airisk.mit.edu

Open the Incidents section of airisk.mit.edu to browse and compare documented real-world AI harms. The site lets you filter incidents by risk domain, type and severity of harm, number of people affected, alleged developer, time, and national security impact. Use it when you need concrete examples of AI harms rather than abstract risk categories — for instance, to ground a policy discussion, benchmark a company's disclosed risks, or find precedents for a specific harm type.

What the incident data contains

The MIT AI Risk Initiative describes its incidents dataset as a way to "explore and compare real-world AI incidents." Each record is organized around dimensions that let you move from a single case to broader patterns:

• Risk domain — the category of risk the incident falls under
• Type and severity of harm — what kind of damage occurred and how serious it was
• Number of people affected — the scale of impact
• Alleged developer — the organization linked to the system involved
• Time — when the incident occurred
• National security impact — whether the incident carries national-security relevance

These fields are the comparison axes. You can hold one constant (say, a risk domain) and compare across another (severity, or time) to see how harms cluster.

Step-by-step: from a single incident to a pattern

1. Start in the Incidents view. This is the entry point for real-world harms, distinct from the risk repository and governance sections.
2. Pick a comparison axis. If you're investigating a specific harm type, filter by risk domain first. If you're tracking a company, filter by alleged developer.
3. Narrow by severity or scale. Add the harm severity or number-of-people-affected filters to separate minor incidents from high-impact ones.
4. Check the time dimension. Sorting or filtering by time reveals whether incidents in your domain are increasing, concentrated in a period, or isolated.
5. Flag national-security relevance if your question involves security policy, since this is a separate field from general severity.
6. Record what you find — the site presents incidents for exploration and comparison; the filtering is how you turn a list into evidence.

Expected result: a filtered set of incidents you can cite as concrete examples, plus a sense of how your chosen risk domain behaves across severity, scale, and time.

Connecting incidents to the rest of the site

The incidents data is one of several linked datasets, and the AI Risk Navigator is the tool that connects them: the site says the Navigator "connects our datasets so you can move between them and surface the patterns that matter." That matters for incidents because a single harm usually has three companion views:

A practical workflow: find an incident in your domain, note its risk category, then check the expert survey to see whether that category is ranked as high-severity, and the governance database to see which laws address it. The company disclosures view adds a fourth check — whether major companies publicly acknowledge that risk at all.

What the incident data is useful for

• Grounding abstract risk discussions. Instead of arguing about hypothetical harms, point to documented incidents with severity and scale attached.
• Testing disclosure gaps. Compare incidents in a risk domain against company disclosures to see whether firms acknowledge the risks that are actually materializing.
• Prioritizing governance work. Cross-reference incident frequency and severity with the expert consensus survey to decide which risks deserve attention first.
• Policy and legal mapping. Use the risk domain of an incident to filter the governance database for relevant laws and frameworks.

The site's stated purpose is to "help you identify, prioritize, and manage the risks from AI," and the incidents dataset serves the identify step most directly — it is the empirical record against which the other datasets can be checked.

Common sticking points

• Don't treat the incident list as exhaustive. The site calls its risk repository "living," implying ongoing updates; incident coverage should be read the same way.
• Severity and national-security impact are separate fields. An incident can be severe without national-security relevance, so filter both if your question spans security policy.
• Alleged developer is not a finding of liability. The field is labeled "alleged," so attribute carefully when citing.
• Use the Navigator for cross-dataset questions. If you find yourself manually jumping between incidents, risks, and governance pages, the Navigator is designed to move between them in one place.

The MIT AI Risk Initiative catalogs AI risks across a living repository of risk types, a ranked set of priority risks based on expert consensus, a database of real-world incidents, a mapping of global AI laws and frameworks to the risks they address, and a review of how large companies disclose these risks. The site is built for people who need to identify, prioritize, and manage AI risk — policy analysts, governance teams, and researchers — and its tools are organized around four questions: which risks matter most, what the risks are, what harms are actually occurring, and who is responding.

The four datasets and what each covers

How the pieces connect

The AI Risk Navigator is the entry point that links these datasets, so you can move between a risk category, the incidents associated with it, the laws that address it, and the companies that disclose it. That linkage is the site's main structural idea: a risk is not just a label but a node connecting evidence of harm, responsible actors, and regulatory coverage.

A practical example: if you are assessing exposure for a deployed hiring tool, you could start from the risk repository to identify relevant risk domains, check the incidents database for documented harms in that domain filtered by severity and affected population, then use the governance tracker to see which laws and frameworks cover that domain at your lifecycle stage. The company disclosures view tells you whether large firms in the same space acknowledge the risk at all — a useful gap check between expert priorities and public disclosure.

Priority risks and expert consensus

Rather than presenting all risks as equally weighted, the initiative publishes expert consensus on which risks are expected to have the most severe harms, who is most vulnerable, and who should be responsible for addressing them. This matters if you need to justify prioritization decisions: the survey gives you a citable basis for ranking risks rather than relying on internal judgment alone. The disclosure comparison against 24 risks then shows where corporate acknowledgment diverges from that expert ranking.

Who the site is for and how to use it

The work is cited by academic and governance figures, including Yoshua Bengio (Université de Montréal, LawZero, Mila), who describes it as filling a critical information gap, and practitioners at Partnership on AI, Oracle, and Credo AI who point to its taxonomy and evidence-based approach for policy discussion and value-chain risk management. If your task is building a risk register, briefing policymakers, or benchmarking disclosures, start with the priority survey for ranking, the repository for definitions, and the Navigator to trace risks to incidents and laws. The site also maintains a blog and newsletter for updates to the living datasets.

The MIT AI Risk Initiative (airisk.mit.edu) is a research initiative that produces authoritative data and frameworks to help people identify, prioritize, and manage risks from AI. It is most useful if you need structured, evidence-based material on AI risk — for example, to ground a policy discussion, benchmark company disclosures, or map which laws cover which risks — rather than general commentary. Its core offerings are a living repository of AI risks, an incident database, a governance/law mapping, a company disclosure tracker, and the AI Risk Navigator that connects these datasets.

What the site provides

The homepage organizes its work around five questions, each backed by a tool or dataset:

The AI Risk Navigator

The Navigator is the site's integration layer: it connects the datasets above so you can move between them and surface patterns that matter across risk, incidents, governance, and company disclosures. If you only have time for one entry point and your question spans more than one dataset, start here.

Who it is for

The site is built for people who need to act on AI risk information rather than just read about it. Named users in the site's testimonials point to three audiences:

• Policy and governance professionals — Madhu Srikumar of the Partnership on AI describes the taxonomy as helping "the policy community have more grounded discussions" in a field "characterized by significant uncertainty."
• Legal and compliance teams — Kevin Fumai of Oracle cites the "practical, evidence-based approach" as offering "immediate benefits to anyone integrating risk management across the AI value chain."
• Researchers and technical teams — Yoshua Bengio calls it "a valuable database for accessing current and comprehensive information on a wide range of AI-related risks," and Dr. Ian W. Eisenberg of Credo AI frames it as "open academic infrastructure the AI governance ecosystem needs."

How to use it in practice

A practical sequence for a typical task — say, preparing an internal briefing on a specific risk domain:

1. Scope the risk. Use the risk repository to find the relevant risk domain and how it emerges.
2. Check for real-world evidence. Query the incident database filtered to that domain, and note severity, number of people affected, and alleged developer.
3. Check the legal picture. Use the governance mapping, filtered by sector or lifecycle stage, to see which laws and frameworks address that risk.
4. Compare with corporate practice. Check the company disclosure tracker to see whether large companies acknowledge that risk publicly, and how that compares with expert consensus.
5. Cross-reference. Use the Navigator to move between these views and identify gaps — for example, a risk experts rank as severe but that few companies disclose or few laws cover.

The expected result at each step is a citable, structured view rather than a narrative summary, which is what makes the material usable in governance and policy settings.

What to keep in mind

• The site presents itself as a source of data and frameworks, not advice; the judgments about what to do with the findings remain yours.
• The risk repository is described as "living," and the incident and disclosure data are tracked over time, so figures you cite should be dated to when you accessed them.
• The site does not state pricing or login requirements in the material reviewed here, so check the individual tool pages before assuming open access to every dataset.

For a first visit, the Navigator gives the broadest view; for a narrow question, go directly to the matching dataset in the table above.

The MIT AI Risk Initiative tracks company AI risk disclosures by reviewing the public documents of 200 of the world's largest companies and recording which of 24 AI risks each one acknowledges. The results are published as a browsable dataset that you can compare against the Initiative's expert survey of which risks matter most. Use it when you want evidence of where corporate disclosure and expert priorities diverge, rather than a compliance score or a legal judgment about any company.

What the disclosure dataset covers

The site describes the company disclosures tool as showing "which of 24 AI risks 200 of the world's largest companies acknowledge in their public documents, and how that compares with what experts say matters most."

That definition sets the boundaries of the dataset:

• Unit of analysis: a company, drawn from the world's largest firms by the site's own selection.
• Evidence base: public documents, not interviews, audits, or internal records.
• What is recorded: acknowledgment of a risk — whether the company names or addresses it — not the quality or effectiveness of any mitigation.
• Fixed risk list: 24 risks, which lets you compare companies and risks on the same footing.

A company appearing in the dataset therefore means its public materials mention a given risk. It does not mean the company manages that risk well, and absence from the dataset does not prove a company ignores the risk — only that the reviewed documents did not acknowledge it.

How the comparison with expert consensus works

The disclosure data is designed to be read next to the Initiative's expert survey, which reports "expert consensus on which AI risks are expected to have the most severe harms, who is most vulnerable, and who is responsible for addressing them."

The comparison answers a practical question: are companies disclosing the risks that experts rank as most severe? The site frames the disclosure tool around exactly this gap — how corporate acknowledgment "compares with what experts say matters most." Two patterns are worth looking for:

1. High-severity risks with low disclosure. Experts rank a risk as severe, but few of the 200 companies acknowledge it. This is the clearest signal of a disclosure gap.
2. Widely disclosed risks that experts rank lower. Many companies mention a risk, but it sits lower in the expert ranking. This can indicate disclosure shaped by reputational or legal fashion rather than severity.

Because both datasets use the same risk taxonomy, the comparison is like-for-like. That shared structure is the main reason the two tools are useful together rather than separately.

How to explore company disclosures and mitigations

The site does not publish a step-by-step manual for the disclosure tool, so the following is a practical sequence based on what the tool is described to contain. Expect the exact filters and labels to match the live interface.

1. Open the company disclosures tool from the "Our tools and data" section of the homepage.
2. Pick a risk from the 24-risk list to see which companies acknowledge it. Starting from a risk rather than a company is usually faster when you are testing the disclosure gap.
3. Pick a company to see the set of risks it acknowledges across its public documents.
4. Read across to the expert survey for the same risk, and note whether the severity ranking and the disclosure count point in the same direction.
5. Check the mitigation and response material. The site's navigation includes "Responses," "Mitigations," and "Governance" alongside "Company Disclosures," so related information on how risks are addressed sits in adjacent collections rather than only in the disclosure view.
6. Use the AI Risk Navigator to move between datasets. The site describes it as connecting "our datasets so you can move between them and surface the patterns that matter" — useful when you want to follow one risk from expert ranking to company disclosure to incidents.

What to verify before relying on a result: confirm the document set and date range behind a company's entry, since disclosure is a snapshot of public materials at review time. A company that published a new AI policy after the review will not show it.

Where this fits alongside the other tools

The disclosure dataset is one of several connected collections. The homepage lists tools for expert priority rankings, a living repository of AI risks, real-world incidents, and global AI laws and policies mapped against the risks they address. The disclosure view is the corporate-behavior layer: laws show what is required, incidents show what has gone wrong, and disclosures show what companies say about risk.

For a governance or policy audience, the most defensible use is triangulation — a risk that experts rank as severe, that appears in incident data, and that few companies disclose is a stronger case for attention than any single dataset alone. The Initiative's stated purpose, "authoritative data and frameworks to help you identify, prioritize, and manage the risks from AI," matches that layered reading rather than any single score.

Limits to keep in mind

• Disclosure is not performance. The dataset records acknowledgment in public documents; it does not measure whether mitigations work.
• The company set is fixed at 200 of the largest firms, so smaller developers and open-source actors are outside the frame.
• The 24-risk list is the Initiative's taxonomy. Comparisons are valid within the site's framework; mapping to another taxonomy requires care.
• Public documents change. Treat any figure as a snapshot and check the review date before citing it in a decision.

The MIT AI Risk Initiative maps global AI laws, policies, and frameworks against the specific risks they address, then lets you filter that mapping by risk domain, sector, lifecycle stage, value chain actor, legislative status, and technical scope. Use it when you need to answer a question like "which rules actually cover this risk, and where?" rather than "what does one law say?" The mapping lives in the Governance section of the site and is connected to the initiative's other datasets through the AI Risk Navigator.

What the mapping actually connects

The site describes its governance data as "Global AI laws, policies, and frameworks, mapped against the risks they address." That means each entry is treated as a law or framework on one side and a set of risks on the other, with the link between them being the thing you browse. This is a coverage view, not a legal analysis: it tells you which instruments touch which risks, not whether a given instrument is strong, enforceable, or sufficient.

The risks themselves come from the initiative's living repository of AI risks, so the governance mapping inherits that taxonomy. If you want to understand a risk category before checking coverage, read the risk entry first — the law mapping will make more sense against that definition.

The six filters and what each one is for

The practical move is to combine filters rather than use one at a time. Filtering by risk domain alone gives you a long list; adding legislative status separates binding rules from aspirational frameworks, and adding value chain actor tells you whether the obligations land on the party you care about.

A worked example

Suppose you need to know what exists for a risk in a specific sector, aimed at deployers, and actually in force. Start with the risk domain filter to pull the relevant risk category, add the sector to drop instruments aimed elsewhere, set value chain actor to deployers, and set legislative status to enacted instruments only. What remains is a short list you can read individually. Reverse the order — start from a law you already know and read its mapped risks — when you are auditing an existing instrument rather than searching for coverage.

How this fits the rest of the site

The governance mapping is one of several connected datasets. The site also publishes expert consensus on which risks are expected to cause the most severe harms and who is responsible for addressing them, a living repository of AI risks, a database of real-world incidents filterable by risk domain, harm type and severity, people affected, alleged developer, time, and national security impact, and a company disclosure dataset covering which of 24 AI risks 200 of the world's largest companies acknowledge in public documents.

The AI Risk Navigator connects these so you can move between them. A useful sequence: check expert consensus for priority, check company disclosures for whether firms acknowledge that priority, then use the governance mapping to see what rules cover it. Gaps between the three are often the most informative output.

Limits to keep in mind

• Coverage is not compliance. A law appearing against a risk does not mean it is enforced or effective.
• The risk taxonomy is the initiative's own, so results are only as granular as that taxonomy.
• Legislative status changes; treat status filters as a snapshot and verify against the primary source before relying on it.
• The site presents itself as producing data and frameworks for identifying, prioritizing, and managing AI risk. It is a research resource, not legal advice — for a decision with legal consequences, confirm against the actual text of each instrument.