So, What Do We Do Now?

On 29 October 2026, nineteen people spent an afternoon in San Francisco on one question: what should be in place before AI systems become too capable to simply direct? By five o’clock the room had four projects, and seventeen people had written down what they would give.

29 October 2026, San FranciscoChatham House ruleHosted by Julia Bossmann, Ellen Pao and Skye Lee

Two tables took the question from opposite ends

The room held founders who have sold companies, people who run or fund things, researchers, and people from civil society. Most had not met before lunch. Each table put its worries on the wall, clustered them, voted on what it could move within a year, wrote how-might-we questions, voted again, and drew up a concept sheet for each winner: the problem, who it is for, what is different, what it needs. At four, each table pitched its two projects to the other.

When things go wrong

A model with hidden objectives, a cyberattack on the grid or the banks, panic that outruns the facts. This table ended the day with two playbooks.

A playbook for households and neighborhoods, delivered before anyone needs it

Disaster playbooks exist for earthquakes and storms. Nobody has written one for an AI disruption. This project compiles and adapts what exists and gets it to individuals, neighborhoods, community organizations and churches ahead of time, through the groups people already trust. The sticky note that drew the most votes all day said “know your neighbors now, really,” and the table wanted that inside the playbook.

The favorite delivery idea was a cheap thumb drive mailed to every household, carrying the playbook, offline maps, offline Wikipedia, radio repeater lists and a small offline AI model. Most people would throw it away. A few per neighborhood would keep it, and when the internet goes down, whatever people have on hand ends up leading. Three days later, one participant sent the group a full design for it, costed per household.

A binder that says who in government does what when an AI incident hits

When an earthquake hits San Francisco, the mayor has a physical binder, inches thick, listing every stakeholder and what each one does when. Nobody at the table knew of one for AI incidents. This project writes it: step-by-step instructions per scenario and per agency, on paper so it still works when the internet doesn’t. A think tank drafts example playbooks and agencies adopt them.

The hardest question, tested against a two-day regional water outage, is whether to write for specific scenarios or stay broad. A generic playbook from outsiders earns a polite thank-you and nothing more.

When things go well enough

Well enough that the question becomes whose values the models carry, since what a model is aligned to is currently decided inside a handful of companies. This table ended the day with a civil-society project and a research institute.

Reweaving brings the values nobody is asking about into AI training

Its tagline is bringing IRL to RL: take real-world experience from the people, and the species, underrepresented in the centers of power, and feed it into the reinforcement learning that shapes models. The route in is civil society, because those organizations can bring people to the table. None of us are farmers; it takes a farmers’ association to convene farmers.

The core proposal is a community process for declaring the tech futures people actually want, leading to a national citizens’ assembly on AI ahead of the next presidential election. Many people outside the industry read as anti-AI partly because nobody has asked them what they want AI to do. The first job is a process that participants, the convening organizations and the labs would all trust. Nobody has designed one yet.

A research institute outside the labs for how model values form

One-off experiments have already shown that people outside the labs can shift what models value. This project turns that into a tested method and a body of theory, and gives the few dozen researchers who work on these questions in isolation a home. It does technical research on how models’ motivations form and how to influence them, and it builds the field. It sits outside the labs, stays lab-agnostic and answers to civil society.

The table called the field model psychology, warned against assuming models think like humans, and named the risk: if this works, it is a powerful tool, so one of the institute’s first research questions is who should have access to it. Its verdict on the name “Research Institute” was that it is the most generic title ever. Suggestions welcome.

Seventeen people wrote down what they would give

Thirty-nine cards. Most offers are introductions (to funders, to emergency responders, to a city’s chief technology officer, to organizations that translate animal communication or map what it means to be human), research (user research, scenario design, a synthesis of public opinion on AI, continued work on model motivations), and time. One is money. Every project got offers from both tables.

The room rated it 9.1 out of 10

Fifteen people answered the feedback form, organizers included, since the format was new to all of us. The average rating was 9.1 out of 10 and the net promoter score 67. Everyone who answered wanted the participant list shared, most wanted a working group, and eight signed up for a six-week strategy group. Asked who should be in the room next time, they named more people with technical machine-learning backgrounds, people who work in government, a youth organizer, and a journalist who isn’t there to report.

The full write-up is due by the end of November

It covers the day under the Chatham House rule and will be linked from this page. After that comes a short update on each project at one month and at three months. Nothing is promised beyond what the cards say.

Who ran it

Hosted by Julia Bossmann, Ellen Pao and Skye Lee. Beth Lingard and Adrien Lanusse facilitated the two tables, with Aleksandra Czuchaj. Under the Chatham House rule, this page names the organizers and nobody else.