Our view of global AI progress. The numbers and the reasons.
The AIAF Index is our AI-assisted editorial assessment of selected evidence about AI progress worldwide. We look across models, agents and practical applications—not just one company or country.
Global scope, selective evidence. We cannot observe every AI system. The Index expresses AIAF’s judgment; it is not an official global benchmark, a safety rating or a percentage of the way to AGI.
Two ways to understand the same subject
The Index gives a compact editorial assessment across four progress dimensions, with a separate amber WTF surprise meter. Progress Watch explains what AI can do now, what still fails and what changed. Read them together: the number invites a question; the evidence helps answer it.
AIAF OVERALL INDEX
72.3/100
Component assessment date
Method v2 · Four dimensions · Evidence under review
Method v2 · Four-part progress Index: (81 + 68 + 76 + 64) ÷ 4 = 72.25, displayed as 72.3/100. Each progress dimension contributes 25%. WTF is a separate surprise meter with no weight in this average. Equal weighting is an editorial choice, not a scientifically established weighting.
Method changed 17 September 2026: surprise is now separate from progress. The previous five-part 75.4/100 and original unsupported 72/100 remain in the history. The new 72.3 is a recalculation of unchanged 14 September judgments—not a decline in AI progress. Component evidence remains under review; scores across the two methods are not directly comparable.
Capability81
What this means for you
Can it do your task?
Look for a demonstration on work similar to yours—not just an impressive showcase.
Background reading; not validation of this historical score.
Adoption and real-world impact64
What this means for you
Does it improve your work?
Look for measured time savings, better quality or lower costs.
Score evidence under review.
Separate from the progress Index · No composite weight
WTF · The surprise meter
88/100
Historical editorial judgment · 14 September 2026 · Evidence under review.
How far did verified reality exceed credible expectations? This measures surprise, not capability, adoption, usefulness or safety. As people adjust their expectations, surprise can fall even while AI improves.
Hypothetical example: experts expect AI to help propose treatments; instead, a system achieves an independently verified medical breakthrough much sooner than their dated forecasts. The unexpected leap—not merely the importance of the achievement—creates the surprise. This is an illustration, not a claim that it happened.
A future scored assessment needs a named, dated expectation, linked evidence of the actual result, limitations and a written rationale. The historical 88 lacks that validation; it is not a precise measurement.
Four progress dimensions and a separate surprise meter
Capability: the breadth and difficulty of tasks AI demonstrates it can perform.
Autonomy: how much of a defined workflow AI completes, with intervention and failure rates disclosed.
Reasoning / reliability: consistency in solving unfamiliar problems, including failed attempts.
Adoption and real-world impact: how widely AI is deployed across sectors and regions, and measured changes to time, cost, quality and people's work—including unequal access and harms.
WTF factor / surprise: an explicitly subjective editorial view of how unexpected a development is relative to cited, dated expectations. Surprise does not establish competence or safety.
The intended scale is 0–100, with higher scores expressing a stronger editorial assessment within a dimension. A score of 81 does not mean 81% of all tasks can be done, and 72.3 does not mean 72.3% of the journey to AGI. The historical snapshot lacks published anchors sufficient to explain why each exact number was chosen; it should not be used for precise comparisons.
Before the next scored edition
The composite formula is now the equal-weight mean of the four progress dimensions (method v2; 25% each), excluding WTF surprise, displayed to one decimal place. If any element is not assessed, no composite is published. Before changing component scores, we will publish the assessment scope, evidence cutoff and dimension-specific scoring anchors. Each dimension must include linked sources, conditions, failures, confidence and a written scoring rationale. Missing evidence will be marked “not assessed”, never replaced with an invented value.
Updates must show what changed and why. Method changes will be versioned; the historical launch score will not be silently recalculated or presented as a comparable baseline. No new scored-release date is confirmed.
Scope clarification: “Real-world impact” is now labelled “Adoption and real-world impact” to make deployment explicit. The historical score of 64 is unchanged and has not been reassessed against the expanded description.
Imagine the ceiling · Not a prediction
What would 100/100 look like?
Not just a smarter chatbot. A world in which highly capable AI is deeply embedded in work, decisions and everyday systems.
100/100HYPOTHETICAL CEILING
Progress is not the same as adoption. The AIAF Index combines capability, autonomy, reliability, adoption and impact. Subjective surprise is displayed separately and does not affect the overall score. It does not measure the percentage of people using AI or the percentage of the journey to AGI.
Proposed qualitative anchors—not validated thresholds or a reassessment of today’s scores.
Dimension
What 100 would look like
What we would need to see
Capability
AI performs a very broad range of economically important cognitive tasks at strong human-expert level, including unfamiliar situations.
Independent, representative tests across professions, languages and difficult new tasks—not selected demonstrations.
Autonomy
Agents complete complex, extended workflows, coordinate tools and recover from failures with very little intervention.
Repeated end-to-end results over long periods, with intervention rates, failures and permission boundaries disclosed.
Reasoning / reliability
Performance remains dependable in novel, ambiguous and adversarial conditions; consequential errors are rare and measured.
Reproducible testing, calibrated uncertainty and evidence that reliability holds outside the original test setting.
Adoption and real-world impact
AI is routinely used across major sectors and diverse economies, with substantial, independently verified effects on work and daily life.
Representative deployment and outcome data: geographic reach, sector penetration, productivity, access and effects on workers—including harms.
Separate WTF / surprise meter
Verified developments dramatically exceed credible, previously documented expectations. Not included in the overall Index.
A published reference point and evidence of the surprise. This remains subjective: surprise can fade while capability and adoption grow.
Not a safety certificate. More capability can coexist with harm, displacement and concentrated power.
Not proof of consciousness. Performance does not establish subjective experience, AGI or ASI.
Not an inevitable destination. No arrival date or probability is implied. Progress and surprise need not peak together.
A ceiling, not a finish line. Under the current equal-weight formula, 100 overall would require all four progress components to reach 100; the separate WTF meter could be low or high. These descriptions are an editorial thought experiment; measurable thresholds and intermediate scoring anchors still need validation. They do not validate the historical component judgments or the recalculated 72.3/100. Any revised methodology will be versioned, with old scores preserved and comparability explained.
Review and update policy
Reviewed weekly. Changed only with evidence.
Scheduled review: every Monday, Malaysia time. The first review was due on and was completed late on . The latest review was completed on . The next review is due . Our editorial checks may identify significant developments for an additional review between weekly reviews.
A review does not automatically change a score. Each completed review will publish its date, sources examined, outcome and next review due. The score's assessment date remains separate from the date of the latest review.
Component assessment date: 14 September 2026. Method v2 adopted 17 September 2026. Current overall: 72.3/100 using four unchanged progress components; separate WTF: 88/100. Neither has been evidence-validated. The former five-part 75.4/100 is retained in the history, not a comparable trend point. New component scores require scoring anchors and traceable evidence.
Review record
19 October 2026 — weekly evidence review
Evidence reviewed on 19 October, Malaysia time: NIST’s 16 October announcement and October version 1.0 joint incident-readiness document; METR’s 14 October assessment proposal; Anthropic’s 15 October evaluator announcement, including its 16 October clarification; Google Research’s 7 October workplace-learning report; and UNESCO’s 5 October museum report about a 17 September event. Older evidence is context, not new news.
Capability: METR proposes assessments of harmful behaviour, monitoring and AI-assisted research acceleration. This is a route to evidence, not a new independently validated capability result. Gap: comparable external task results and dimension-specific scoring anchors.
Autonomy: NIST, NSA and CISA published ten voluntary practices for incidents during agent training and evaluation, spanning preparation through reporting. Guidance for managing unauthorised actions is not a measurement of how frequently they occur. Gap: representative incident denominators and measured control effectiveness. The source’s “super intelligence” terminology is not treated here as proof that ASI or consciousness has been achieved.
Reasoning / reliability: Announced evaluator access can improve scrutiny, but does not establish reliability. Anthropic’s clarified account says it ultimately determines which information falls within predefined redaction categories; METR retains certain rights to describe access and information received. METR’s broader proposal explicitly does not certify safe development. Gap: published findings, limitations and evidence that remediation worked.
Adoption and real-world impact: Google’s three-month trial with 133 patent lawyers reports improved AI-assisted drafting, but no discernible average unassisted judgment gain among juniors. UNESCO describes museum case studies and the need for provenance, curatorial judgment and institutional capacity. Neither establishes global adoption, general job displacement or representative cultural outcomes. Gap: broader longitudinal outcome data, including harms and unequal gains.
WTF / surprise, separate from the Index: New oversight practices and access arrangements are consequential, but do not supply a predeclared expectation baseline against which to score surprise. Gap: dated expectations and traceable surprise anchors. No automatic increase follows an alarming headline or a new institutional label.
Outcome: scores not recalculated; methodology/evidence validation pending. Current method v2 display remains 72.3/100; WTF remains 88/100 separately. Component assessment date remains 14 September 2026. Historical methods and forecast originals are unchanged. Next review due .
Practical implication — AIAF analysis: Ask who can stop an agent, what evidence survives an incident and who authorises restart. Require measured outcomes before treating either a safety partnership or improved assisted output as proof of durable human skill or dependable control.
12 October 2026 — weekly evidence review completed: reviewed OpenAI’s 7 October GPT-6 release and deployment system card, and Google Research’s 7 October report on a three-month randomized trial with 133 patent lawyers at 11 firms. OpenAI says GPT-6 is rolling out across ChatGPT and reports stronger performance on selected internal evaluations, while classifying the released models as “High” capability in cybersecurity and biological/chemical domains but below its “High” AI self-improvement threshold. The same system card reports statistically significant regressions in some safety categories and warns that changing policies, graders and datasets limit comparisons. Google reports that AI access improved assisted patent-drafting quality, but junior lawyers showed no discernible average gain on a later unassisted judgment task; the study is narrow, short and not representative of every profession. Together, these sources inform capability, reliability, adoption and workforce-impact questions, but provider deployment counts, internal evaluations and one occupational field trial do not validate the historical component scores or establish population-wide outcomes. Scores not recalculated; methodology/evidence validation pending. Display remains 72.3/100 under method v2; WTF 88/100 remains separate. Component assessment date remains 14 September 2026. Next review due 19 October 2026.
5 October 2026 — weekly evidence review completed: reviewed Mastercard’s 30 September Agent Pay Trust Framework and intelligence-services announcement, the W3C Digital Credentials Working Draft of 4 September, and OpenAI’s 29 September description of always-on Dots and their safeguards. The sources show continuing commercial adoption of AI agents, more explicit separation of identity from delegated authority, and proposed controls for credentials, execution and behavioural risk. They do not provide representative deployment outcomes, independent reliability measurements or validated evidence for the historical component scores. A transaction risk signal is not proof of user authorization; the W3C document remains a draft focused mainly on credentials about people. Scores not recalculated; methodology/evidence validation pending. Display remains 72.3/100 under method v2; WTF 88/100 remains separate. Component assessment date remains 14 September 2026. Next review due 12 October 2026.
28 September 2026 — weekly evidence review completed: reviewed the latest version (v3, revised 22 September) of the arXiv survey on recursive self-improvement and Google Research’s 24 September multi-agent video announcement, alongside the evidence recorded on 23 September. These sources describe proposed improvement frameworks and company-reported research results; neither establishes general autonomous self-improvement or validates the historical Index components. The framework highlights unresolved needs for independent evaluation, durable gains across later tasks, explicit human release authority and rollback limits in physical or clinical settings. Scores not recalculated; methodology/evidence validation pending. Display remains 72.3/100 under method v2; WTF 88/100 is separate. Component assessment date remains 14 September 2026. Next review due 5 October 2026.
23 September 2026 — overdue evidence review completed: reviewed Google’s generative-UI study and technical report, Joby’s autonomous-flight announcement and human-supervision limits, Anthropic’s embedded-evaluation proposal, Palo Alto Networks’ continuous frontier-AI defence announcement, and Alibaba’s reported model-optimisation loop and infrastructure roadmap. These sources are relevant to capability, autonomy, reliability and adoption, but the evidence is selective, much of it is company-reported, and the historical component scores still lack validated scoring anchors and traceable rationales. Scores not recalculated; methodology/evidence validation pending. Current display remains 72.3/100 under method v2, with WTF 88/100 separate. The missed 21 September publication was not backdated; its cause is not established in the public record. Next review due 28 September 2026.
17 September 2026 — owner-approved method v2: separated WTF from progress. (81 + 68 + 76 + 64) ÷ 4 = 72.25, rounded to 72.3/100. Four dimensions have 25% weight each; WTF has zero composite weight. No components reassessed. This is a methodology change, not evidence of regression. Previous method v1: (81 + 68 + 76 + 64 + 88) ÷ 5 = 75.4/100.
15 September 2026 — calculation correction: replaced 72 with 75.4 using (81 + 68 + 76 + 64 + 88) ÷ 5. Components unchanged; no new progress assessment. The original 72 had no documented formula.
15 September 2026 — review policy established. Schedule configured; first weekly evidence review pending. This is a policy update, not a completed score assessment. If a review is delayed, the actual completion date and reason will be disclosed here.
AI Progress Watch
What AI can do now. What still fails. What changed.
We follow selected, verifiable developments worldwide. Coverage includes models, agents and applications across organisations and countries; public evidence is uneven and coverage is not exhaustive.
What changed: the development, date and task involved.
What the evidence shows: source links, test conditions and results.
What still fails: limitations, failed attempts and missing evidence.
Why it matters: AIAF's labelled interpretation for life and work.
Index relevance: which dimension it informs. A score changes only after the scoring review, not automatically with a headline.
The 15 September briefing helps explain autonomy and reliability: what did an evaluation establish, what were its limits, and when should a person take over? It is an evidence explainer, not a new scored assessment or breaking-news claim.
The initial numerical Index proposal was followed by a Progress Watch-only presentation to clarify its scope. Following owner review, both now appear together. The 14 September scores remain historical editorial judgments with the original limitations disclosed. The previously proposed 30 September scored-beta target remains withdrawn; the overall calculation was subsequently corrected to an equal-weight 75.4/100 on 15 September, while component evidence remains under review.