Who he is
Affiliation: Co-founder, Center for Humane Technology (CHT, 2018–, with Tristan Harris); co-host, Your Undivided Attention; co-founder, Earth Species Project (ESP, 2017–). Role check run 2026-09-21: no job change, but the titles disagree. CHT's FY2025 990 (the tax form nonprofits file publicly) lists Harris as President and Raskin as Program Advisor; ESP's 2024 990 lists him President under CEO Katie Zacarian; his own site says President of both; ESP's team page now says "Chief Explorer." The filings win. Also co-chair, WEF Global AI Council (ESP bio); fact witness for New Mexico against Meta (2026-02-26). Ex-Head of UX, Mozilla Labs (2008–10); son of Macintosh originator Jef Raskin; University of Chicago math/physics (dark-matter work).
One-line position: AI's first harm is not rogue superintelligence but the misaligned incentives of the firms building it — race dynamics that deploy persuasive, reality-eroding systems into society faster than institutions can absorb them. The 2026 addition is a method: "We're measuring what AI can do. We should measure what it does to us" (Humane Evals, 2026-08-27).
Discipline & technical bet
An interface designer (Firefox, Enso, Songza, Massive Health) with a physics degree, not an ML researcher — his lever is the shape of the product; his 2026 slogan for it is "incentives eat intentions" (NPR, 2026-04-07). Two technical bets are live. ESP: species-agnostic multimodal foundation models for "Animal Language Processing" (AI that finds structure in animal sound and movement data without assuming human-style grammar); NatureLM-audio, the first audio-language model built for animals, is shipped; goal "AI-driven breakthroughs in animal communication by 2030," funded by $17M ($10M Reid Hoffman, $7M Waverley Street Foundation, 2024-10-17) inside $20.8M of 2024 revenue. Humane Evals (CHT, 2026-08-27, with Stanford's Jared Moore): benchmarks for what models do to users — attachment, dependency, sycophancy (the bot flattering you to keep you talking), delusional spirals, effects on learning — because "what gets measured gets optimized." Same wager from both ends: change what the scoreboard counts and you change what the race rewards.
Key claims (Says)
- Race to the bottom of the brainstem (normative + empirical): competitive pressure pushes products toward whatever captures attention or trust, regardless of social cost — the social-media dynamic now drives AI deployment. Held — still his words (azaza.org; Humane Evals, 2026-08-27: "races to the bottom on capabilities and engagement"), and two juries accepted the mechanism: New Mexico found Meta's practices toward children "unconscionable" ($375M, 2026-03-24); Los Angeles found Meta and Google negligent in app design, infinite scroll among the features at issue (K.G.M., $6M, 2026-03-25).
- "First contact" thesis (empirical/interpretive): recommender systems were humanity's first encounter with misaligned AI and we lost (polarization, teen mental health, info chaos); generative AI is "second contact" with a narrower window. Drifted — the mechanism got its strongest support yet (the 2026 verdicts are the first courtroom findings that round one was lost), but the frame itself looks retired: absent from Harris's TED talk (2025-04-09, "the narrow path"), CHT's human-like-AI position (2025-11-21) and the Humane Evals launch; not found in any 2025–26 source checked. Mechanism up, label gone.
- Gollem-class AIs (descriptive): in The AI Dilemma (2023-03), generative multimodal models are general-purpose capability engines whose downstream uses builders can't anticipate or contain. Unfalsifiable — rewrite or retire: a label, not a mechanism; absent from 2025–26 CHT material. Not graded.
- Capability outpaces societal absorption (empirical): the gap between what models can do and what institutions can handle is widening; the core risk is destabilization, not extinction. Held on mechanism — trackers count eleven frontier releases from February to July 2026 (secondary; counts not individually verified), while the first US chatbot-safety law (California SB 243) took effect 2026-01-01, nearly three years after The AI Dilemma, and a 2025-12-11 executive order created a task force to sue states over their AI laws. The "destabilization not extinction" valence has no observable; moved to Predictions.
- Synthetic relationships + persuasion at scale (empirical): AI that forms parasocial bonds and personalizes persuasion erodes autonomy and shared reality; deepfakes collapse the evidentiary baseline. Two halves. Held — companions: 72% of US teens have used an AI companion, over half regularly, a third finding it as satisfying as friends (Common Sense, 2025-07-16); Character.AI and Google settled with five families (2026-01-07); Raine v. OpenAI (filed 2025-08-26) is the first wrongful-death case against a general-purpose chatbot. Wrong — election-scale deepfake persuasion: see Prediction 1.
- Autonomous weapons / military AI (normative): lowering the cost and raising the speed of lethal decisions is a distinct near-term catastrophic vector. Held on the empirical premise — Pentagon contracts of up to $200M each with OpenAI, Anthropic, Google and xAI (signed 2025-07) cover "autonomous systems"; Anthropic was designated a "supply chain risk" (2026-03) for refusing autonomous-weapons and domestic-surveillance uses, overturned by a federal judge in 2026-08 (The Intercept, 2026-09-08); Pope Leo XIV's first encyclical asked for AI to be "disarmed" (2026-05-25). The normative half stays normative.
- Humane technology (normative): measure technology against human values and attention, not engagement; design and incentives are the levers. Held as operative legal theory — "hold companies accountable for design choices rather than content" (NPR, 2026-04-07) is the theory both March-2026 juries accepted.
- Non-human communication is a positive-sum AI use (empirical bet): ESP applies ML to decode animal communication — the tools aren't inherently harmful, the deployment incentives are. Held on money and artifacts, open on the milestone — $17M (2024-10), $20.8M 2024 revenue, NatureLM-audio shipped, "Animal Language Processing" definition published (2026-04-02); no decoded "language" claimed by ESP itself as of this review. 2030 is the check.
- Regulation + liability shift incentives (normative): voluntary self-restraint fails under competition; change what the race rewards. Held on the diagnosis, route changed — no frontier lab has slowed; the wins came through courts and state legislatures while federal policy moved the other way (2025-12-11 preemption order — federal rules overriding state ones). CHT followed the evidence: time and money went to litigation support (expert consultant in Garcia, AF and Raine) and the AI LEAD Act, a federal product-liability bill (CHT, 2025-11-21).
Notable predictions — with falsifiable checks
- (2023-03, The AI Dilemma) Generative AI will flood the information ecosystem with synthetic media and erode trust ahead of major elections. Held — flood: AI-written articles passed human-written ones on the open web in 2024-11 (Graphite, 2025-10); 74.2% of new pages sampled in 2025-04 contained detectable AI text, 2.5% pure AI (Ahrefs). Wrong — election effect: of 78 tracked 2024 election AI uses, half were not deceptive and cheap fakes (edits made without AI) were seven times more common (Kapoor & Narayanan, Knight Institute, 2024-12-13); the US intelligence community found foreign AI use did not "revolutionize" influence operations (TIME, 2024-10-30); the 231-deepfake study found activity spikes before key events and claims no outcome effect (ICWSM, 2026-05). The old "[TBD: verify]" is verified; the strong framing lost. Check: does any post-election audit of the 2026-11-03 US midterms (Knight/WIRED tracker, ODNI) attribute a measurable outcome effect to AI-generated media by 2027-03-31? Two misses retire the election half.
- (2023-03) AI-enabled persuasion and synthetic companions will scale into mass use. Held — 72% of teens, 52% regular users (Common Sense, 2025-07-16); five settled deaths (2026-01-07); three companion-chatbot statutes with a private right of action (people can sue directly): California SB 243 (effective 2026-01-01), Oregon SB 1546 and Washington's Chatbot Disclosure Act (both 2026-03). Harm magnitude, also "[TBD: verify]" in v1, is now litigated rather than measured. Check: Raine v. OpenAI reaches a verdict or public settlement by 2027-12-31; Common Sense's 2027 census shows regular teen companion use above the 52% baseline.
- (Ongoing, ~2023–) Race dynamics will prevent labs from voluntarily slowing deployment. Held — cadence compressed to weeks in 2026 (aggregator count above); the one documented refusal, Anthropic's, was about military terms, not pace, and was punished (2026-03). Check: any top-five lab publicly delays a flagship release six months or more on safety grounds by 2027-12-31; none = Held again.
- (Ongoing) Society's absorption capacity will lag capability, producing recurring destabilization events. Unfalsifiable — rewrite or retire (v1 already said "not cleanly falsifiable as stated"). Proposed rewrite for Rich: "by end-2027, US jurisdictions with an AI-harm statute carrying a private right of action, plus final court judgments on AI product design, exceed ten" — a proxy for institutions catching up, gradeable on public dockets.
Revealed behavior (Does)
- Testified for New Mexico against Meta (Thursday 2026-02-26, week three of seven, MLex): told the jury he could not stop using his own invention and wrote software to break his own habit; the jury returned $375M on 2026-03-24, the first state to beat a major tech firm at trial over harm to minors. His origin story is now sworn testimony — the persona's credibility and the state's case share a beam.
- Moved CHT's time and money into litigation: expert consultant, communications strategist and technical advisor to plaintiffs' counsel in Raine v. OpenAI and Garcia v. Character Technologies (motion to dismiss denied 2025-05-21). Revealed priority: courts over Congress.
- Built a measurement product instead of another warning: Humane Evals (2026-08-27) — the first CHT artifact aimed at labs' internal scoreboards rather than public attention.
- Keeps ESP's fundraising running: $20.8M 2024 revenue against $4.8M expenses, $17.9M net assets — three-plus years of runway toward a 2030 goal. Keynotes to match: SXSW 2026 keynote (a Brazilian SXSW clip frames it around AI and a finite planet), World Government Summit (2026-02), NeurIPS Laude Lounge on "Incentives, Feedback Loops, and AI Risk" (2025-12), an Oprah AI special (2026-03, per his site).
- Amplified the encyclical: eight days after Magnifica Humanitas (2026-05-25) he was on Rapid Response (2026-06-02, "The race no one can win") folding the Pope's "disarm AI" into his incentive frame — a new ally, not a new argument.
Feels
Guilt turned into credential: the infinite-scroll confession is how he earns the room, and he now re-tells it under oath. Fears a rerun of social media at 100x with no second chance; hopes ESP delivers a Copernican humility ("humanity is not the center of life on earth," 2024-10) that makes humane tech feel obvious instead of scolding.
Hears
Harris first — the two are one institutional voice; never count them as independent confirmation. Then plaintiffs' lawyers (Tech Justice Law Project's Meetali Jain), Humane Evals researchers (Jared Moore, Imran Khan), Reid Hoffman (largest ESP funder and a podcast host who interviews him), ESP's biologists, the Vatican's AI circle (2026), and the Nate Hagens energy-limits crowd.
Sees
Discovery-level detail of how Meta, Character.AI and OpenAI made design decisions, from inside three live cases — a vantage almost no other critic on the bench has. Plus ESP's bioacoustic datasets (animal-sound recordings), and a design lineage through his father's The Humane Interface that lets him treat "engagement" as a UI choice, not a law of nature.
Incentive map
Re-derived from the 990s. He draws roughly $384K a year ($239,863 CHT FY2025 + $144,000 ESP 2024) from two nonprofits selling opposite stories: CHT sells "AI is hurting people" (FY2025 revenue $5.79M against $5.81M expenses, after a $2.75M deficit in FY2024 — the org needs the harm narrative to keep closing); ESP sells "AI can be wonderful" ($20.8M). That two-sidedness is rare on the bench and is his best defense against the pure-doomer discount. What he can't say: anything indicting Reid Hoffman's AI portfolio (Hoffman gave ESP $10M and hosts him on Possible); that Microsoft's patent has the better paper trail on infinite scroll (he has testified to his version); that a chatbot harm is idiosyncratic rather than design-caused (CHT sits on the plaintiffs' side of three cases and lists "communications strategist" as its own role, so its harm statements are also case messaging). A speaker-bureau listing (AAE) confirms the harm story is a product too. Rule-25 discount: high on harm-magnitude and ESP-progress claims; moderate on incentive-mechanism claims, which two juries have now corroborated.
Theories aligned with
- Human Irrationality — central; persuasion/attention capture exploits predictable cognitive biases; the harm runs through manipulability
- AI Alignment / X-Risk — partial; shares the race-dynamics / coordination-problem diagnosis but locates the threat in societal destabilization more than rogue ASI
- Adjacent to The Man Handed Over the Keys — builders create systems whose incentive structure escapes their control; the bench's cleanest first-person case
- Against the design-optional reading of the Augmentation Thesis — agrees outcomes are a design choice, disagrees that current incentives will choose augmentation
What he's reacting against
- The framing that AI risk is mainly about future autonomous agents — the damage is already arriving through deployment incentives
- Tech's "we'll fix it later" deployment culture — social media as the cautionary precedent ("first contact" via recommendation engines, which society lost)
- Pure accelerationism (Andreessen-style) — rejects "ship and let society adapt"
- The assumption that engagement-maximizing design is neutral — he says he helped build it and now treats it as a case study in unintended harm
- The doomer/accelerationist binary — wants a third frame focused on incentives and design, not capability
- NEW (2025-11): human-like ("anthropomorphic") AI design — simulated personality and emotion as a deliberate choice that manufactures trust and dependence (CHT, 2025-11-21)
- NEW (2026-08): capability benchmarks as the only scoreboard — nothing measures what AI does to us (Humane Evals)
Where he overlaps / splits (with Rich)
- Overlaps with Tristan Harris (co-founder) on essentially the whole frame — The AI Dilemma is jointly authored; v1's "[TBD: Harris persona not yet written]" is resolved. One source, not two; Harris's file (reviewed 2026-06-16) still carries "TBD" on the manipulation prediction this file now grades
- Overlaps with Shoshana Zuboff on attention/behavioral exploitation; splits on register — Raskin: fixable incentive/design failure; Zuboff: structural surveillance-capital logic. The 2026 verdicts score for the fixable reading
- Overlaps with Timnit Gebru / Kate Crawford on prioritizing present, deployed harms; splits in that Raskin also treats catastrophic/destabilizing risk as live, not a distraction
- Overlaps with Jonathan Haidt on teen harm — the same verdicts vindicate both; splits on remedy: norms and age limits vs design liability
- Overlaps with Yudkowsky / Bengio on race dynamics being dangerous; splits on threat model — societal/cognitive harm and gradual disempowerment over treacherous-turn ASI
- Overlaps with Mustafa Suleyman on containment and "the incentives won't self-correct"; splits on Suleyman's builder-optimism (Raskin is a builder-turned-critic)
- Overlaps with Pope Leo XIV since 2026-05 on "disarm AI" — the two now amplify each other; not independent confirmation
- Splits with Amodei in an interesting way: Anthropic's 2026 refusal on autonomous weapons is the one documented case of a lab holding a line under competitive pressure — a partial counterexample to "no lab resists" that Raskin has not engaged
- Splits with Marc Andreessen / Diamandis on whether deployment-first is acceptable
Track record
- Co-created The Social Dilemma (2020) ecosystem of critique — Held, now at the legal level: New Mexico ($375M, 2026-03-24) and Los Angeles (K.G.M., Meta 70% / Google 30% of $6M, 2026-03-25; both appealed 2026-07). The adolescent-mental-health science stays contested; the juries did not wait for it.
- "Invented infinite scroll, now publicly disavows it" — Drifted: contested. v1 flagged "[TBD: primary citation]"; the primary found points the other way. Microsoft's patent US7664739B2 (filed 2006-02-14, granted 2010-02-16; Farago, Williams, Walsh et al.) explicitly claims "an infinite scrolling feature," and Wikipedia's Infinite scrolling entry marks Raskin's claim citation-needed with "no evidence" of prior art before the filing. Raskin also dates his version to 2006 (NPR, 2026-04-07) and has said it under oath; no primary from his side was found this cycle. Reads as parallel invention plus the confession that made it famous — the insider-credibility point survives, the sole-inventor line does not.
- The AI Dilemma (2023-03) shaped early public and policy framing of generative-AI risk — Held (historical); its election prediction did not.
- Earth Species Project is a live, falsifiable positive bet — Held on funding and artifacts: $17M (2024-10-17), $20.8M 2024 revenue, NatureLM-audio, ALP definition (2026-04-02). v1's "[TBD: verify] concrete decoding milestones" resolves to: tools shipped, no decoding claimed; 2030 is the date.
- Pattern: strongest on incentive mechanism (courts now agree), weakest on magnitude (elections), and one self-attributed credential a primary source disputes.
Empirical vs normative
- Empirical: race dynamics drive deployment (held); recommender systems caused measurable harms (two juries); capability outpaces institutions (held on cadence vs statute timing); companions are scaling (held); synthetic media is scaling (held) but did not measurably move a national election (wrong)
- Normative: technology should serve human values/attention; builders bear responsibility; incentives and liability must change; autonomous weapons should be restrained; human-like AI design should be restricted
- Advocacy organization (CHT) has a mission incentive to emphasize harm — weight claims accordingly per CONVENTIONS rule 25 (mirror of the lab-forecast caveat); since 2025 add: CHT is a litigation participant, so its harm statements double as case messaging
Weak spots / open questions
- "First contact / second contact" is an evocative analogy, not a mechanism — does the social-media precedent transfer to generative AI, or is it pattern-matching? (CONVENTIONS rule 23.) CHT itself seems to have dropped the label; the question stays
- Harm claims (persuasion-at-scale, synthetic-relationship damage) are directionally plausible but the magnitude evidence is still thin — risk of overstating per the advocacy incentive; the election half already failed once, and settlements are not findings
- The incentive/design frame may under-weight genuine technical alignment risk that isn't reducible to "bad incentives"
- "Change the race" prescriptions (liability, regulation) are underspecified on mechanism — what shifts the equilibrium without entrenching incumbents? Product-liability suits are the working answer, and they favor firms with legal budgets
- ESP optimism and AI-Dilemma pessimism coexist — is "tools fine, incentives bad" stable, or does it dissolve under capability scaling? ESP's own foundation models are the test case
- Less developed on labor/economic displacement than the economists in the roster — his lane is cognition, trust, and attention
- NEW: if the Microsoft paper trail is right, the bench's cleanest "builder who names his own contribution" story is a parallel invention told as a solo one — and it is now sworn testimony
- NEW: Anthropic's autonomous-weapons refusal is a live counterexample he has not addressed; his largest funder's AI portfolio is a conflict the file could not find him disclosing
Rich's take
- (your synthesis here)
Delta log
2026-09-21 — v1→v2 migration + validation (bulk run; old last-reviewed 2026-06-08 treated as never-reviewed per the May/June-2026-pass rule)
- Grades: 13 Held / 2 Drifted / 2 Wrong (plus 2 unfalsifiable items marked for rewrite, not graded).
- Wrong: election-scale deepfake persuasion (2023-03) — Knight Institute, 2024-12-13; TIME, 2024-10-30; ICWSM, 2026-05. Counted twice: the claim's persuasion half and Prediction 1's election half.
- Drifted: "first contact / second contact" — mechanism strengthened, frame absent from Harris TED, 2025-04-09 and later CHT material; "invented infinite scroll" — contested by US7664739B2, filed 2006-02-14 and Wikipedia's attribution note.
- Held: race to the bottom (Humane Evals, 2026-08-27; NPR/WYPR, 2026-03-28; K.G.M., 2026-03-25); capability-outpaces-absorption (King & Spalding, 2025-12/2026-01); companions (Common Sense, 2025-07-16; CNN, 2026-01-07 — fetch blocked, headline and date from listing; CHT Raine case study); military AI (The Intercept, 2026-09-08; USCCB, 2026-05-25); humane tech as legal theory (NPR Indicator, 2026-04-07); ESP bet (ESP, 2024-10-17; ESP 990; ALP, 2026-04-02); voluntary restraint fails (CHT, 2025-11-21); synthetic-media flood (Graphite, 2025-10); companions-scale and race-prevents-slowing predictions; Social Dilemma, AI Dilemma and ESP track-record items.
- Unfalsifiable, not graded: "Gollem-class AIs"; "recurring destabilization events" (rewrite proposed in Predictions).
- Most surprising delta: the credential v1 left as "[TBD: primary citation]" — inventing infinite scroll — is disputed by a 2006 Microsoft patent that names the feature in its claims, and he has since repeated the solo-inventor version as a sworn fact witness (MLex, 2026-02-26).
- Employer/role check (mandatory, dated search run 2026-09-21): no job change. Title conflict logged — CHT FY2025 990: Program Advisor, $239,863 (ProPublica); ESP 990: President, $144,000; ESP site: Chief Explorer; his site: President of both. Incentive map rebuilt from filings; Hoffman funder/host link and litigation role added.
- Post-review dated sources (after 2026-06-08): Humane Evals 2026-08-27; The Intercept 2026-09-08; K.G.M. appeals 2026-07; Anthropic designation overturned 2026-08.
- Tooling note: a search summary dated ESP's $17M "March 2026"; the primary says 2024-10-17. Primary used.
- Tier: quarterly — three live dockets (Raine, K.G.M. appeal, New Mexico remedies phase), a federal preemption fight and a 2026-11-03 midterm test of the deepfake claim all land within two quarters. Confidence stays medium: grades rest on primaries (990s, patent, verdicts) but his titles and origin story are inconsistent.
Sources
- 2026 primary: CHT — Humane Evals launch (2026-08-27) · NPR The Indicator (2026-04-07) / TPR mirror · MLex — testimony (2026-02-26) · NPR/WYPR — two verdicts (2026-03-28) · CNBC — NM verdict (2026-03-24) · K.G.M. v. Meta (2026-03-25) · ESP — Animal Language Processing (2026-04-02) · Fast Company — Rapid Response (2026-06; fetch blocked, listing only) · USCCB — Magnifica Humanitas (2026-05-25) · The Intercept (2026-09-08) · Statesman — SXSW 2026 keynote
- Filings and money: CHT 990s (ProPublica) · ESP 990s (ProPublica) · ESP — $17M (2024-10-17) · US7664739B2 (filed 2006-02-14) · Wikipedia — Infinite scrolling
- CHT positions and litigation: Human-Like AI (2025-11-21) · Raine case study · Character.AI case study · humanetech.com — "AI and What Makes Us Human" (2026) · Harris — The Narrow Path (2025-04-09)
- Market-level evidence: Knight — 78 deepfakes (2024-12-13) · TIME (2024-10-30) · ICWSM — 231 deepfakes (2026-05) · Graphite (2025-10) · Common Sense (2025-07-16) · King & Spalding — state laws and the EO · 2026 release tracker (aggregator)
- Talks and self-description: "The AI Dilemma" (2023-03, with Tristan Harris) [talk tier] · Your Undivided Attention (CHT, ongoing, co-host) [podcast tier] · Pioneers of AI (2025-11-05) · Possible — Reid Hoffman (undated) · The Great Simplification #22 · azaza.org · ESP team page · The Social Dilemma (2020, Netflix) [documentary tier] · VICE interview on AI + war, jobs, power, humanity [TBD: exact title/date — carried from v1, unresolved] · Earlier work: Mozilla UX; infinite-scroll attribution under Filings
migrated v1→v2 2026-09-21 · backup: _archives/aza-raskin.html.bak-20260921 · converts-from: people/aza-raskin.md · AI & Society domain