Stop AI

Stated Theory of Change

Stop AI claims its path to reducing AI risk is: civil disobedience at AI company offices -> arrests and media attention -> criminal trial as public platform (Necessity Defense) -> public awakening to extinction risk -> mass nonviolent movement (3.5% of population) -> permanent global ban on AGI/ASI.

This is modeled on Just Stop Oil and Erica Chenoweth's research on nonviolent resistance. The intellectual architecture, provided primarily by Remmelt Ellen, adds a precautionary principle argument: since experimental proof of AGI safety is impossible before building AGI, and since mathematical modeling of AGI's downstream effects is impossible, and since evolutionary convergence would push AGI toward conditions lethal to humans, the probability of extinction exceeds 99%, and therefore development must be permanently banned.

The secondary demand -- citizens' assemblies to regulate narrow AI -- appeared only in one Kirchner interview and never became part of official messaging, suggesting it was an afterthought rather than a core position.

Revealed Theory of Change

Stop AI's actions reveal a much simpler and less strategic theory of change than the stated one: get attention by doing dramatic things. The evidence:

1. No legislative engagement whatsoever. Despite claiming a three-front strategy including "influencing decision-makers and legislation," there is zero evidence of any contact with any lawmaker. The entire operation was protests and arrests.

2. Protests targeted the same building repeatedly. Nearly every action was at OpenAI's SF office. The Feb 2025 arrest-generating protest occurred on a Saturday when the office was closed. The goal was not to disrupt OpenAI's operations but to generate media coverage and arrest footage.

3. The Necessity Defense strategy was aspirational, not operational. The trial has no reported resolution. The legal theory (that blocking a door is justified to prevent extinction) has no precedent in AI and a poor track record in climate activism.

4. The real output was media attention. Stop AI generated coverage in The Atlantic, Politico, Time, Wired, KQED, ABC7, MIT Technology Review, Business Insider, and many others. For an organization with essentially zero budget, this media footprint is extraordinary. The Altman subpoena was genuinely creative and viral.

5. The escalation was not strategic but emotional. The progression from protests to door-chaining to hunger strikes to mic-grabbing to violence was not a calculated escalation strategy but the trajectory of individuals under increasing psychological pressure from believing the world would end imminently.

The revealed theory of change was: generate media attention through increasingly dramatic actions, which would somehow lead to a mass awakening. The "how" between "people see us on the news" and "3.5% of Americans join the movement" was never specified.

Key Assumptions

1. Nonviolent civil disobedience can achieve a permanent global ban on AGI.

• Evidence for: JSO achieved its first demand (no new UK oil licenses) with <1% of UK population. The Chenoweth research is real and well-cited.
• Evidence against: JSO's demand was a specific policy within an existing regulatory framework. "Permanently ban AGI worldwide" is not comparable -- it requires defining AGI in law, global enforcement, and overcoming enormous economic incentives. No technology remotely this economically valuable has ever been banned. The closest analogues (nuclear weapons, gain-of-function research) are only partially restricted despite decades of effort.
• If wrong: The entire tactical approach is misdirected. Even wildly successful protests would not produce the demanded policy outcome.
• Verdict: This assumption is almost certainly wrong. The analogy to JSO breaks down on specificity and achievability of the demand.

2. Apocalyptic messaging recruits more people than it alienates.

• Evidence for: Stop AI did attract ~200 Signal members and media attention. The extinction framing creates urgency.
• Evidence against: The Kirchner crisis empirically demonstrated the costs. Elmore, Torres, Ellen, and PauseAI all argue the apocalyptic framing drives away potential allies and recruits a skewed population vulnerable to radicalization. Post-crisis, Stop AI's own new leadership abandoned the fear-based approach.
• If wrong: The core recruitment strategy alienates far more people than it attracts, and attracts disproportionately vulnerable individuals.
• Verdict: The evidence strongly suggests apocalyptic messaging is net negative for movement building. Stop AI's own leaders concluded this post-crisis.

3. The absence of institutional infrastructure is an acceptable trade-off for agility and authenticity.

• Evidence for: Stop AI's informal structure allowed rapid action and attracted media attention that more bureaucratic organizations could not.
• Evidence against: The crisis was a direct consequence of no governance, no financial controls, no mental health support, and no succession planning. The "agility" led to an organizational collapse that nearly destroyed the group and damaged the broader movement.
• If wrong: The grassroots authenticity that generates sympathy is inseparable from the organizational dysfunction that enables crises.
• Verdict: The Kirchner crisis is a definitive case study in why direct-action organizations need institutional infrastructure.

4. The court system can be used as a platform for the extinction risk argument.

• Evidence for: The Altman subpoena generated massive attention. Climate activists have occasionally used Necessity Defense trials for publicity.
• Evidence against: No reported trial outcome. Two defendants took plea deals. Kirchner is missing. Reichstadter was in jail. Only Kaufmyn appeared at the last hearing. The trial collapsed as an organizational vehicle.
• If wrong: The legal strategy produced criminal records and financial burden without the intended public platform.
• Verdict: The trial strategy may have succeeded on the awareness dimension (the subpoena went viral) while failing on the legal dimension (no Necessity Defense victory or public testimony).

Strengths

Extraordinary media attention per dollar spent. For an organization with effectively zero budget, Stop AI generated coverage in major national and international outlets. This is genuine impact on awareness, even if the awareness generated was not always positive.

The Altman subpoena was a masterpiece of tactical creativity. It demonstrated that creative, targeted actions can produce outsized attention and actually advance the legal/media strategy.

Filled a genuine gap in the advocacy spectrum. PauseAI does legal protests, ControlAI does insider lobbying. Stop AI was the only organization willing to break the law for the cause. Whether this is net positive for the movement is debatable, but the "radical flank effect" (where extremists make moderates seem reasonable) is a documented phenomenon in social movements.

Authentic grassroots energy. The founders literally lived in homeless shelters for the cause. Whatever else you think of Stop AI, the commitment was genuine and visible. This authenticity generated sympathy and media interest that polished advocacy organizations cannot replicate.

The hunger strike model worked for awareness. Reichstadter's 30-day hunger strike at Anthropic generated international solidarity and significant coverage. As a pure awareness tactic, it was effective.

Weaknesses and Risks

The demand is politically unactionable. A "permanent global ban on AGI" cannot be legislated because AGI cannot be defined in law, enforcement would require unprecedented international coordination, and the economic incentives against it are enormous. The demand has no plausible policy pathway, making all the activism performative rather than instrumental.

Apocalyptic rhetoric predictably produces radicalization. Weiss-Blatt's warning (Dec 2024) and Torres' warnings both preceded the crisis by months to years. The causal chain is clear: if you tell vulnerable people that everyone they love will die in 1-3 years and that the only way to stop it is direct action, some of those people will escalate beyond the boundaries you set. Kirchner's trajectory was predictable, and Reichstadter's "bullet through their head" tweet shows the problem was not limited to one individual.

The organizational model was fundamentally unsafe. Communal living + illegal activity + no governance + no financial controls + no mental health support + apocalyptic ideology is a recipe for cult dynamics and individual crises. Every element was a known risk factor.

The crisis damaged the broader movement. The "unhinged group" framing in Politico and elsewhere gave ammunition to AI industry advocates seeking to discredit all AI safety activism. PauseAI's immediate distancing shows how toxic the association became.

Post-crisis, the organization barely exists. Two remaining leaders, plans to vacate the house, no funding, a tainted brand. Stop AI's future organizational capacity is near zero.

Alienated every potential ally. Disrupted EA Global, heckled Bengio, criticized Yudkowsky, split from PauseAI. The only relationships remaining are with coalition partners (Stop the AI Race, Extinction Rebellion) who do not share Stop AI's specific demands.

Cross-References

vs. PauseAI: Same origin community (PauseAI Discord), opposite approach. PauseAI maintains legal boundaries, policy engagement, and institutional funding. Stop AI rejected all three. The split was over illegal direct action. PauseAI's disavowal after the crisis was immediate and explicit. The contrast is instructive: PauseAI with ~EUR 715K in funding and 600+ volunteers has achieved more legislative engagement (60 MPs, 25 Congressional offices) than Stop AI achieved with arrests and hunger strikes. The "radical flank effect" argument would suggest Stop AI makes PauseAI look reasonable by comparison, but the "contamination effect" (where moderates are tarred by association with extremists) may dominate.

vs. ControlAI: Opposite end of the professionalization spectrum. ControlAI has 15 staff, 279 lawmakers briefed, slick campaigns, and an inside-game strategy. Stop AI has zero lawmakers contacted. ControlAI's critique of Stop AI is implicit but clear: if the problem is political, you solve it through political institutions, not door-chaining. Yet both share the "ban ASI" end goal.

vs. Just Stop Oil (model): JSO achieved its first policy demand. Stop AI has achieved no policy outcomes. The key difference: JSO targeted a specific, achievable regulatory action (no new oil licenses) within an existing framework. Stop AI targeted a maximalist demand (permanent global AGI ban) with no existing framework. The Chenoweth model works for achievable demands, not utopian ones.

vs. Extinction Rebellion (new mentor): XR's experience with burnout, internal crises, and "regenerative culture" is directly relevant to Stop AI's rebuilding. XR has also pivoted from shock tactics to community building, providing a possible model for Stop AI's evolution.

What Would Change This Assessment

Upward: If the criminal trial produced a Necessity Defense victory or high-profile testimony from Altman, the legal strategy would be vindicated as an awareness tool regardless of organizational dysfunction. If the post-crisis leadership successfully rebuilds with proper governance and the "big tent" approach attracts significant new membership, the crisis could become a founding myth rather than a fatal wound. If a major AI incident validates the urgency of the extinction argument, Stop AI's early warnings would gain credibility.

Downward: If Kirchner causes harm to himself or others, the damage to the entire AI safety advocacy movement would be severe. If the trial resulted in significant penalties without any public platform effect. If the remaining leadership cannot rebuild and Stop AI simply dissolves. If evidence emerges that the "large donations" came from sources that create conflicts of interest.

Fundamental question: Does a radical flank help or hurt the AI safety advocacy movement? The environmental movement literature is divided on this. If the radical flank effect dominates, Stop AI served a purpose even in failure -- it made PauseAI and ControlAI seem moderate and reasonable. If the contamination effect dominates, Stop AI's actions made it harder for all AI safety advocates to be taken seriously by policymakers. The answer probably depends on the audience: politicians may be contaminated, while grassroots supporters may be activated.

Self-Critique

What sources should I have checked but didn't?

• The archived podcast episodes (Techdirt author has copies of "Go to Jail to Stop AI" and "Near Midnight in Suicide City") would provide unfiltered Kirchner content.
• The EA Forum posts that could not be fetched due to bot detection might contain community reactions not captured in news coverage.
• The trial docket would clarify what happened on Jan 29, 2026.

Where is this analysis potentially biased?

• I may be too harsh on Stop AI because the Kirchner crisis provides a clean narrative of failure. Organizations that have crises and recover are common; I may be treating the crisis as more definitively damaging than it is.
• Conversely, I may be too sympathetic to the "radical flank" argument. The evidence that Stop AI made PauseAI look more reasonable is speculative; the evidence that Stop AI damaged the broader movement's credibility is concrete (Politico, industry responses).

What would a thoughtful person who disagrees say? "You're judging Stop AI by institutional standards that don't apply to grassroots direct-action movements. The civil rights movement, anti-apartheid movement, and suffragettes all had internal crises, violent episodes, and radical wings that the mainstream found embarrassing. One member's mental health crisis doesn't invalidate the approach. And Stop AI generated more media attention for AI extinction risk with zero dollars than the entire EA-funded safety ecosystem did with hundreds of millions. If you actually believe AI could kill everyone, chaining yourself to a door is not an overreaction -- it's the minimum rational response."

What's my single weakest claim? That Stop AI's demand is "politically unactionable." I presented this as near-certain, but I should acknowledge: a decade ago, "ban all new oil drilling licenses" seemed equally unrealistic to most people, and JSO achieved it. Political environments shift rapidly and unpredictably. A major AI incident could make a ban thinkable overnight.

What information would most change my view? A compelling quantitative analysis showing that Stop AI's media coverage meaningfully shifted public opinion on AI risk regulation. If the coverage translated into measurable attitude change (not just clicks), the awareness theory of change would be vindicated even if the organizational execution was disastrous.