Who he is
Affiliation: Founder & CEO, Khan Academy (nonprofit, 2006–) — role check run 2026-09-21: still CEO; he signs the 2026-08-24 and 2026-08-27 Khan Academy posts as founder and CEO, and Fortune (2026-04-15) carries the same title. Hat count (enumerated, per the multi-hat rule): seven. (1) Khan Academy, founder & CEO; (2) Schoolhouse.world, founder & CEO (free peer tutoring; College Board release 2026-02-03); (3) Khan Lab School, founder (Mountain View, 2014); (4) Khan World School at ASU Prep, founder-visionary (online honors school, launched August 2022); (5) TED board member and "Vision Steward" since 2025-10-15, succeeding Chris Anderson as steward — he keeps the Khan Academy job full-time; (6) Khan TED Institute, the nonprofit higher-education venture announced with TED and ETS on 2026-04-14; (7) Aspen Institute trustee — a 2014 Aspen election release surfaced in search, but neither it nor the current trustees page could be fetched; carried as a flag, not a fact. Author of The One World Schoolhouse (2012) and Brave New Words: How AI Will Revolutionize Education (and Why That's a Good Thing) (Viking, 2024-05-14). Born 1976; MIT (EECS and math), Harvard MBA (2003). Reported compensation $871,261 for the fiscal year ending June 2025 (Form 990, the tax return nonprofits file, via ProPublica).
One-line position: AI can be the great equalizer in education — giving every student a personal tutor and every teacher an assistant — but only if we build it deliberately for learning, not engagement, and make it accessible to everyone. Status of that position as of 2026-09-21: the "personal tutor" half has been publicly walked back by its author — "For a lot of students, it was a non-event. They just didn't use it much" (Chalkbeat, 2026-04-09); "I just view it as part of the solution; I don't view it as the end-all and be-all" — and then defended against the harsher reading: "Some people heard my recent comments about Khanmigo and interpreted them as 'AI tutoring failed.' That is not what I said, and it is not what I believe" (Sal's Newsletter, 2026-09-09). The equity half survived intact and hardened into an anti-ban stance: banning AI in schools is "a velvet rope disguised as caution tape" (Khan Academy blog, 2026-08-27). He is the one bench persona whose central claim was tested by a randomized trial he himself commissioned — and who published the null result.
Discipline & technical bet
Not a researcher: a product operator who argues in numbers, running a free platform with 189.6M registered users (SY24-25 annual report). The technical wager, 2023 version: a frontier language model (GPT-4 at launch, 2023-03-14) wrapped in pedagogy guardrails — Socratic, refuses to give answers — bolted beside Khan Academy's mastery-practice content, would reproduce the effect of a human tutor at near-zero marginal cost. The wager was revised in 2026, in his own words: "The AI could not just sit next to the content. It had to be woven into it" (2026-08-27). The revised bet is that the engagement system, not the model, is the product: Khanmigo now appears inside practice automatically after a wrong answer, offers "gentle hints" before submission (the original no-answers guardrail "proved too rigid," KA 2026-06-15), and will give students "credit" for a bot-assisted redo toward the five-in-a-row needed to advance (Hechinger, 2026-09-21). Model layer is deliberately multi-vendor now — Azure OpenAI (2024-05-21), a fine-tuned Phi-3 for middle-school math, Google Gemini for the Writing Coach (2026-01-21) and for interactive diagrams built by a six-engineer, six-month Google.org fellowship (2026-08-27). He owns no model; he owns the content graph, the district relationships and the usage data. Second, newer bet: a competency-based bachelor's in applied AI for under $10,000, AI agents doing tutoring and feedback, ETS doing assessment, TED doing content — the Khan TED Institute (2026-04-14). A non-technical thinker with two very specific application-layer bets, one of which has already been scored.
Key claims (Says)
- AI tutor for every student, AI assistant for every teacher — the core Khanmigo pitch: Bloom's "2-sigma problem" (1984) showed tutoring doubles outcomes but doesn't scale; AI can change that. Graded in three parts. (a) The 2-sigma magnitude: Wrong — the 2020 Nickow, Oreopoulos and Quan meta-analysis of randomized tutoring studies "found that the average effect of tutoring was 0.37 standard deviations, or 14 percentile points" and "Among 96 tutoring studies the authors reviewed, none produced a two-sigma effect" (Education Next, 2024-03-07; that piece also records Khan promoting the 2-sigma frame in a May 2023 TEDx talk titled "The Two Sigma Solution"). Standard deviation (SD) here is the yardstick for how big an effect is; 2 SD would be enormous, 0.37 is solid. (b) AI can deliver tutoring's effect at scale: Wrong — retired by its author before the data landed. Khan first (Chalkbeat, 2026-04-09: "non-event"; "Some will; most won't"); then the trial: "These gains resemble those from Khan Academy practice without AI assistance" (Oreopoulos & Low, NBER w35620, August 2026). (c) An assistant for every teacher: Held — Khanmigo for Teachers has been free to all US K-12 teachers since 2024-05-21 on Microsoft-donated Azure infrastructure; "Nearly half of Khanmigo users over the past school year were grassroots teachers using Khanmigo for free in 70+ countries" (SY24-25 annual report). The teacher tool is the part that got adopted.
- Khanmigo as proof of concept — deployed in classrooms; uses the Socratic method rather than giving answers; designed to ask questions back. Drifted — deployed, yes (795 US districts, "nearly 1.5M licensed learners," SY24-25). But the Socratic no-answers design is exactly what students routed around: the median student messaged it "in only 17 percent of the exercise sessions in which they made a mistake," and "Messages that students did send were mostly bare answers or clicks on suggested prompts" (NBER w35620). Khan Academy then loosened the guardrail (2026-06-15) and now has the bot prompt students "to explain how they arrived at an answer" (2026-08-27). The proof of concept proved the concept students don't want.
- Education AI must be built for learning, not engagement — the design principles matter; Khanmigo won't write your essay but will help you think it through. Drifted — mechanism and valence graded separately. Mechanism held: the essay guardrail still stands (Writing Coach "guides students through outlining, drafting and refining" rather than producing text, Google 2026-01-21), and internal tests "reduced giving away the answer by 50%" (KA, 2026-05-06). Valence inverted: a learning-first tool that students don't open teaches nothing, and the 2026 redesign is engagement engineering by another name — auto-appear after mistakes, credit for redos, "We had to make productive struggle harder to sidestep" (Chalkbeat, 2026-08-25). His own summary: "What this has put in huge spotlight — not that it's a shock to educators — is how much motivation matters."
- AI can close equity gaps — students in under-resourced schools get the same tutoring as students with private tutors, if the technology is universally available. Drifted — the distribution evidence runs the other way so far: the PNAS study he co-published with Oreopoulos and Khan Academy's own research staff finds "Higher-achieving students benefit most, in part because they spend more time on CAL and progress through more skills than lower-performing peers" (Eames, Brunskill, Yamkovenko, Weatherholtz & Oreopoulos, PNAS 123(1), 2026). The RCT population was students at least a grade behind; those who stayed the full second year gained 0.14 SD — real, but no more than non-AI practice. And the access half is now being decided by policy against him: New York City removed student-facing AI for 2-K through grade 8 (600,000 students) for 2026-27 (K-12 Dive, 2026-09-02); LAUSD barred all students from generative AI on district devices (K-12 Dive, 2026-09-03). The claim has become conditional on human engagement systems and on winning a policy fight.
- Teachers become more valuable, not less — AI handles differentiation and scaffolding; teachers keep mentorship, motivation, the human dimensions. Held on Says — "AI is going to help. But I think our biggest lever is really investing in the human systems" (2026-04-09); "if I had to pick between amazing teacher and amazing technology. I'd pick an amazing teacher every time" (TED AI Show, September 2024; repeated in TIME, 2026-03-01); "we see that when teachers make it part of the assignments, that drives engagement" (CSMonitor, 2026-08-07). Does is less tidy: the Khan TED Institute is "100% online initially" with "AI agents providing tutoring and feedback" (The 74, 2026-04-14) — teacher-light by design. Logged under Weak spots.
- NEW (2025-12 → 2026-02): AI will displace white-collar workers at a scale most people don't yet see. "Every company benefiting from automation should dedicate 1 percent of its profits to help retrain displaced workers," which "would create a $10 billion annual fund" (NYT op-ed, 2025-12-27, quoted via Slashdot 2025-12-28; the Times itself blocks fetches). "If white-collar work were to shrink even 10%, it's going to feel like a depression"; "I've heard people say you could do the same work with a quarter of the team" (Fortune, 2026-02-12). Live — a market-level claim; check under Predictions. Note the register change: the education optimist is now a labor pessimist.
- NEW (2026-08/09): Bans are the wrong response. "If we respond to the challenges of AI by banning it in schools, we are putting up a velvet rope disguised as caution tape, denying under-resourced students the very tools that can level the playing field" (2026-08-27); on the NYC and LA moves: "I don't think this necessarily solves the problem" (CNBC video title, 2026-09-04 — transcript blocked, title only). Live; this is the direct counter to Haidt's 2027 EdTech-rollback bet.
- NEW (2026-08-27): a capability-skeptic claim from an evangelist. "AI is not good at generating standards-aligned content on the fly, for example"; "We are only now starting to see some of the model capabilities that would allow Khanmigo to interact with visuals." Held (self-issued) — consistent with the Gemini-diagrams work landing only in August 2026.
- NEW (2026-03/04): a sub-$10K competency-based degree can compete with elite credentials. "If someone's able to achieve this, you would put them in the same category as someone who graduated summa cum laude from Harvard or Stanford"; "Just imagine something that, as opposed to $400,000, costs $5,000" (TIME, 2026-03-01). Live; check under Predictions.
Notable predictions — with falsifiable checks
- (2024, v1 wording — not a quote) AI tutors will become a standard part of education within 5 years. Correction at migration: the primary says "in the next five or 10 years, we'll have AI bots capable of truly working in conjunction with human teachers" (Edutopia, 2024-11-27) — v1 shortened the horizon and changed the content. Drifted — the near-term version was retired by its author in April 2026; the 5–10-year version is live. Check: by 2029-12-31, at least one independent RCT (a trial where students are assigned by lottery) shows an AI-tutor increment of ≥0.10 SD over non-AI practice at district scale; interim, by 2027-06-30 Khan Academy publishes redesigned-Khanmigo usage showing the median student engages in more than half of mistake sessions (baseline 17%). Miss both = Wrong on the date; direction logged separately.
- (2024) "Properly designed AI tutoring will improve learning outcomes measurably." Wrong on its first large test — Oreopoulos & Low (NBER w35620, August 2026; Hamilton County, Tennessee; 18 middle schools; school years 2024-25 and 2025-26; pre-registered 2024-04-28 as AEA RCT 13519): "Assignment raises math achievement by 1.3 national percentile ranks per term, or about 0.06 to 0.08 standard deviations over a school year; the implied effect of a full year of active participation reaches 0.14 standard deviations. These gains resemble those from Khan Academy practice without AI assistance." Mechanism and valence separately: the platform beat the control group (Khan's own reading: "Year 1 showed no reliable effect," Year 2 alone 0.08 SD "once implementation matured," 0.14 SD for students who stayed the whole second year — KA blog, 2026-08-24); the AI increment was nil. Khan concedes the second half and claims the first. Both are true.
- (Ongoing) "AI will make quality education universally accessible." Unfalsifiable — rewritten. Observable: the share of US public-school students with district-licensed access to a student-facing AI tutor, and the direction of state/district policy. Status 2026-09: ~1.5M licensed learners in 795 districts (SY24-25) against roughly 600,000 NYC students removed from student-facing AI for 2026-27 and all of LAUSD on district devices; North Carolina cut a proposed $10M Khanmigo appropriation to $500,000 (plus $2.5M for Khan Academy partnerships) after "educators and artificial intelligence experts also questioned whether Khanmigo had been sufficiently evaluated" (NC Newsline, 2026-07-01). Check: by 2028-06-30, licensed-learner count above 3M and no additional top-10 district ban = Held; either condition fails = Drifted; NYC renews and two more top-10 districts follow = Wrong.
- (2023-08-15) "Expect in this coming year, there's going to be probably 20 to 100,000 students" using Khanmigo (Rethinking with Adam Grant). Held (exceeded) — "we are about 160,000 students and teachers using it right now" (TED AI Show, September 2024; mixed count, so the student-only number is unknown but the band was cleared).
- (2024-09) "It's probably going to be on the order of a million" for the coming school year (TED AI Show). Held — "2.0M Total Global Khanmigo Users Who had access to Khanmigo at any point during SY24-25" and ~1.5M licensed learners; sign-ups "climbed 731% YoY" (annual report). Note the definition: access, not active use.
- (2024-01) Generative AI "cutting 90% of teachers' admin tasks" (as quoted by Dan Meyer, 2024-01-17, who countered that only 2% of teachers used it extensively). Unfalsifiable as stated — rewrite: no dataset measures admin-task share. Nearest data is Khan Academy's own pilot teacher self-report, "about four to fives [sic] hours per week" saved (KA, 2026-06-15), against a stated baseline that prep "can consume more than 50% of their workload." Proposed observable: a time-use survey of teachers with Khanmigo access by 2027-12-31 showing ≥30% reduction in prep hours; anything under 30% grades the 90% Wrong.
- (2024-09) "We're probably a year, year and a half away" before voice-conversation tutoring "is more accessible." Unfalsifiable — vague referent (accessible to whom, at what price); not graded.
- (2026-02-12) White-collar displacement visible within ~3 years: "When three years [from now] the stuff is really hitting the fan, you can't wait another year to even have something to start mitigating the problem." Live. Check: by 2029-02-28, US white-collar (professional and business services + information + finance) payroll employment is down ≥3% from its 2025 peak or the entry-level relative-employment gap in AI-exposed occupations (the Canaries series used to regrade Brynjolfsson) exceeds −25%. Neither = Wrong on the date. His "10% shrink = depression" line is graded as market-level only, never on anecdotes about "a quarter of the team."
- (2026-04-14) Khan TED Institute: applications "expected to open in 12 to 18 months," under $10,000, bachelor's in applied AI, accreditation to be sought (ETS release and KA blog, 2026-04-14; Fortune's "12 to 24 months" is the looser number — I trust the parties' own release). Live. Check: by 2027-10-31 applications are open and an accreditor or pathway is named publicly; by 2029-06-30 a first cohort is enrolled at a published all-in price under $10,000. Miss the first = Drifted; miss the second = Wrong.
- (2026-08-24) The redesigned Khanmigo, "part of practice rather than something students have to go and open," ships to "All our school district partners … this school year." Live. Check: by 2027-06-30 Khan Academy publishes mistake-session engagement for the redesign (baseline 17%) and registers or announces a second randomized trial. No number by then = Drifted; a number that doesn't beat 17% by at least half = the redesign thesis is Wrong.
Revealed behavior (Does)
- Money (Form 990 via ProPublica, fiscal year ending June 2025): revenue $117,488,956; expenses $87,533,668; net assets $152,215,344; Khan's reported compensation $871,261. The organization changed its fiscal year from December to June — the "2024" filing is a six-month stub showing $28.1M revenue, and any source quoting that as an annual collapse is misreading it. The SY24-25 annual report gives $128M revenue and $95M expenses on its own basis; the 990 is the primary. Khan Academy is a $117M-revenue organization that sells district AI licenses, not "a small nonprofit funded by donations," which is how its own 2023 launch post described it.
- Where the time went, 2025–26 (grade priorities by time, not adjectives): (1) TED — board seat and Vision Steward (2025-10-15); (2) the Khan TED Institute — a higher-ed credential (2026-04-14); (3) labor policy — the NYT 1%-of-profits op-ed (2025-12-27), Fortune (2026-02-12); (4) the Khanmigo rebuild — "reimagined" platform (2026-01-07), five-district pilot of ~5,100 students (2026-06-15), all districts on the redesign (2026-08-24). The evangelist's hours moved from the AI tutor to credentials and workforce policy; the tutor became an engineering project run by the product team.
- Vendor diversification: OpenAI launch partner (GPT-4, 2023-03-14) → Microsoft (Azure OpenAI Service, free teacher tier, Phi-3 math fine-tune, 2024-05-21) → Google (Gemini Writing Coach 2026-01-21; Reading Coach due later in 2026; six Google.org fellows for six months building Gemini-powered diagrams and a teacher practice tool, 2026-08-27). The v1 "dependency on OpenAI" is gone; a dependency on Big Tech subsidy replaced it.
- Published the null result himself. Commissioned an independent RCT (pre-registered 2024-04-28 with Oreopoulos — the co-author of the tutoring meta-analysis that had already deflated the 2-sigma number), let it run two years, and wrote the blog post explaining the 0.06/0.08/0.14 SD findings the day before Chalkbeat's coverage (2026-08-24). Compared the 0.14 to "the very strong 0.45 effect size we saw in the India study" and called the Tennessee control group, which used competing edtech, "a much higher bar." He told Hechinger the early release did "no harm" (2026-09-21). This is the strongest track-record entry in the file, regardless of the result.
- Pricing: learners $4/month or $44/year; teachers free; districts "Request pricing" (khanmigo.ai, read 2026-09-21). Khan Academy's own 2023-11-17 post already said $4/$44; a 2024-03 Education Next piece said "$9 a month, or $99 per year" — I trust the vendor's page and treat the $9 figure as an early-2023 price or an error.
- Public money: Ohio gives all teachers and students free Khan Academy tools through TechCred (KA blog, 2025-08-19); North Carolina's Senate proposal of $10M for Khanmigo, drafted as a no-bid contract, was cut in conference to $500,000 nonrecurring plus $2.5M for Khan Academy partnerships (NC Newsline, 2026-07-01).
- Internal experimentation at scale: "roughly 20 substantive product tests" over "more than 15 million tutoring threads across a six-month period," yielding a "6.1 percent improvement in next-item correctness," with a paper submitted to the AI in Education conference (KA, 2026-05-06). Small, real, and honest about what "showed no effect."
Feels
Wants it to be true that technology can give every child a tutor. The 2026 wound is the word "non-event" in his own mouth; the newsletter line "That is not what I said, and it is not what I believe" is a man guarding his life's thesis from a headline. Fear: that bans lock poor kids out while rich kids get private access — the "velvet rope." Hope, now transferred: that motivation can be engineered, and that a $10K degree can do for credentials what free videos did for lectures.
Hears
Philip Oreopoulos (his evaluator and co-author, and the economist whose meta-analysis capped tutoring at 0.37 SD); Bill Gates (Gates Notes ran a review of Brave New Words — surfaced in search, not fetched this cycle); district partners' pilot feedback (Taft ISD, Hanover Community, Newark); partner channels at Microsoft, Google and OpenAI; now TED's community and ETS's assessment people. Bloom's 1984 paper is the origin text; Bastani et al. (2025) — high-school students given unrestricted GPT-4 access "solved more practice problems but performed worse on subsequent unassisted" tests — is the paper Oreopoulos and Low cite as the reasoning behind Khanmigo's no-answers design.
Sees
The largest live dataset on what students actually do with an AI tutor — 15M+ tutoring threads in six months, 795 districts, plus the first two-year RCT — before the rest of the field. Almost nobody else can see the gap between "had access" (2.0M) and "asked it a question after a mistake" (17%). He also sees district procurement and state budgets from the vendor's chair, which is why his bans argument is so specific.
Incentive map
Re-derived this cycle from income sources, not tax status. Who pays him: Khan Academy, $871,261 (FY June 2025). What the organization sells: Khanmigo district licenses (custom pricing) to 795 US districts; $4/month consumer subscriptions; state contracts (NC, Ohio); the teacher tier is free because Microsoft underwrites the compute, and the newest features exist because Google donated six engineers. What the new hats sell: the Khan TED Institute is a credential product whose pitch ("as opposed to $400,000, costs $5,000") is his newest thesis — the thesis is the product (the Ng pairing rule). The 1%-of-profits "centralized skill training platform" he proposes is a natural customer for exactly the kind of platform he runs; log the pairing, don't assume the motive. What he can't say: that the teacher tools are the business and the student tutor is the brand; that LLM vendors' free compute is what makes "free for teachers" possible; that a ban on student-facing AI in NYC removes 600,000 potential seats from his market — the equity argument and the sales argument point the same way. What he did say anyway: that his own AI increment came back nil. That earns a real discount on the discount. Rule-25 weight: moderate — higher than v1's "nonprofit model gives more credibility" (a $117M-revenue vendor with state contracts is not an unconflicted witness on AI-in-schools policy), lower than for-profit ed-tech CEOs because he funded and published the test that hurt him.
Theories aligned with
- Bloom's 2-sigma problem (tutoring as the gold standard) — still his frame, now carrying the 0.37 SD asterisk from the meta-analysis and the 0.14 SD ceiling from his own trial
- Education-technology as equity intervention — restated in 2026 as an anti-ban argument
- Adjacent to techno-optimism — education-specific variant; since 2025-12 paired with a displacement-pessimist labor view (1% fund; "feel like a depression") that lands him beside white-collar labor end as a policy voice, not a data source
- NEW: competency-based credentialing — "focus on what you can actually demonstrate while keeping the bar high" (KA blog, 2026-04-14)
What he's reacting against
- Ed-tech that optimizes for engagement over learning — insists AI tutors must be pedagogically grounded, not gamified (and in 2026 discovered that the un-gamified tutor is the one nobody opens)
- AI skepticism that ignores education's structural inequities — opposing AI tutoring means defending a system where only wealthy kids get personal attention
- The assumption that AI in education means replacing teachers — Khanmigo is teacher support, not replacement
- One-size-fits-all education — AI enables personalized learning at scale
- NEW: school AI bans and moratoria (NYC 2026-09-02; LAUSD 2026-09-03) — "velvet rope disguised as caution tape"
- NEW: the "AI tutoring failed" reading of his own April 2026 comments (newsletter, 2026-09-09)
- NEW: the cost of college — "as opposed to $400,000, costs $5,000"; siloed corporate certificates ("you get a Google certification and then you go to Microsoft and they could care less about that," TIME 2026-03-01)
- NEW: unrestricted "answer machines" — the trial paper cites the Bastani et al. (2025) finding that free GPT-4 access made students worse on later unassisted tests as the reasoning behind his no-answers design
Where he overlaps / splits (with Rich)
- Overlaps with Ng on AI education as democratization; splits on audience — Ng teaches AI practitioners, Khan teaches K-12. 2026 update: the split narrowed — both now run credential products for adults (Ng chairs Coursera–Udemy; Khan launched an applied-AI bachelor's), and Khan's loudest 2026 audience is displaced workers, not children
- Overlaps with Karpathy on education being high-leverage; splits on approach — Karpathy teaches technical depth, Khan teaches at scale with AI assistance. 2026 update: both moved their own time off the AI-tutor product — Karpathy to pre-training, Khan to credentials and labor policy
- Overlaps with Li on human-centered AI; splits on domain — Li works across AI applications, Khan is education-specific
- Splits with Haidt on technology and children — Haidt warns about screens; Khan argues AI tutoring is fundamentally different from social media. 2026 update — the live pair on the bench: Haidt praised the product ("Khan Academy uses AI very well. And if we could have a device that just did Khan Academy and nothing else, I could see that having a positive impact on education," Ezra Klein Show, as quoted by Khan Academy 2025-06-10), then predicted "one [rollback] will come for EdTech in 2027, in particular for the '1:1 device policies'" and noted that "Britain's poorest children are given chatbot tutors" (After Babel, 2026-07-13). NYC and LAUSD (September 2026) are the first data points on Haidt's bet; Khan's "velvet rope" is the direct counter. Both now name motivation and human attention as the binding constraint — converging on mechanism, split on prescription. Haidt's "years of testing" demand (TED, 2026-04, per the Haidt persona file — not re-fetched here) was aimed at AI therapists, not tutors; but the Khanmigo RCT is the testing, and it came back null on the AI increment — the persona who ran the test is the one Haidt is arguing with
- Splits with skeptics who argue AI can't teach effectively — v1: "Khanmigo deployment data (contested, early) suggests learning gains." Wrong as written — the data are no longer contested or early; the RCT found the platform helps and the AI adds nothing measurable. The skeptics were right on the increment; Khan was right that the platform beats the control; both were wrong if they thought the answer would be dramatic
- Splits with Andreessen's infinitely-patient-AI-tutor vision — Khan built it, and the median student asked it nothing in 83% of the sessions where they got something wrong
- NEW overlap with Brynjolfsson, Autor, Acemoglu on displacement — Khan is now a labor-policy voice ("10% … feel like a depression"; 1% fund), adding no data but a prescription; the bench's labor personas should treat him as a policy advocate, not a source
Track record
- Khan Academy: 189.6M registered users, 104.9M yearly active learners, 190+ countries, 55+ languages (SY24-25 annual report) — v1's "~150M" was the 2024 book-jacket figure; the claim "most successful education platform in history by scale" Held
- Khanmigo launched with OpenAI partnership (GPT-4, 2023-03-14) — among the first educationally serious AI tutor deployments Held; by SY24-25, 2.0M users with access and ~1.5M licensed learners across 795 districts
- Brave New Words (Viking, 2024-05-14) positioned him as the leading voice for AI in education Held — and he is the leading voice on its limits too: "For a lot of students, it was a non-event" was re-quoted by Fortune, The 74, the Christian Science Monitor and Hechinger within five months of Chalkbeat printing it
- Khan Academy's nonprofit model gives more credibility than for-profit ed-tech — but dependency on philanthropic funding creates constraints Drifted — see Incentive map: $117M revenue, district sales, state contracts, Microsoft/Google subsidies; the credibility now rests on his publishing behavior, not the tax status
- NEW: the RCT. Commissioned, pre-registered (2024-04-28), ran two years, published as NBER w35620 (August 2026) and blogged by Khan (2026-08-24). Result: platform works (0.06–0.08 SD ITT — "intent to treat," everyone assigned, whether or not they used it; 0.14 SD for full-year participants), AI increment nil. Scored honestly by the persona within days
- NEW: Khan Academy-run studies. Newark: 13 North Ward schools, ~8,000 students, "a six-to-eight-point gain on NJSLA" — "close to triple the average state gain and 1.5 times the district gain" (KA blog by Bogdan Yamkovenko, 2025-11-21; in-house, not independent). PNAS 2026 with Oreopoulos and Stanford's Brunskill: 0.031 SD at 6.6 hours a year, "0.085 SD at the recommended 30 min per week"; higher achievers benefit most
- NEW: forecasting pattern. User-count calls (2023, 2024) both cleared; effect-size calls (2-sigma, "measurable" AI gains) both missed; a 90%-admin-tasks call was never measurable. Excellent on adoption, over-optimistic on learning magnitude, unusually willing to score himself
- NEW: institutional weight. TED board and Vision Steward (2025-10-15); Khan TED Institute with ETS (2026-04-14); AFT president Weingarten toured a Newark school to see Khanmigo (Chalkbeat, 2026-07-14); Schoolhouse.world "delivered 30,000 SAT bootcamps to 118,000 learners" in the College Board pilot (2026-02-03)
Empirical vs normative
- Empirical: Khan Academy practice raises math scores modestly (0.03–0.14 SD depending on dose and population); the AI tutor adds no measurable increment when students can ignore it; students ignore it in 83% of mistake sessions; teachers adopt the assistant tools; higher achievers use it more; AI is "not good at generating standards-aligned content on the fly"
- Normative: every student deserves personalized learning; AI should serve equity; education AI must be designed for learning, not engagement; bans are a velvet rope; companies that automate owe 1% of profits to retraining; a $10K competency degree should count like a Harvard one
- v1 said the empirical evidence was "early and self-reported"; as of 2026 the central piece is late and independent, and it cuts against the headline. The normative vision is unchanged and now carries a policy fight
- Conflict note (CONVENTIONS rule 25): moderate — a $117M-revenue vendor arguing against bans on his own product category, offset by having funded and published the null result
Weak spots / open questions
- Khanmigo early data is promising but not yet independently validated — the pedagogical claims need rigorous RCTs at scale. Held — and answered. The RCT ran; it validated the platform and not the AI. The weak spot was the right question; the answer went against the thesis
- Dependency on OpenAI's API — what if costs rise or priorities change? Drifted — mitigated, then replaced. Multi-vendor since 2024-05-21 (Microsoft) and 2026-01-21 (Google); the new exposure is subsidy dependence: the free teacher tier exists because Microsoft donates compute and the newest features because Google donated engineers
- AI tutoring that asks questions back may frustrate students expecting quick answers — adoption depends on culture change. Held — confirmed by the trial. Students "largely stopped using it" when it asked rather than told; Khan Academy loosened the guardrail (2026-06-15) and Khan's own explanation was "a lot of kids might be too embarrassed. They wouldn't know how to do it" (CSMonitor, 2026-08-07)
- Equity promise requires universal free access — which requires sustained philanthropy or public funding. Held. The philanthropy arrived (Microsoft, Google.org); the public funding went the other way (NC cut; NYC and LAUSD removed access)
- "AI tutor" framing understates the complexity of education — motivation, social dynamics, emotional development. Held — in his own words. "How much motivation matters" (2026-08-25); "our biggest lever is really investing in the human systems" (2026-04-09)
- NEW: Says "human systems," builds an AI-first, online-only degree. The Khan TED Institute's design ("AI agents providing tutoring and feedback," "100% online initially") is the K-12 thesis he just retired, re-deployed for adults with less teacher in the loop, not more. Check the Institute's staffing model when applications open
- NEW: the "nonprofit = credible" discount is stale. Grade him as a vendor with a good publishing record, not as a neutral party on bans
- NEW: no independent evaluation of the teacher tools' time savings exists — the only numbers are pilot self-reports; the adopted half of the product is the unmeasured half
- NEW: "2.0M Khanmigo users" means "had access at any point"; active-use numbers for the student tutor have never been published by Khan Academy. Ask for them at every review
- NEW: the Khan TED Institute's legal entity, funding and Khan's title in it are undisclosed (ETS calls it "a nonprofit higher education collaboration established by three organizations"); the Aspen trustee seat is unverified this cycle; the "350 staff" figure circulating from an interview clip has no fetched primary and is not in the file
Rich's take
- (your synthesis here)
Delta log
2026-09-21 — v1→v2 migration + validation (bulk run)
- Grades: 11 Held / 6 Drifted / 4 Wrong (plus 3 unfalsifiable items rewritten with observables, not graded).
- Wrong: Bloom's 2-sigma magnitude (average tutoring effect 0.37 SD; none of 96 randomized studies reached 2σ — Education Next, 2024-03-07); "AI can deliver tutoring's effect at scale" — self-retired ("non-event," Chalkbeat, 2026-04-09) and then nulled ("These gains resemble those from Khan Academy practice without AI assistance," Oreopoulos & Low, NBER w35620, August 2026); the 2024 "measurable improvement" prediction (same trial); v1's "deployment data (contested, early) suggests learning gains" line.
- Drifted: Khanmigo-as-Socratic-proof-of-concept (guardrail loosened, KA, 2026-06-15); learning-not-engagement (mechanism held, valence inverted; redesign is engagement engineering, Chalkbeat, 2026-08-25); AI closes equity gaps (higher achievers benefit most, PNAS 123(1), 2026; NYC/LAUSD removals, K-12 Dive, 2026-09-02); the 5-year tutor prediction (transcription corrected to "five or 10 years," Edutopia, 2024-11-27; near-term version retired); "nonprofit gives more credibility" (re-derived incentive map from the FY June-2025 Form 990); the OpenAI-dependency weak spot (multi-vendor, Google, 2026-01-21).
- Held: assistant-for-every-teacher (free tier since 2024-05-21; nearly half of users are teachers, SY24-25 annual report); teachers-more-valuable (Says); both user-count forecasts (2023, 2024); the scale, launch and book track-record entries; four of five v1 weak spots (validation, question-asking friction, access-needs-funding, motivation) — each confirmed by the trial or by his own words.
- Most surprising delta: the author of "an AI tutor for every student" retired the super-tutor claim in April, then his own pre-registered two-year RCT found the AI added nothing over plain Khan Academy in August — and he wrote the explainer himself (KA blog, 2026-08-24) the day before Chalkbeat, while re-engineering the product so students can't sidestep it.
- Employer/role check (mandatory): no job change — still founder and CEO of Khan Academy. Hats missed by the June-2026 pass: three — TED board member and Vision Steward (PR Newswire, 2025-10-15), the Khan TED Institute (ETS, 2026-04-14), and the labor-policy platform (NYT op-ed 2025-12-27 via Slashdot; Fortune, 2026-02-12). Schoolhouse.world, Khan Lab School and Khan World School were also absent from v1. Compensation and revenue entered from the 990 for the first time.
- Post-review sources (the defective-pass rule): the June-2026 v1 review stamped 2026-06-16 mentioned none of: the April "non-event" interview, the TED role (8 months old), the Institute (2 months old), the NYT op-ed (6 months old), or the Gemini partnership (5 months old). Treated as never reviewed; every 2025–26 item above is new to the bench.
- Haidt cross-check (lead from the sibling run, verified on the primary After Babel, 2026-07-13): the 2027 EdTech-rollback prediction and "chatbot tutors" line are real; the "years of testing" demand targets AI therapists, not tutors. Intersection logged in Overlaps: NYC/LAUSD are Haidt's leading indicators, Khan's RCT is the testing, and Khan's "velvet rope" is the counter.
- Tier: quarterly — the redesigned Khanmigo's first usage data, NYC's one-year moratorium review, LAUSD's committee recommendations and the Institute's application window all land within two quarters; the Khan–Haidt pair resolves on 2027 data.
- Confidence unchanged at medium: sources are strong and primary, but the persona's position is mid-pivot and the Institute is a promise, not a product.
Sources
- Primary — the trial: Oreopoulos & Low, "One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment," NBER w35620 (August 2026) · EdWorkingPaper 26-1551 (same paper; PDF names Hamilton County, TN, school years 2024-25 and 2025-26) · AEA RCT Registry 13519 — pre-registration (2024-04-28; PI Oreopoulos; ~3,300 Tier 2/3 students planned) · Sal Khan, "What I found compelling in a new randomized trial…" (2026-08-24) · Chalkbeat, "Students rarely engaged with … Khanmigo, study finds" (2026-08-25) · Hechinger Report / KQED, Barshay (2026-09-21)
- Primary — his own 2026 statements: "Khanmigo's First Chapter Changed How I Think About AI" (2026-08-27) — "velvet rope," "woven into it," "not good at generating standards-aligned content" · Sal's Newsletter, Fall 2026 (2026-09-09) — "That is not what I said" · Chalkbeat, Barnum, "Why Sal Khan is rethinking how AI will change schools" (2026-04-09) · Christian Science Monitor (2026-08-07) · CNBC video, "I don't think this necessarily solves the problem" (2026-09-04; title only, transcript blocked) · TIME Charter (2026-03-01) · Fortune (2026-02-12) · Slashdot summary of the NYT op-ed "We Need a Better Plan for Worker Training" (op-ed 2025-12-27; NYT blocks fetches)
- Primary — product and organization: Khan Academy Annual Report SY24-25 (189.6M registered; 2.0M Khanmigo users; 795 districts; 731% YoY) · Form 990s via ProPublica (FY June 2025: revenue $117,488,956; expenses $87,533,668; Khan $871,261) · "Khan Academy Reimagined" (2026-01-07) · "Building a Better AI Tutor" — 15M threads, 6.1% (2026-05-06) · "Built in the Open" — 5 districts, ~5,100 students, guardrail "too rigid" (2026-06-15) · Khanmigo pricing (read 2026-09-21) · Khanmigo launch post (2023-03-14, updated 2023-11-17; GPT-4; $4/$44) · Newark three-year study (2025-11-21, in-house) · Eames et al., PNAS 123(1), 2026 (via RePEc; PNAS blocks fetches)
- Primary — partners and new hats: TED leadership release (2025-10-15) · ETS — Khan TED Institute (2026-04-14) · KA — "Introducing the Khan TED Institute" (2026-04-14) · The 74 — five things (2026-04-14) · Fortune (2026-04-15) · Microsoft Education — free Khanmigo for Teachers, Azure OpenAI, Phi-3 (2024-05-21) · Google — Gemini Writing/Reading Coach (2026-01-21) · Google — Google.org fellows, interactive diagrams (2026-08-27) · College Board — Schoolhouse.world SAT tutoring (2026-02-03) · ASU Prep — Khan World School launch (2022-04-22) · KA — Ohio TechCred (2025-08-19)
- Policy and market context: K-12 Dive — NYC moratorium, 2-K–8, 600,000 students (2026-09-02) · K-12 Dive — LAUSD restriction (2026-09-03) · NC Newsline, Jallow — $10M to $500K (2026-07-01; read via Yahoo syndication, NC Newsline blocks fetches) · Chalkbeat — Claude for Teachers, 61% of teachers using AI (2026-07-14) · Haidt, After Babel — EdTech rollback 2027 (2026-07-13) · KA quoting Haidt on the Ezra Klein Show (2025-06-10) · von Hippel, Education Next — 2-sigma vs 0.37 SD (2024-03-07)
- Earlier primary (predictions dated here): TED2023, "How AI could save (not destroy) education" (April 2023) · KA summary with quotes (2023-05-01) · Rethinking with Adam Grant transcript (2023-08-15) · The TED AI Show transcript (September 2024) · Edutopia interview (2024-11-27) · Dan Meyer on the 90%-admin prediction (2024-01-17) · Penguin Random House — Brave New Words (Viking, 2024-05-14) · Wikipedia (bio dates; used only for birth year, degrees, Khan Lab School 2014, Schoolhouse 2020)
- Carried from v1: Book — Brave New Words (2024) · khanacademy.org · Khanmigo (launched 2023, in partnership with OpenAI) · TED talks (multiple, on Khanmigo and AI in education) · media coverage of launch and deployment
migrated v1→v2 2026-09-21 · backup: _archives/sal-khan.html.bak-20260921 · converts-from: people/sal-khan.md · AI & Society domain