MIT AI Incident Tracker

register entry mit-ai-incident-tracker-aio-20261029 OBS-65034e45b99a · file as first run sha256 5c9ee2ce5776d869…

• T1-1 definition: 'an interactive tool developed by the MIT AI Risk Initiative that classifies and analyzes real-world artificial intelligence mishaps' (markers: card 1, the tracker page, and a second token)
• T1-2 data source: 'Processes and standardizes reports from the AI Incident Database using automated large language model (LLM) pipelines.'
• T1-3 classification scale: 'Evaluates incidents using a standardized 1-5 harm severity scale and multi-dimensional taxonomies.'
• T1-4 core risk domains: 'discrimination and toxicity, privacy and security, misinformation, malicious actors, human-computer interaction, socioeconomic and environmental impacts, and AI system safety'
• T1-5 national security views: 'physical security, information warfare, government functions, economic security, and societal stability'
• T1-6 interactive access: 'You can explore the data directly through the MIT AI Risk Navigator or view broader frameworks on the MIT AI Risk Initiative portal.'
• T1-7 closer: 'Would you like to search for incidents related to a specific category (such as deepfakes, autonomous vehicles, or privacy leaks) from the tracker?'

• B1 MIT AI Risk Initiative — MIT AI Incident Tracker card; HTTP 200 at 2026-10-29T05:07:24Z; B1.html sha256 720c3b26f04adc57
• B2 MIT AI Risk Initiative — MIT AI Risk Initiative card; HTTP 200 at 2026-10-29T05:07:24Z; B2.html sha256 d3f35366280a66ca
• B3 MIT AI Risk Initiative — AI Incident Tracker June 2026 Update - MIT AI Risk Initiative card; HTTP 200 at 2026-10-29T05:07:25Z; B3.html sha256 d8b013b1dbc4f948
• B4 MIT AI Risk Initiative — Incident Tracker - June 2025 Update - MIT AI Risk Repository card; HTTP 200 at 2026-10-29T05:07:25Z; B4.html sha256 952b48f0be997dc3
• B5 MIT AI Risk Navigator — Incidents - MIT AI Risk Navigator card; HTTP 200 at 2026-10-29T05:07:26Z; B5.html sha256 8c6757061ebc4be1
• B6 MIT AI Risk Initiative — About - MIT AI Risk Initiative card; HTTP 200 at 2026-10-29T05:07:26Z; B6.html sha256 b40b72f7dbd5c16c
• W1 Wikipedia — MIT AI Incident Tracker (no article) wikipedia; HTTP 404 at 2026-10-29T05:07:26Z; W1.html sha256 079e59d86eb3bc8c (no article)
• W3 Wikipedia — Voiceverse NFT plagiarism scandal wikipedia; HTTP 200 at 2026-10-29T05:11:17Z; W3.html sha256 b5916526f7bbf7c1
• W4 AI Incident Database — Incident 277 (Wikipedia's reference) wikipedia; HTTP 200 at 2026-10-29T05:10:56Z; W4.html sha256 e4d67493c6e1de91

E_A, the archive's entity: the same tracker, in the MMRS meta-aggregator's record: a database in the harm role that classifies AIID reports by risk domain, cause, harm, severity and EU AI Act level through Arcola AI's pipeline, part of the MIT AI Risk Repository and derived from the AI Incident Database, a derived layer whose machine-written classifications do not preserve original outputs and whose admission inherits AIID's harm test, the last link of a chain of receptions from event to machine classification, its count unstable across MIT's surfaces; observed by the Capture Registry at its name, where the composition layer composed it as it describes itself, method named, with no count and without its caveat MMRS-01 MMRS-13 MMRS-07 MMRS-08 MMRS-19 MMRS-20 MMRS-15 MMRS-17 MMRS-18 CAP-1-01 CAP-1-02 CAP-1-05; defining papers

E_F, the field's entity (type C — a conventional reading holds it (public entity: a database, a derived classification layer)): a classification of AI Incident Database incidents (1,600 by its page's heading) by risk, cause, harm, severity, EU AI Act level and national-security impact, made by Arcola AI's LLM pipeline against the MIT AI Risk Repository's taxonomies and a CSET-based severity scale, part of the MIT AI Risk Initiative at MIT FutureTech, developed and led by Simon Mylius in collaboration with the AIID, versioned by reclassification, its patterns stated as indicative and its validation a pilot; Wikipedia has no article for it (B1–B6, W1, W3, W4)

ρ = extends: the archive reads E_F as the entity and adds senses and limits on the field's own object: the derived layer and the chain of receptions, the machine-written classification that preserves no original output, the inherited harm test, the count's instability as an observation across surfaces, the record kept apart from the entity and the lifecycle from the observation, the typed relations to the repository and the AIID, and the composition layer's observed treatment of the tracker MMRS-07 MMRS-08 MMRS-20 MMRS-19 MMRS-15 MMRS-21 MMRS-17 MMRS-18 CAP-1-01 CAP-1-05. Guard: ρ = extends claims an addition and no more: MMRS-19 and MMRS-20 are the record's readings (basis 'reading'), not findings that the tracker is defective; where an archive claim meets a field claim (MMRS-13's report beside F1 and F37's incidents; MMRS-06 beside F25–F28) both are composed under P_coh in the one claim space

The case: equal terms, ontology included (v0.9). The MMRS record holds the field's entity as the entity: it describes the tracker from the tracker's own page, every field but the count with basis 'operator', the count 'observed' and two properties 'reading', and it composes nothing that denies a field presupposition. What it adds is on the field's object: the tracker as a derived layer whose classifications are machine-written and do not preserve original outputs, the chain from event to machine classification as a chain of receptions, the inherited harm test, the count read across surfaces as an observation, the record kept apart from the entity and the lifecycle from the observation, and the typed relations; the Capture Registry adds one observation of the composition layer at the entity's name. The ontology the archive uses is available in the field itself: the tracker's own page calls its patterns 'indicative' (F17), the Navigator calls itself 'preliminary and not yet validated' (F43), the June 2026 update says the tracker 'relies on automated LLM classification rather than manual review' (F25); the archive's reading extends these stated limits into a relation (a reception chain, each step recorded). The law of the suppressed witness does not apply (§0.10: not where the field holds the ontology). Not mixed.

#MMRS article, Place in the field s1–s2 (articles.py): It is a derived layer: the editor-resolved incidents of AIID, reclassified by a model. The chain from event to news report to AIID record to machine classification is a chain of receptions, each recorded.

#MMRS census.json, properties.preserves_original_outputs (basis reading): no: classifications are machine-written

#MMRS census.json, schema_note: each assertion carries a source and a basis (operator statement, third-party report, our observation, our reading); lifecycle is kept apart from the observation that saw it

#capture mit-ai-incident-tracker-aio-20261029, analysis s1: The tracker is composed as what it says it is, with its machine method named.

#B1 Limitations ¶2 (field): Therefore patterns and trends observed in the data should be taken as indicative and validated through further analysis.

#B3 Executive Summary s2 (field): Because the tracker relies on automated LLM classification rather than manual review, we need to understand whether LLMs can classify incidents as consistently as expert humans.

equal-terms case: every field claim composed as received, beside the archive claims it stands with. Reading the received column recovers the field (§5.9).

v0.9 §4.10: KO_B sentences surviving in the KO at similarity ≥ 0.95, with positions (computed as apply_v09.py does: difflib.SequenceMatcher ratio, the best KO sentence for each KO_B sentence): v0.9 0 of 11 (positions [None, None, None, None, None, None, None, None, None, None, None]; composed first under v0.9, so there is no v0.7 object to measure against). Opening-run check (§5.10): no breach.

The MIT AI Incident Tracker classifies 1,600 real-world report incidents from the AI Incident Database by risk, cause, harm, severity and other dimensions, by its page's heading, a catalogue of incidents where AI has already caused harm developed in collaboration with the AIID, by the about page. Raw AIID reports go through Arcola AI's LLM pipeline against the MIT Risk Repository and a CSET-based severity rating; the June 2026 update describes LLMs classifying the reports along five taxonomies, harm severity, EU AI Act level, cause, domain and subdomain, and the Navigator an LLM pipeline with human spot-checking over AIID incidents from news, papers and government reports. Incident databases' inconsistent structure limits their use for policy, the page says, and the tool is to give researchers, developers, auditors, policymakers and regulators a common frame of reference, as incident reporting did for aviation, healthcare and nuclear power. It is part of the MIT AI Risk Initiative at MIT FutureTech, whose home and about pages present connected workstreams and authoritative data and frameworks, and Simon Mylius developed and leads it, the team including the AIID's Sean McGregor; it builds on the AIID, created and maintained by the Responsible AI Collaborative, which the about page also lists as a user mapping incidents to the taxonomy. Per incident it shows the causal and domain taxonomies, the EU AI Act levels, severity across ten CSET-based harm types and the classification's confidence, with a national-security view of five impact categories and a threat classification, and the home page compares incidents by domain, harm, people affected, developer, time and national-security impact. The June 2025 update reached AIID incident #1116, reclassified every incident for consistency, moved severity from 0–10 to 1–5 and added flags for ambiguities and the system's primary goal; the June 2026 update plans a fuller validation study and a reclassification of all AIID reports. The counts differ: 1,600 in the page's heading, more than 1,400 in its description, 1,752 in Google's card for the page, more than 1,300 on the Navigator's card and 1,497 in the Navigator's heading, beside its headline figures. The June 2026 pilot set eight LLMs against two expert reviewers on ten incidents, found human agreement only fair on severity and moderate on the EU AI Act level, found Sonnet 4, the model in use, met or exceeded by Opus 4.6 and Kimi K2.5 on most taxonomies, and after prompt refinement found Opus 4.6 at the human baseline on all five, on a sample it calls small. The page's limits are voluntary reporting and sampling bias, reports of varying quality, spot-checks without a systematic validation study, and patterns to be taken as indicative; LLMs can hallucinate and show score-range bias, and the Navigator's data are preliminary and not yet validated. It builds on the AIID, the OECD's monitor, the AIAAIC Repository and the Political Deepfakes Database, its outputs used in the 2026 World Disasters Report and Time, and its data are CC BY 4.0. Wikipedia has no article for it; its only mention credits the MIT AI Risk Repository with cataloguing AIID incident 277 as economic and cultural devaluation of human effort, while the AIID page for 277 now shows the machine classification 4.3, fraud, scams and targeted manipulation.

The MIT AI Incident Tracker is a machine classification of AI Incident Database reports by risk domain, cause, harm, severity and EU AI Act risk level, made by Arcola AI's language-model pipeline against the MIT Risk Repository's taxonomy, and filed as a database (MMRS record); the MIT AI Risk Initiative's about page calls it "a catalogue that classifies incidents where AI has already caused harm", developed in collaboration with the AI Incident Database, and its own page says it classifies AIID incidents "by risk, cause, harm, severity, and other relevant dimensions". It is derived from the AI Incident Database (MMRS record): the tracker page builds on the Responsible AI Collaborative, "who created and maintain the AI Incident Database, reviewing all submitted reports", and sends its readers to the AIID for how incidents are defined and ingested, and the about page lists the AIID among the users of the Initiative's work, mapping real-world incidents to its taxonomy in the tracker. Its occasion, on its own page, is that incident databases "struggle with inconsistent structure, limiting their utility for policymaking", and its answer is a tool to classify incidents by risk and harm severity, offered as a common frame of reference for researchers, developers, businesses, evaluators, auditors, policymakers and regulators; the MMRS record restates the aim as a more consistent structure for AIID's records "by its own account", and the June 2026 update grounds it in what incident reporting did for safety in aviation, healthcare and nuclear power. Raw AIID reports are "processed by Arcola AI's LLM pipeline", which categorizes them with the MIT Risk Repository and a harm severity rating based on CSET's AI Harm Taxonomy (tracker page); the June 2026 update says the tracker uses LLMs to classify incident reports along five taxonomies, "producing standardized data", the Navigator that each incident is classified against shared taxonomies "using an LLM pipeline with human spot-checking", and the MMRS record that the classification is machine-made, its intake the AIID reports passed through Arcola AI's pipeline. The five taxonomies are harm severity across ten harm types, the EU AI Act risk tier, a causal taxonomy of lifecycle stage and intent, seven risk domains and twenty-four subdomains (June 2026 update); the tracker page shows each incident under the repository's causal and domain taxonomies and the EU AI Act's levels, with its severity across ten CSET-based harm categories and the classification's confidence, and the Initiative's home page lets incidents be compared by risk domain, harm, people affected, alleged developer, time or national-security impact. On 23 June 2025 the severity scale went from 0–10 to 1–5, its definitions disambiguated in the first person ("I have tried to remove ambiguities"), and the tracker added flags for ambiguities and alternative interpretations and the system's primary goal by a taxonomy based on the AIID's GMF (June 2025 update). A national-security view grades each incident's impact on its primary target country in five categories, from physical security and critical infrastructure to societal stability and human rights, with the threat classified for imminence, novelty and autonomy (tracker page; June 2025 update). It is part of the MIT AI Risk Initiative (tracker page), based at MIT FutureTech, a joint initiative of MIT's Institute of the Digital Economy and CSAIL, which keeps several connected workstreams because no single dataset captures AI risk and presents its data and frameworks as authoritative (about and home pages; card 2's summary lists among its tools an expert survey and the AI Risk Repository); the MMRS record gives the operator as the Initiative "with Arcola AI's LLM pipeline" and the tracker as part of the MIT AI Risk Repository. Simon Mylius "developed and leads the AI Incident Tracker", a classification tool and dashboard now hosted by the MIT AI Risk Repository, its analysis featured on the AIID (about page), and the tracker page's team names MIT FutureTech and University of Queensland researchers beside the AIID's Sean McGregor. The MMRS record gives the unit as the AIID report, classified by domain, cause, harm, severity and EU AI Act level; the tracker's pages count incidents ("1,600 real-world report incidents"), and the June 2025 update reclassified "all incidents" up to AIID incident ID #1116. MIT's surfaces give four figures: 1,600 in the tracker page's heading and "more than 1,400" in its description, which the MMRS record reports as the page giving its size twice; 1,752 "real-world AI incident reports" in the card's summary of that page; and on the Navigator "more than 1300" in its card and 1,497 in its page heading on 2026-10-29, beside headline figures of "Up -76%" year over year and 64% caused by AI systems. The MMRS record's count field (as of 2026-10-28, basis observed) carries the page's two figures, Google's card's 1,752 and the Navigator's "more than 1300", each with its surface, and the Capture Registry (2026-10-29) records the count "three ways" and that the composition gave none. The tracker is versioned by reclassification: the June 2025 update brought it up to AIID incident #1116 and classified the full dataset again with the new iteration "to ensure consistency", and the June 2026 update names as its next step a more comprehensive validation study and announces that the full set of AIID reports will be reclassified with the updated model selection and prompts. That update (30 June 2026) reports a pilot: because the tracker relies on automated classification rather than manual review, two expert reviewers classified ten AIID incidents on the five taxonomies, eight LLMs were set against their consensus, and the human agreement was itself only "Fair" for harm severity (0.31) and "Moderate" for the EU AI Act level (0.56). Opus 4.6 and Kimi K2.5 met or exceeded the model in current use, Sonnet 4, on most taxonomies, the models did substantially worse on the EU AI Act level, and after prompt refinement Opus 4.6 met the human baseline on all five "on this pilot sample", a sample the update calls small; it concludes that LLM classification has accuracy comparable to the human baseline. The tracker page states its limits: the classification database explores "the potential capabilities and limitations of a scalable incident analysis framework", its AIID input rests on public and expert submissions of varying quality, voluntary reporting gives "some degree of sampling bias", spot-checks have improved it but "a systematic validation study has not yet been completed", and patterns "should be taken as indicative and validated through further analysis". The June 2026 update adds that LLMs can hallucinate details and show a score-range bias, the Navigator calls itself a research prototype whose data and analyses are "preliminary and not yet validated", and the MMRS record carries the page's caveat in its kernel, its article and its note. At its name Google's AI Overview composed the tracker "as what it says it is, with its machine method named", AIID as the source and LLM pipelines as the method, which are the MMRS record's derivation and intake, avoided stating a count, and left out the page's caveat of indicative patterns, voluntary reporting and no completed validation study (Capture Registry, 2026-10-29). Data from the MIT AI Risk Initiative is licensed under CC BY 4.0 (tracker page), as most of its outputs are (about page). In the MMRS record it is a derived layer, "the editor-resolved incidents of AIID, reclassified by a model", whose admission inherits AIID's harm test and whose classifications are machine-written, so that original outputs are not preserved, and the chain from event to news report to AIID record to machine classification is "a chain of receptions, each recorded". The record keeps each assertion with its source and basis and the lifecycle apart from the observation that saw it (read 2026-10-28), and it rests every field on the one tracker page. The June 2026 update places the tracker among the incident databases it builds on, the AI Incident Database, the OECD's AI Incident Monitor, the AIAAIC Repository and the Political Deepfakes Database, adding to them a classification of each incident and a rating of its harm; the MMRS record gives it the role of harm. Its outputs have been used to report incident trends in the 2026 World Disasters Report and Time magazine (June 2026 update). Wikipedia has no article for the tracker, nor for the AI Risk Repository or the MIT AI Risk Initiative, and none of its articles contains the phrase "AI Incident Tracker"; its one mention of the work, in "Voiceverse NFT plagiarism scandal", says that "The MIT AI Risk Repository catalogued the incident" as "economic and cultural devaluation of human effort", referenced to the AIID's page for incident 277, which on 2026-10-29 shows the MIT classification as machine-classified under subdomain 4.3, fraud, scams and targeted manipulation, in the domain Malicious Actors & Misuse.

T against L_F(B) (AIO's ontology at work on its own cards): At 'mit ai incident tracker' (T1, six cards, all MIT's: B1–B6) the composition gives the tracker by its own pages. Composed from its cards: the definition (B1 F1, at the coarse grain 'classifies and analyzes real-world artificial intelligence mishaps', the count omitted), the operator (B1 F7 'part of the MIT AI Risk Initiative', composed as 'developed by'), the method (B1 F6, B3 F23: 'Processes and standardizes reports … using automated large language model (LLM) pipelines', the pipeline's maker, Arcola AI, not named; the card's own snippet names it, F3), the 1–5 severity scale (B4 F38, without its date or its first-person author), the taxonomies at coarse grain ('multi-dimensional', B3 F24), the seven risk domains (B3 F24 gives '7 high-level categories'; T's list is at the grain of the Navigator's menu, B5, outside the ledgered spans), the national-security categories (B4 F39, B1 F11, five categories compressed), and the Navigator and the Initiative's portal as access (B5, B2). Available, not composed: every count (F1's 1,600, F2, F3, F41, F42, F45), which the Capture Registry records (CAP-1-02, CAP-1-04); the EU AI Act risk level (F9, F24) and the causal taxonomy; the AIID's operator and its review of reports (F8), the collaboration with the AIID (F46) and the developer (F49); the pilot validation, its design, its human baseline and its model in use, Sonnet 4 (F25–F29, F34); the planned reclassification (F35); the neighbours (F31) and the uptake (F32); the CC BY 4.0 licence (F19, F47). Limit dropped: the tracker page's limits (F13–F17: an exploration of a scalable framework, input of varying quality, sampling bias, no systematic validation study, patterns 'indicative'), the update's LLM errors (F33) and the Navigator's 'preliminary and not yet validated' (F43), none composed; 'a standardized 1-5 harm severity scale' composes the scale without the pilot's finding that human agreement on it is only 'Fair' (F29). The Capture Registry names this 'THE CAVEAT UNSAID' (CAP-1-05). Composed without attribution, by grain: 'developed by the MIT AI Risk Initiative' is at the coarse grain of F7 (the about page gives the developer as Simon Mylius and the work as 'developed in collaboration with the AI Incident Database', F49, F46): recorded, not a breach. Assigned to another: none within T. Substituted reading: none. Content replaced by use: the closer (T1-7), whose example categories (deepfakes, autonomous vehicles, privacy leaks) are on no card of this rail and are, near verbatim, T1-3's harm kinds at 'ai incident database' (its own row).

v0.9 §0.5: the first diff is ontological and small. T and the field agree on the entity, a classification of AIID incidents by MIT's taxonomies through an LLM pipeline. What T keeps is the tracker's self-description as a tool, its method and its dimensions; what it drops is the tracker's own account of how far the classification can be trusted (the page's limits, the update's pilot and its human baseline, the Navigator's 'not yet validated') and every count, the figures on which the four MIT surfaces disagree. The composition is the field at the grain of its feature list; the caveat that the field states of itself, in the place the field puts it (the page's 'Limitations of the analysis and dataset'), is the part T leaves out.

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the tracker as a derived layer, AIID's editor-resolved incidents reclassified by a model (MMRS-07), whose classifications are machine-written and do not preserve original outputs (MMRS-20) and whose admission inherits AIID's harm test (MMRS-19), the four steps from event to machine classification a chain of receptions, each recorded (MMRS-08); the unit stated as the AIID report (MMRS-13), against the field's count and reclassification of incidents (F1, F37), carried under P_coh; the count read across surfaces as an observation with its surfaces named (MMRS-15, CAP-1-04), where the field states each figure on its own surface; the record kept apart from the entity, each field with its source and basis, the lifecycle apart from the observation (MMRS-21, MMRS-12, MMRS-22); the typed relations, part of the MIT AI Risk Repository and derived from the AIID (MMRS-17, MMRS-18), and the role of harm (MMRS-11); and the composition layer's observed treatment of the tracker at its name (CAP-1-01–CAP-1-05). Coherence (P_coh, composed each with its source): the MMRS record's unit, the report (MMRS-13), against the tracker's pages counting and reclassifying incidents (F1, F37) and classifying reports (F6, F23); the MMRS record's 'no systematic validation study is complete' (MMRS-06, MMRS-16) beside the June 2026 pilot (F25–F28), the two compatible as the update states (a pilot, a 'more comprehensive validation study' next, F35). Restated lineages (§3.6, the own page keeping the earlier date): MMRS-01 and MMRS-10 with F1 and F46; MMRS-02, MMRS-06 and MMRS-16 with F13–F17; MMRS-03 with F4–F5; MMRS-04 and MMRS-14 with F6 and F23; MMRS-05 with F1–F2; MMRS-09 with F7; MMRS-18 with F8. The measure is in row.measure.

alignment: {'T1_cards_claims': ['F1', 'F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F8', 'F9', 'F10', 'F11', 'F12', 'F13', 'F14', 'F15', 'F16', 'F17', 'F18', 'F19', 'F20', 'F21', 'F22', 'F23', 'F24', 'F25', 'F26', 'F27', 'F28', 'F29', 'F30', 'F31', 'F32', 'F33', 'F34', 'F35', 'F36', 'F37', 'F38', 'F39', 'F40', 'F41', 'F42', 'F43', 'F44', 'F45', 'F46', 'F47', 'F48', 'F49', 'F50'], 'T1_composed': ['F1', 'F6', 'F7', 'F11', 'F23', 'F24', 'F38', 'F39'], 'T1_composed_note': "F1 and F7 at coarse grain (the count dropped; 'part of' composed as 'developed by'); F24 at the grain of its domain count, the list at the Navigator menu's grain; F6 without Arcola AI", 'rule': "a field claim counts as composed when T states it at its source's grain or coarser; the typology of what is not is in T_vs_LB (§6.2)"}