Encode (formerly Encode Justice)

Stated Theory of Change

Encode's stated theory runs: AI poses both near-term harms (bias, deepfakes, surveillance) and catastrophic risks (bioweapons, autonomous systems, loss of human control). Young people bear the greatest stake in these outcomes and therefore have unique moral authority to demand action. By mobilizing youth voices and building bipartisan coalitions, Encode can create political will for state-level AI safety legislation, which then pressures national standards through the "Brussels effect" -- companies adopt the strictest jurisdiction's requirements globally rather than fragmenting their products.

The mechanism is advocacy --> legislation --> enforcement --> safer AI development. The organization explicitly rejects both the "pause everything" framing (they are "pro-AI, pro-innovation") and the "AI can self-regulate" framing (they champion mandatory requirements with legal consequences).

Revealed Theory of Change

Encode's actions largely track their stated theory, with a few important reveals:

The youth movement is the legitimacy base, not the operational engine. The real policy work -- drafting SB 1047, testifying before legislators, filing amicus briefs, negotiating bill amendments -- is done by a small team of professional policy staff (Calvin, Gandhi, Billen, Christian) with serious credentials (Stanford JD, CAIS Action Fund, Senate Judiciary). The ~1,000 student members provide the narrative power ("young people demand action") and some local organizing, but the strategic decisions and policy expertise come from the professional team. This is not a criticism -- it is arguably the right structure -- but it means the "youth-led" branding overstates member agency.

They pick battles with overwhelming public support. Gandhi's "moral high ground" framing is revealing. Their biggest wins (TAKE IT DOWN Act on deepfake porn, AI weapons guardrails, child safety) are issues where public opinion is 80%+ in favor. Their more controversial work (SB 1047 on frontier model liability) got vetoed. The strategic lesson they seem to have absorbed: fight on terrain where "who could oppose this?" applies, and use those wins to build institutional credibility for harder fights later.

The rebrand signals a shift from identity politics to technocratic bipartisanship. Removing "Justice" from the name, scrubbing blog posts about ICE and racism, removing the Cori Bush internship from LinkedIn, and adopting "pro-AI, pro-innovation" language all point in the same direction: Encode is deliberately distancing itself from progressive identity politics to maximize bipartisan appeal. The TAKE IT DOWN Act with Ted Cruz and the Utah work with a Republican freshman are the proof of concept.

The CAIS/longtermist connection is real but understated. Nathan Calvin came directly from CAIS Action Fund. Claire Larkin came from Institute for Progress. Encode receives funding from FLI and SFF. They collaborate with Secure AI Project (Nick Beckstead, ex-Open Phil). The AI 2030 platform was signed by Daniel Kokotajlo. This means Encode is a bridge between the EA/longtermist policy network and the progressive advocacy world -- but they do not advertise this connection, probably because "funded by the same people as everyone else in EA" would undermine their "independent youth movement" brand.

Key Assumptions

1. State-level legislation can meaningfully constrain frontier AI development.

• Evidence for: SB 53 became law and applies to companies with >$500M revenue. Companies (even Meta, Google, OpenAI) indicated willingness to work within its framework rather than exit California. The Prop 12 pork precedent (upheld by SCOTUS) confirms states can regulate products sold within their borders regardless of where they are produced.
• Evidence against: SB 1047 was vetoed. Trump's executive order attempted to preempt all state AI laws. If frontier models become cheap to train, state-level regulation targeting large-compute models becomes irrelevant. Calvin himself acknowledges: "if it's the case that you have really dramatic algorithmic improvement such that you can train capabilities really cheaply, I just don't think any state law is really going to do the trick."
• If wrong: Encode's legislative wins become symbolic rather than substantive. They would need to pivot to federal advocacy or international coordination.

2. Youth voices provide unique political leverage on AI issues.

• Evidence for: Revanur's personal narrative (15-year-old founds movement) generates enormous media coverage (TIME, CNN, BBC, WEF Davos, 60 Minutes). Politicians find it harder to dismiss youth advocates than think-tank lobbyists. The "moral high ground" on child safety issues is real.
• Evidence against: The actual policy work is done by professional staff, not youth. The membership base's influence on legislation is unclear. Youth movements historically struggle with sustainability as founders age out.
• If wrong: Encode would function essentially the same as any other small DC policy shop, just with better PR. The "youth" angle would be marketing rather than mechanism.

3. Bipartisan framing can survive hyper-partisan environments.

• Evidence for: TAKE IT DOWN Act, Utah HB 286, and NDAA provisions all demonstrate genuine cross-aisle achievements. AI policy may be one of the rare issues where a bipartisan window exists (both parties have factions that fear corporate AI power).
• Evidence against: The Washington Examiner investigation frames Encode as a progressive front regardless of bipartisan achievements. Trump's attempted preemption of state AI laws shows the executive branch actively opposes state-level regulation. As AI becomes more politicized, the bipartisan window may close.
• If wrong: Encode gets coded as a "Democratic" organization and loses access to Republican legislators who might otherwise be allies.

4. Transparency and liability requirements incentivize genuine safety, not just compliance theater.

• Evidence for: SB 53's whistleblower protections create a channel for genuine concerns. Calvin's argument that "reasonable care" standards from tort law already apply is legally sound. The TAKE IT DOWN Act's platform accountability provisions create enforceable obligations.
• Evidence against: Companies routinely game compliance regimes (publish reports that say nothing, conduct safety testing that doesn't challenge their release decisions). Without an independent regulator with technical capacity, transparency requirements may produce voluminous disclosures of low quality.
• If wrong: Encode's legislative wins produce paperwork without safety improvement.

Strengths

Unusual legislative track record for its size. A 7-person org with a $500K-$1M budget that has co-drafted landmark state legislation, passed federal law, filed amicus briefs against OpenAI, and influenced NDAA provisions is remarkably productive. The cost-effectiveness (policy impact per dollar) appears high.

Genuine bipartisan capacity. Working with Ted Cruz on TAKE IT DOWN and with a Utah Republican on HB 286 while being funded by Omidyar is a rare skill. Most policy orgs are locked into one partisan ecosystem.

Bridge between communities. Encode connects the EA/longtermist AI safety community (via CAIS, FLI, SFF, Secure AI Project) with the progressive advocacy world (via Omidyar, Heising-Simons, Design It For Us, March On) and the mainstream policy world (via Senate connections, WEF presence). This bridging function may be their most distinctive contribution to the AI safety ecosystem.

Professional staff with serious credentials. Calvin (Stanford JD, CAIS, Senate Judiciary), Larkin (IFP), Snyder (Yale, biosecurity), Christian (7 years in CA legislature). These are not student volunteers playing at policy -- they are experienced professionals who happen to work at a youth-branded org.

Compelling narrative. A college student founding an organization that takes on a half-trillion-dollar company and wins (SB 53 passed despite OpenAI's opposition) is an extraordinarily powerful story for building public support.

Weaknesses and Risks

Financial opacity is a real problem. An organization that champions transparency for AI companies maintains complete opacity about its own finances. The c4 + fiscal sponsorship structure means there is no way to verify the independence claim ("no corporate or AI exec funding"). This creates a legitimate trust deficit.

Governance vacuum. No board, no conflict of interest policy, no formal decision-making process. The Swetha Revanur / Hebbia conflict is undisclosed. For an org that files legal briefs and makes political donations, the absence of governance structures is a material weakness.

Founder dependency. Sneha Revanur IS the brand. Her personal narrative drives all media coverage and opens all doors. But she is still a college student. What happens when she graduates? Does she become a full-time employee? Does leadership formally transition to the professional staff? Youth-founded organizations frequently collapse at this transition point.

The rebrand may backfire. Scrubbing progressive history (Cori Bush internship, ICE blog posts) while claiming nonpartisanship creates an honesty problem. If more revelations emerge about past progressive positioning, the "nonpartisan" brand takes damage. Being transparent about the evolution would have been more credible.

No technical capacity. Encode relies entirely on external sources (CAIS, academic researchers, labs) for the technical basis of their policy positions. They cannot independently evaluate whether a proposed safety requirement is technically sound. This makes them dependent on the very experts and institutions they seek to regulate.

Membership claims may be inflated. "1,000 members in 40+ states and 30+ countries" could mean 1,000 email signups. The NC chapter profile shows a handful of active students. The distinction between supporters, signups, and active members is never clarified.

Cross-References

CAIS / CAIS Action Fund: Direct pipeline. Nathan Calvin was senior policy counsel at CAIS Action Fund before joining Encode. The SB 1047 coalition (Encode + CAIS AF + Economic Security CA) shows how these organizations work as a legislative team. Encode provides the advocacy/public face; CAIS provides the technical/research basis.

Secure AI Project: Active collaboration on Utah HB 286. Nick Beckstead (ex-Open Phil/Coefficient Giving CEO) runs SAIP. Both orgs were investigated together by the Washington Examiner. This collaboration connects Encode to the EA/longtermist policy network.

Institute for Progress (IFP): Claire Larkin came from IFP, suggesting convergence between the "progress studies" and "AI safety advocacy" communities.

PauseAI / ControlAI: Encode occupies a distinctly different space. Where PauseAI advocates for halting frontier development, Encode advocates for "rules of the road" within continued development. The "pro-AI, pro-innovation" rebrand explicitly rejects the pause/moratorium framing.

Midas Project: Tyler Johnston (Midas Project founder) was also subpoenaed by OpenAI. The two organizations share an adversarial relationship with OpenAI on the restructuring question.

OpenAI: Antagonistic relationship. OpenAI subpoenaed Encode, attempted to weaken SB 53, and OpenAI CSO Jason Kwon publicly accused Encode of being linked to Elon Musk (which Encode denied). Calvin himself notes he "uses and gets value from OpenAI products" -- the criticism is targeted at their political behavior, not their technology.

What Would Change This Assessment

• If Encode's actual budget turned out to be substantially larger than estimated (say, >$2M), it would suggest major undisclosed funders and undermine the scrappy-underdog narrative.
• If corporate or AI executive funding were discovered despite the public claim to the contrary, it would be severely damaging to credibility.
• If Sneha Revanur stepped back and Encode maintained its effectiveness, it would prove the model is institutionally robust rather than personality-dependent.
• If state AI regulations proved to be easily circumvented or preempted by federal law, Encode's entire legislative strategy would need rethinking.
• If a major internal conflict or staff departure occurred with public criticism, it would reveal whether the governance vacuum is a practical problem or merely a theoretical one.
• If Encode started accepting lab funding or softening positions on frontier lab regulation, it would suggest the incentive pressures of the AI policy ecosystem are reshaping the organization.

Self-Critique

What sources should I have checked but didn't:

• The FLI podcast transcript with Sneha Revanur (only the description was available). This could contain the most candid views on x-risk.
• The Safety Summit report PDF (unreadable binary). This could be their most substantive policy contribution.
• Sneha's CNN, Guardian, and 60 Minutes appearances (not fetched by scout). These might contain more candid statements than the print interviews I read.

Where is this analysis potentially biased:

• I may be too sympathetic to Encode because the OpenAI subpoena incident makes them a sympathetic underdog. I should be more skeptical about whether their actual policy impact justifies the attention.
• The absence of substantive criticism could mean the org is too small to criticize (not that it is uncriticizable). I may be over-interpreting this absence as evidence of quality.
• I may be underweighting the partisan criticism because it comes from conservative sources (Washington Examiner, InfluenceWatch). The factual documentation in those sources is legitimate even if the framing is partisan.

What would a thoughtful person who disagrees say: "Encode is an astroturf operation that uses young people as cover for a professionally-run progressive advocacy shop funded by liberal billionaires. The 'youth-led' branding is marketing. The 'nonpartisan' claim is contradicted by the staff's progressive backgrounds and funding sources. The rebrand is a cynical attempt to hide their political identity. Their legislative wins are modest (SB 53 is a disclosure requirement, not a safety mandate) and will be easily circumvented. The real policy work in AI safety is being done by technical organizations, not activists who can't evaluate whether a safety test is meaningful."

My single weakest claim: That Encode's legislative approach will meaningfully improve AI safety outcomes, rather than producing compliance paperwork. The gap between "companies must publish transparency reports" (SB 53) and "companies must actually build safe AI" is large, and it is not clear that Encode's disclosure-focused approach bridges it.

What information would most change my view: Full financial disclosure. If Encode's actual budget, donor list, and spending breakdown were public, many of the uncertainty-driven concerns in this analysis would resolve in one direction or another.