AI-Wiki

Inside the 2026 AI Index Report

19 min read

Sajadieh / Stanford HAI — Inside the 2026 AI Index Report (talk + panel, 2026-05-27)

The AI Index, currently in its ninth year, tracks, collates, distills, and visualizes data relating to artificial intelligence.

What is changing in the field of AI? Join Stanford HAI’s AI Index Lead Sha Sajadieh for an exclusive look into this trusted source of AI intelligence.

Learn about the year’s major AI breakthroughs, changes in the workplace, policy shifts, and an evolving public sentiment for this technology. From technical advances to ethical considerations, education trends to economic impact — get the unbiased, data-driven insights that executives, policymakers, and global researchers rely on.

(Channel description, Stanford HAI.)

A ~1:12:52 talk + panel-Q&A from the Stanford HAI YouTube channel, published 27 May 2026. Auto-generated English captions (ASR, single track). The session has three parts:

  1. 0:00 – 30:46 — Sha Sajadieh (Editor-in-Chief, AI Index, 2026 edition) presents the report’s headline framing.
  2. 31:36 – 1:11:30 — moderated panel with Raymond Perrault (Co-chair, AI Index Steering Committee; co-founder of the report; SRI International emeritus) and Russell Wald (Executive Director, Stanford HAI; AI Index Steering Committee).
  3. 1:11:30 – 1:12:52 — closing.

This is the talk-track for the 2026 AI Index Report. Statistical claims overlap heavily with the published report; the genuinely new material is the panel Q&A, which is the focus of the body summary below.

TL;DR

Five substantive contributions over and above the published report:

1. The absorption gap as the single observation running through the whole 2026 report. Sajadieh: “AI is scaling faster than the systems around it are built to absorb. The governance frameworks, evaluation methods, the data infrastructure, our educational institutions, all of these are meant to help us understand, prepare, and work with AI. But AI is moving far faster than they can keep up.” This is the talk-track’s compression of the report’s headline framing — the same “data does not point in a single direction; it reveals a field that is scaling faster than the systems around it can adapt” line that the co-chairs use in the published front matter. The talk frames it earlier and more aggressively: the absorption gap is the whole story; everything else is detail.

2. The acceptability-threshold research gap — Perrault’s most substantive original contribution. “I sound like a broken record on this, but the one example we have is self-driving cars. And we have been expecting self-driving cars for 20 years now … the reliability we expect out of these systems is very high because people will die if it isn’t. So what we don’t have is — if you want to take an AI technology and put it in a legal system or a medical system or a finance system, how well does it have to work to be acceptable? … I’m hoping that we’re going to get to the point where we get a sense that, you know, an application is fine at 90% reliability or it’s not fine at 99.9 reliability. So anybody who has insights into any measures of that kind — I would really love to hear about them. I have not found it.” This is the wiki’s first named research gap for what reliability threshold must AI clear to be acceptable in regulated domains? — a measurement question the AI Index cannot yet report on because nobody is measuring it. Wald reinforces with the hired-a-human-lawyer-who-gave-me-wrong-advice framing: society already accepts some error rate from humans; the unanswered collective question is at what rate it accepts the same from human-machine systems.

3. The academic-vs-industry transparency tension — Wald: “Our report notes the lack of transparency, and we sit here in an academic environment where we are driven by science. Science is driven by open publication, by open science and the means to be able to peer review and share discovery together so that we can actually create a field to understand this. And so there’s a lot that we hear from industry without transparency to understand what is happening. … Some of these companies that used to publish more and used to be pushed for this have gone into this complete mode of proprietary approaches that have made it much more difficult for verification.” The report’s empirical anchor (80 of 95 notable models released without training code; Foundation Model Transparency Index average declined this past year) gets the editorial argument layered on top of it: industry-frontier shift renders academic peer-review structurally weaker. Wald: “If this is a true future, I think academia being a partner in this is a benefit … versus the vanguard of a few companies that are going to build this and we will be purely at their mercy.”

4. The AI-Index founding mission and the facts-only discipline — Perrault’s reflection on the report’s origins. “After the first one of these reports in 2015 or 16, Yoav Shoham […] thought that there was room for a more frequent report on the facts, not projections, on the facts on the ground. And that’s what led to the AI index report. Our first report came out like a year later. It was 60 pages long … that’s remained the mission. Report on the facts. Don’t do deep analysis. Don’t do projections. Leave that to others. And don’t recommend to governments what they should be doing — that again is something we try to stay away from.” This is the wiki’s first first-person founder-account of the AI Index’s editorial discipline — relevant when comparing the AI Index against more interpretive landscape reports like FTSG Convergence Outlook 2026 (explicitly normative) or MGI Race Takes Off (interpretive arena-framework). Perrault also clarifies the editorial discipline in real time during the panel: “This is a question about the future, right? The AI index takes no position about the future. So whatever you’re getting up here is from members as you know private citizens, not of the index as an institution. In fact, the index is ancient already because it’s retrospective of what happened last year, not what’s going to happen.”

5. The World-Bank / IMF global-AI-adoption measurement initiative — Perrault discloses: “Yolanda Gil and I were at a meeting of the World Bank and the IMF a few weeks ago where they are putting together a project to measure adoption of AI worldwide by working directly with the major companies. And there were people in the room from all the big US companies and they’ve apparently all agreed to share information about their users and what they do with them and so on. So we’re looking forward to that. I don’t think that will be available in time for the next report. Uh but maybe the one the one after that. Um but that would not include open-source models.” This is the wiki’s first forward-look from inside the AI Index’s data-acquisition pipeline — a structural answer to the “the data is incomplete” limitation that runs through every Index edition. Useful as a baseline for tracking the 2027 / 2028 editions.

Sha Sajadieh’s headline framing (0:00 – 30:46)

The presentation walks the report’s chapter highlights in roughly the same order as the published report’s 15 Top Takeaways (see the 2026 AI Index Report page for the full list). The talk-specific compressions worth noting:

  • The three-numbers scale-and-speed anchor: “88% of organizations globally say they use AI in at least one business function. 53% of population level adoption of generative AI tools, and global corporate investment has more than doubled year-over-year. And again, this is not just about people using ChatGPT more often. It’s about AI moving through our daily lives, through our institutions and our markets at the same time very quickly.”
  • The lopsided system compression for the concentration story: “AI may be able to spread quickly but the ability to build it, fund it and shape where it goes next is really highly concentrated.” US $286B private investment in 2025 (+150% YoY, 23× China’s $12B); ~2,000 newly-funded AI companies (5× next country); 5,400+ data centres (10× next country); 59 notable models vs China 35 vs South Korea 8. Followed immediately by: “The US lead is real. But as you can see, even from here, it’s becoming less comfortable.”
  • The jagged frontier live-talk framing: “a top model like Gemini Deep Think will win the gold medal at the International Mathematical Olympiad, but can it tell time on analog clocks? … the top models perform and get it right only about half of the time. So the same system that looks really brilliant in one evaluation that can ace a a problem that’s, you know, a mathematical problem that’s really hard for most people uh can also look incredibly brittle in in another evaluation setting.” This is the report’s Top Takeaway #4 turned into a single sentence — useful as a teaching anchor for jagged-frontier.
  • The agents-reliability framing: OSWorld task success 12% → 66% in a year, “still failing roughly one in three attempts.” Sajadieh’s interpretation is the talk-specific addition: “when we’re asking or we we want to deploy these technologies in organizations and into our daily lives and have them take actions … reliability, safety and the concern that they’re actually going to do the job that we set out for them becomes even more important.”
  • The softening of the talent pipeline compression for ai-employment-effects — Sajadieh declines to make a “bold claim” on aggregate AI labor displacement; the data shows it’s “narrow, mixed, and complicated.” But the 22-25 software-dev cohort number (-20% from 2024) gets surfaced as “a softening of that talent pipeline” — a phrase the published report does not use; useful as a less-loaded way to talk about the empirical finding in interview / executive contexts.
  • Public sentiment chapter — Wald’s should-almost-be-chapter-one aside later in the panel echoes Sajadieh’s own framing: “this matters beyond just thinking about public sentiment because public sentiment in turn shapes adoption. It shapes regulation. It shapes how confident people are in the systems that are being deployed. So a technology like AI can be powerful and it can be really economically valuable but if people don’t trust it and they don’t trust the institutions that are managing it that becomes its own kind of bottleneck.” The data point Sajadieh emphasises: “in China 84% of respondents say AI will profoundly change their lives in the next three to five years; but in the US that’s only 38%.”
  • The country leading in development vs country leading in adoption observation — Sajadieh’s own talk-specific framing: “the country that’s leading in terms of technology development can not be the same as the country that’s leading in terms of adoption.” This is closer to an editorial claim than most AI-Index material; useful as an analytical handle for enterprise-ai-adoption.
  • The governance-fragmentation compression: US shifting “away from international cooperation and regulation towards federal preemption of state laws and deregulation”; in other countries, “new national AI laws, new frameworks come into play. And I think what’s exciting here is that more than half of those are coming from countries that previously did not participate in AI policy at all.” Plus the AI-sovereignty as organizing thesis framing: “it’s not just about regulate this, it’s about capacity. You know countries are thinking about how much agency do I have over the compute, the data, the infrastructure around this technology.”

Panel: Perrault + Wald (31:36 – 1:11:30)

Audience question — Where does the optimism on AI-as-job-creator come from? (51:24–53:14)

Sajadieh’s answer captures a talk-only observation worth recording: “organizations are now standing up teams, groups, roles around AI governance … the deployment of these systems also begs the question of how within an organization are they going to be governed and that needs its own infrastructure and it needs its own setup. So that’s one way jobs can be created.” Perrault adds the Jevons-paradox framing: “improvement in productivity may overall lead to increase in employment … if you increase the size of an industry by making it more affordable and more deployable, you will end up creating more jobs. At this stage I don’t think we know where that will be in AI but it’s been the case in the past.” Wiki-relevant: this is the first source in which an AI-Index principal explicitly invokes Jevons as the labor-effects frame.

Audience question — Where else do you see fully-autonomous agentic systems coming into enterprise / government? (54:04–56:34)

Perrault: “I have to say I don’t see any fully autonomous systems coming to much scale in any areas that are commercially important. … we don’t know how well they need to work. … with programming for example you know we have a pretty good understanding of what quality we need to get out of what Claude tells us. But the paradigm as I [understand it] is mostly that this is input to what then expert programmers look through and modify and all that. It’s not fully autonomous by any stretch of the imagination.” Plus the Air Canada chatbot example as the automation-without-acceptability-threshold counter-case. This is the operational corollary of the acceptability-threshold gap: until society can name the reliability bar per domain, fully-autonomous deployment outside narrow contexts isn’t feasible.

Audience question — Jack Clark’s 60% prediction of fully-automated AI research within 2 years (56:34–1:00:35)

Perrault: “I still have to say that the reliability problems with code in particular are so high. Nobody wants to be given code that doesn’t work or that isn’t discovered to having a flaw until years later. I mean, that’s what the whole current fuss over the new version of Claude is about. … if what he means is the there will be considerable acceleration in the generation of new AI systems with a human in the loop then sure and I’ll buy that. If it’s if we’re talking about fully automated systems I still have doubts.” Wald reframes the disagreement as a transparency-and-academic-verification argument: see TL;DR contribution #3 above. The same panel position recurs on the AI-companions / opacity audience questions later.

Audience question — AI companions and the erosion of relational capacity (1:06:26–1:09:30)

The questioner — an intimacy coach and AI governance engineer — flags a measurement gap: the report covers AI companions and the worst-harm end (children, sexual abuse), but not the slow erosion of human relational capacity through prolonged interaction. Sajadieh’s answer is editorially careful (“this decline in relational capacity is not unique to AI technology. Social media — we did not necessarily invest in increasing our relational capacity. In fact, we’ve been very reactive to what social media has done to society”) and worth recording as a precedent for how the AI Index handles questions about phenomena the data doesn’t yet support a claim on. Also names a public-attitude observation: “not everyone across the globe is sold on AI companionship … folks are skeptical. … I think because of the lessons of social media and what we have seen in the past few years people are more skeptical.”

Audience question — Public mistrust in the US: is it driven by model opacity? (1:02:52–1:06:02)

Wald pushes back on the framing: “if you’re we’re measuring global public opinion, well there’s still opacity for the models in global public opinion as well. Why is it that Singapore has such a more positive view compared to other western industrialized nations? They’re dealing with the same issues that we are. So I think there’s a lot of other factors that go beyond that particular issue. … I think there is a little bit of the halves and the have nots in some of this reporting.” Then: “to your point though on the opacity issue, I do think that is an issue personally and I do think government should push for more transparency in this space. We hear these claims and they’re claims that are measured against these brittle benchmarks … and because of the brittleness of these benchmarks, we don’t know how these how it’s going to work in actual application.” The wiki-relevant element is the halves-and-have-nots framing as Wald’s editorial position on the public-trust polarization data — relevant for responsible-ai and any future AI sovereignty concept.

Audience question — Adoption of DeepSeek vs US frontier models in industry? (1:09:47–1:12:25)

Perrault: “I do not believe there are public figures about the number of installations of Deep Seek. … what happens to open-source software — people download it, they install it on their machines and it doesn’t get recorded anywhere.” Followed by the World-Bank / IMF measurement-initiative disclosure (TL;DR contribution #5 above). The substantive observation: the AI Index’s measurement framework structurally cannot capture open-source model adoption — the data partnership Perrault is announcing only covers the major US companies’ user data, not weights downloaded and self-hosted. This is a methodological caveat worth flagging for any future open-source adoption concept page.

Why this matters in the corpus

Three sub-corpus roles for this source:

  1. The first talk-track ingest paired with a same-month report ingest — the wiki holds the published AI Index 2026 Report (partial-front-matter ingest, pp. 1–19) and now its authoritative live presentation. This is a precedent for report + talk-track dual ingests where the talk-track is treated as a distinct source whose contribution is the editorial-interpretation layer (panel framings, gap acknowledgements, forward-looks) that the published artifact doesn’t carry. Distinct from the MGI virtual event precedent in that MGI’s panel revealed material genuinely not in the report (Apollo-vs-omniscaler-capex anchor, Sastry’s omniscaler-is-not-a-conglomerate framing); the AI Index talk’s new material is editorial / interpretive rather than additional empirical.

  2. The wiki’s first named acceptability-threshold research gap — Perrault’s “how well does it have to work to be acceptable in legal / medical / finance?” is a measurement question that recurs implicitly across multiple wiki sources (the Dell’Acqua jagged-frontier paper, the jagged-frontier concept page, the AI Index 2026 robotics-vs-real-world chapter highlights). Named in this source for the first time. Plausible candidate for promotion to a concept page if a second source addresses it.

  3. The wiki’s first first-person founder-account of the AI Index’s editorial discipline — Perrault on the facts on the ground, not projections, no recommendations to governments mission. Distinguishes the AI Index epistemic stance from interpretive landscape reports like FTSG or MGI Race Takes Off in a way that informs how the wiki should weight these three reports relative to each other when synthesising the 2026 landscape.

The W&W tagging — digital-sensing/digital-scouting (the report is an annual technology-trends scout) + digital-sensing/digital-scenario-planning (the talk’s where AI is heading framings feed organizational scenario inputs) + digital-seizing/balancing-digital-portfolios (the acceptability-threshold gap is fundamentally a portfolio-balancing question — which AI applications belong in legal / medical / finance, at what reliability bar?) + contextual/external-triggers (public sentiment, sovereignty politics, governance fragmentation as macro forces organizations respond to) — mirrors the MGI virtual event ingest’s four-cell profile, which is the right comparator for this source’s W&W footprint.

ASR notes

  • YouTube’s auto-generated transcript ASR-clean was light: spelled-out timestamp prefixes injected by YouTube’s screen-reader (e.g. “2 minutes, 8 seconds and they’re part of a broader committee”) stripped at clean-up; otherwise transcript left as captured. Surface artifacts: “Ray Perau” in the ASR (clearly Raymond Perrault on the Stanford HAI Steering Committee); “Ozade” as a questioner name (unverified).
  • The ASR-channel attribution “Stanford HAI” is the canonical channel name and goes into author: per §Source-page conventions specific to videos.
  • Caption track quality flag: kind: asr (auto-generated, not human-curated).

Linked entities and concepts

Entities (promoted, source_count bumped):

  • Sha Sajadieh — second source; promotes her source_count from 1 to 2; bumps last_confirmed and accessed_at.
  • Stanford HAI — third source (also publishing channel for this video); bumps source_count, last_confirmed, accessed_at.
  • Raymond Perrault — second source (was 2026-04-30 AI Index report); panel principal here; bumps.
  • Russell Wald — second source; panel principal; bumps.
  • AI Index — third source; bumps.
  • Yolanda Gil — mentioned by Perrault as joint World-Bank / IMF emissary; small bump.

Mentioned but not promoted to entity pages on this ingest:

  • Jack Clark (already an entity via prior multi-source presence) — referenced in panel for the 60% prediction of fully-automated AI research in 2 years; reverse-edge would be one-step-removed (it’s a citation-by-name, not a primary source-to-source relationship).
  • Yoav Shoham (already an entity) — Perrault credits him with the founding-momentum push for the AI Index in 2015–16. Useful as anchor for One Hundred Year Study on AI if/when promoted.

Concepts (last_confirmed bumped, no content change unless flagged):

  • enterprise-ai-adoption — 88% organizational adoption; the governance-roles-as-new-jobs framing on the audience question.
  • generative-ai — 53% population adoption.
  • ai-employment-effects — 22-25 software-dev -20% from 2024; the softening of talent pipeline compression.
  • jagged-frontier — IMO gold + analog-clock framing; OSWorld 12%→66% with still-1-in-3 failures.
  • ai-benchmarks — top-4 Arena Elo within 25 points (was 97); benchmark saturation against human baselines; brittleness per Wald.
  • responsible-ai — 362 vs 233 incidents; FMTI decline.
  • foundation-models — 80/95 notable models released without training code; FMTI average decreased; industry sets the frontier at 91%.

Concepts the wiki may want to promote following this source (deferred):

  • AI acceptability threshold — Perrault’s named gap (what reliability bar is required for AI deployment in legal / medical / finance?). Single-source; defer to second-source promotion.
  • AI sovereignty — Sajadieh’s framing of compute / data / infrastructure agency as governments’ organizing thesis. The 2026 report’s analytical framework section also touches this; the talk’s compression here is editorial. Defer to second-source.

Source

Reading scope

Full ~73-minute transcript read end-to-end during ingest, including audience-participation segments. Three substantive panel framings (acceptability threshold, academic-vs-industry transparency, World-Bank / IMF measurement initiative), one audience-participation correction (Wald’s pushback on the opacity-drives-mistrust framing), and the Sajadieh editorial compressions of report headlines (88% / 53% / $286B / 2.7% / 89% / 362 / 80/95 / 84% / 31% — see TL;DR section above) are surfaced in this page. Statistical claims that overlap with the published AI Index 2026 Report are not re-listed exhaustively; cross-reference the report page for the full Top-15 Takeaways list.

Relationships

Supports

  • Artificial Intelligence Index Report 2026 — **Talk-track for the published report.** Same author (Sha Sajadieh, EiC) presents the report's headline framing live; statistical claims overlap heavily (88% organizational adoption, 53% population GenAI adoption, $285.9B US private investment, 2.7% top-US/top-China gap as of March 2026, 89% AI-talent-inflow decline since 2017 with 80% in the last year, 362 AI incidents up from 233, 80/95 notable models released without training code, 84% of US high schoolers using AI for school work + <50% schools with policies + 6% teachers feeling prepared, US 31% trust in government to regulate effectively — lowest of any country surveyed). The talk-track adds material not in the report itself: a 42-minute panel Q&A with Raymond Perrault (co-founder, steering-committee co-chair) and Russell Wald (Stanford HAI Executive Director) covering AI-Index founding mission (*facts on the ground, not projections, no recommendations to governments*), the **acceptability-threshold research gap** (Perrault: *we know self-driving cars need very high reliability because people die; we don't know how reliable AI must be to be acceptable in legal / medical / finance*), the academic-vs-industry transparency tension (Wald's argument that frontier-lab proprietary mode hampers academic peer-review/verification), and Yolanda Gil + Perrault's World-Bank/IMF measurement initiative — agreement-in-principle from major US companies to share user/usage data for global-AI-adoption measurement, *not in time for next report*
  • The race takes off in the next big arenas of competition (MGI virtual event) — **Sibling live-event-of-annual-landscape-report pattern at major-institution altitude.** Both render an annual landscape report into a live executive briefing + panel Q&A format; identical four-cell W&W profile (`digital-sensing/digital-scouting` + `digital-sensing/digital-scenario-planning` + `digital-seizing/balancing-digital-portfolios` + `contextual/external-triggers`). MGI emphasises the *arena-creation* axis (18 future arenas adding $18T market cap, nine omniscalers redistributing per-arena value at ~20×); AI Index emphasises the *absorption gap* axis (capability moving faster than governance / evaluation / data infrastructure / education can keep up). Both converge on the US-China narrowing, infrastructure concentration (omniscaler-capex / TSMC-foundry), and the public-trust gap as 2026's organizing facts.
  • Convergence Outlook 2026: The new timing of creative destruction — **Two annual 2026 horizon-scans converging on the same empirical landscape from different vantages.** FTSG / Webb anchor the *convergence-as-power-redistributor* thesis (rule 3: convergences redistribute power and value across — not just within — industries) and the *creative-destruction-from-within* normative argument. The AI Index talk anchors the same empirical phenomena via *facts on the ground, not projections*: US-China gap narrowing (FTSG's *redistribute power* rule playing out at the model-frontier altitude), industry concentration of frontier models at 91% (FTSG's *capital concentrates rapidly* enabling condition operationalised), the public-trust polarization (FTSG's *legitimacy of existing order erodes* enabling condition empirically measured). Different framings, same 2026 landscape.

Graph View

Backlinks