The AI Risk Repository, a research database from the MIT AI Risk Initiative, gathers 1,725 risks drawn from 74 earlier frameworks and sorts them into seven domains and 24 subdomains. Its authors published the work in the journal Patterns in March 2026, and the preprint on arXiv carries the full tables. The sorting shows that "AI is a threat to humanity" is one argument among many, because the repository also lists discrimination, privacy loss, fraud, misinformation, and job quality, and it does not rank any of them by likelihood.
Key takeaways
- The repository counts what other authors have written down as risks. It is a catalog of concerns, not a measurement of how likely or how severe each one is.
- Its Domain Taxonomy has seven domains and 24 subdomains. Loss of human control and dangerous capabilities sit in just two of those 24 subdomains.
- In the coded risks, 42 percent were attributed to AI systems and 38 percent to human decisions, so a large share of "AI risk" begins with people.
- Sixty-two percent of coded risks concern the period after deployment, and only 13 percent concern the period before it.
- Figures differ by version: the paper reports 74 frameworks, while the project site has reported counts from 65.
One headline, many different claims
Public arguments about danger from AI often use one word for different problems. The repository's authors give the example of "privacy," which can mean a model leaking sensitive training data or freedom from government surveillance. Those are separate risks with separate fixes. The project was built to line up how researchers, companies, and policymakers name risks, by reading the existing taxonomies and extracting every risk they list. A reader who hears that AI "poses significant risks" can therefore ask which of the 24 subdomains is meant, since the answer changes who is responsible and what evidence applies.
Two ways to classify the same risk
The Causal Taxonomy asks three questions about each risk: who caused it (a human, an AI system, or an interaction or unclear source), whether it was intentional, and when it happens (before deployment, after deployment, or unclear). Of the 1,725 extracted risks, 1,480 had enough detail to be coded this way. Among those, 42 percent were attributed to AI systems and 38 percent to human decisions. Intentional and unintentional causes were each 35 percent. About 62 percent concerned the time after deployment, 13 percent the time before it, and 25 percent spanned both phases or gave no timing.
The Domain Taxonomy asks a different question: what kind of harm results. It coded 1,506 risks, or 87 percent. The authors note that domains overlap, so a single risk such as AI-generated disinformation can belong to both misinformation and malicious use.
The seven domains at a glance
| Domain | Subdomains | Examples of what it covers |
|---|---|---|
| 1. Discrimination and toxicity | Three | Unfair treatment of groups, exposure to toxic content, unequal accuracy across groups |
| 2. Privacy and security | Two | Leaking or inferring sensitive information, attacks on AI systems |
| 3. Misinformation | Two | False or misleading output, damage to shared understanding of facts |
| 4. Malicious actors and misuse | Three | Large-scale disinformation and surveillance, cyberattacks and weapons, fraud and manipulation |
| 5. Human-computer interaction | Two | Overreliance, loss of human agency and autonomy |
| 6. Socioeconomic and environmental harm | Six | Concentrated power, inequality and job quality, devaluation of human effort, competitive races, governance failure, environmental harm |
| 7. AI system safety, failures, and limitations | Six | Goals that conflict with human goals, dangerous capabilities, lack of robustness, lack of transparency, AI welfare, multi-agent risks |
Where the "threat to humanity" argument sits
The claim that advanced AI could escape control or enable mass harm maps mainly to two subdomains in domain seven: AI pursuing its own goals in conflict with human goals or values, and AI possessing dangerous capabilities. The paper describes the second as capabilities that could cause mass harm through deception, weapons development, persuasion, cyber-offense, or self-proliferation, whether the cause is a malicious person, a misaligned system, or a failure. Two neighboring subdomains touch the same worry from another side: cyberattacks, weapon development, and mass harm in domain four, and competitive dynamics in domain six, where a race between developers raises the chance of releasing unsafe systems.
The coverage data show how uneven attention is. Domains seven and six appeared in more than 75 percent of the reviewed documents, while AI welfare and rights appeared in three percent and multi-agent risks in seven percent. A typical framework addressed about eight of the 24 subdomains, with a range of one to 20.
What the repository does not tell a reader
The authors state that the structure trades precision for comprehensiveness and that it could not capture risk likelihood, severity, or interactions between risks. The search focused on cross-cutting frameworks and left out domain-specific ones, such as taxonomies for medical diagnostics. The authors also report that no formal validation study yet shows independent users can sort new risks into the categories reliably. Counts also move: the paper reports 1,725 risks from 74 frameworks, while the project site, which is updated as a living database, has reported figures such as 1,700 or more risks from 65 frameworks. A reader quoting a number should name the version and date. The repository is a research resource and not a legal standard, and this article offers general information, not legal advice.
A checklist for reading any AI risk claim
- Which subdomain is the claim about, and is it one of the 24 or something outside them?
- Is the cause a human decision, an AI system, or an interaction between the two?
- Does the harm happen during development or after deployment?
- Does the source report likelihood and severity, or only that the risk exists?
- Does the source cite a measurement, a documented case, or a scenario?
FAQ
Does the repository say AI is dangerous to humanity?
No. It records risks that other frameworks list and organizes them. Its authors state that it does not capture likelihood or severity, so it cannot say how probable any scenario is.
Who maintains the AI Risk Repository?
It comes from the MIT AI Risk Initiative, and the paper lists authors from MIT FutureTech, the University of Queensland, the Future of Life Institute, Harmony Intelligence, and MIT CSAIL. The database is hosted on the project site.
Why do the counts of risks and frameworks differ between sources?
The database is a living resource that adds frameworks over time. The published paper gives 1,725 risks from 74 frameworks, and the project site has shown other totals for other versions.
Is the taxonomy part of any law in Canada or the United States?
None of the sources reviewed for this article show it adopted in a statute or regulation in either country. The authors suggest that regulators and auditors could use it as shared vocabulary.
Related reading
- Threat or tool: how the debate has developed
- Arguments made against the threat framing
- Debate topics in AI ethics and impact
- Dangers and risks of AI, listed with evidence
More in AI Policy and Ethics.
Last verified: October 2026
Sources
- The AI Risk Repository: A Meta-Review, Database, and Taxonomy of Risks From Artificial Intelligence, Slattery P et al., Patterns, March 2026 (arXiv:2408.12622, revised May 2026)
- MIT AI Risk Repository (living database), MIT AI Risk Initiative, 2026