert disagreement wall: where credentialed experts diverge, and why

2. Supporting research, design sourcing, topic curation, and online presence with a community of volunteers through a digital track. Existing resources will be collected and made available through a website. AI safety and governance experts will be consulted to ensure topical relevance and accuracy, and online exhibits will be prototyped before physical development.

3. Execute and deliver an in-person pilot popup exhibit in Berlin hosted by the Berlin AI Safety Office in November. We've partnered with the Foresight Institute Berlin AI Node and planning further partnerships with SAIGE, Telli and AI.berlin.

4. Provide the template for reproducing the exhibit to other museums and locations globally.

5. Plan further for a larger exhibit in the Futurium museum.

Theory of change: An immersive museum exhibition increases public awareness of safety and governance factors limiting x-risk prevention, supports field building, and provides an intersection point for researchers, policy makers, and the public.

Why Berlin? Berlin has a rich art and museum culture, and is an important center for policy around technology. It is the perfect place to build support and awareness, surface factors in communication and design, and launch the pilot.

Without this funding: The pilot either doesn't happen or with much lower quality. Exhibit development continues on a limited volunteer basis with no professional input or support with design.

Success criteria for the pilot phase:

• Pilot exhibition open in Berlin in November, with 500+ in-person visitors

• 1,000+ high-value people reached through the online exhibitions

• Exhibition template documented for replication (design files, install guide, cost model)

• 2+ committed placements or advanced conversations for 2027 (conference popups or museum residencies), including a scoped next step with Futurium

Budget: The minimal amount ($39,000) delivers the Berlin pilot: the project lead salary of 3 months at ~1 FTE to design, implement, run, and review the project, a short engagement with a professional museum designer, exhibit materials and fabrication, compute for interactive exhibits and curation (LLM API costs), and popup logistics and <5% admin overhead.

The larger budget ($54,000) supports an extensive full engagement with a professional museum designer to ensure the experience, presentation, and communication align with industry standards and the local market landscape. See budget sheet for more information.

Team:

• Justin Shenk (project lead) is an AI safety researcher, BlueDot Impact AGI Strategy facilitator, and former ML startup technical co-founder. He contracted with Redwood on control evals and runs AI Salon Berlin. He previously designed and developed AI computer vision demos at CVPR, ICML and NeurIPS on behalf of Intel.

• Archana Vaidheeswaran (volunteer advisor) is Program Director at Algoverse. Advising on program design and volunteer mobilization.

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

• Exhibit quality doesn't make a splash - experiences on par with existing online demos or that fail to communicate the topic importance won't leave a lasting impression.

• Insufficient marketing/comms, both market research to understand the audience, and advertising to attract visitors.

• Placement beyond the pilot is harder than expected - museums and venues might have competing commitments or interests, limiting scalability.

Our findings would still likely be useful to AI safety community outreach though.

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

None.