MIT AI Risk Repository
Browse AI risks
242 risk entries extracted from 74 frameworks, coded by domain, subdomain, causal entity, intent and timing. Filter, then export the current selection with its licence and citation attached.
-
-
-
—
-
—
-
51.10.00 · Risk Category
"The rational agent framework is pervasive in the study of artificial intelligence. It typically assumes that a well-delineated entity interacts with an environment through action and observation channels. This is not a realistic assumption for physicalistic agents such as robots that are part of the world they interact with (Soares and Fallenstein, 2014, 2017)."
-
51.11.00 · Risk Category
"An artificial intelligence may be copied and distributed, allowing instances of it to interact with the world in parallel. This can significantly boost learning, but undermines the concept of a single agent interacting with the world."
-
53.04.05 · Risk Sub-Category
Indirect AI contributions to existential risks
Other diffuse societal harms
-
-
54.01.00 · Risk Category
"A major role of the current AI ethics movement is to draw attention to overlooked side-effects, costs, and harms of building and deploying AI systems, particularly as they befall existing marginalized groups:"
-
55.01.00 · Risk Category
"Scientific progress: AI could lead to very rapid scientific progress which would likely have long-term impacts, but it’s very unclear if these would be positive or negative. Much depends on the extent to which risky scientific domains are sped up relative to beneficial or risk-reducing ones, on who uses the technology enabled by this progress, and on how it is governed."
-
"Physical - Physical injury to an individual or group, or damage to physical property."
-
"Harassment/abuse/intimidation - Online behaviour, including sexual harassment, that makes an individual or group feel alarmed or threatened."
-
"Self-harm - Intentional seeking out or sharing of hurtful content about oneself that leads to, supports, or exacerbates low self-esteem and self-harm."
-
"Sexualisation - Sexual interest in a technology or application."
-
"Livelihood loss - An individual or group’s loss of ability to support themselves financially or vocationally due to natural disasters, lack of demand for products/services, cost increases, etc, resulting in inability to procure food, reduced employment prospects, bankruptcy, foreclosure, homelessness, etc."
-
"Chilling effect - The creation of a climate of self-censorship that deters democratic actors such as journalists, advocates and judges from speaking out."
-
"Damage to public health - Adverse impacts on the health of groups, communities or societies, including malnutrition, disease and infection conditions."
-
"Public service delivery deterioration - Poor performance of a public technology system due to malfunc- tion, over-use, under-staffing etc, resulting in individuals, groups, or organisations unable to use it in a manner they can reasonably expect."
-
59.06.00 · Risk Category
"The development and operation of an AI system can require significant amounts of (computational) power. If not considered in the hardware selection, this can become an issue in development and operation."
-
59.25.00 · Risk Category
"The first axis pertains to the life cycle of the AI system, as AI hazards may materialize during various phases of an AI system’s life cycle. For instance, issues triggered by bias in training data emerge during the data collection and preparation stages. On the other hand, data drift serves as an example of an AI hazard that arises during the AI system’s operation. Additionally, certain AI hazards may span multiple phases of the AI system, such as ”lack of data understanding”. This is because a proper understanding of the data by the AI developer is required in the data collection and prepar
-
"A majority of them possess an initial stage devoted to the planning and scoping of the AI system."
-
—
-
—
-
—
-
"Conclusively, the AI life cycle model terminates with the maintenance and monitoring stage, which aligns with the referenced models."
-
"The second axis of the taxonomy pertains to the mode of an AI hazard, which determines with what methods to assess and treat AI hazards. We distinguish among three distinct classes: technological, socio-technological, and procedural."
-
60.02.00 · Risk Category
-
-
60.03.00 · Risk Category
"This section considers a range of systemic risks, in the sense of “broader societal risks associated with AI deployment, beyond the capabilities of individual models” (636). Note that this is not identical with how the European AI Act uses ‘systemic risks’ to refer to general - purpose AI models with a high impact on society, based on criteria such as training compute and the number of users."
-
61.01.00 · Risk Category
-
-
61.02.00 · Risk Category
-
-
62.02.00 · Risk Category
—
-
"A risk may be triggered by a human, where the AI serves merely as a tool, or by the AI acting autonomously with no human intervention, or it may involve a combination of both, with the human delegating some parts of decision-making to the AI. For risks where AI is the entity, these risks are exacerbated by an increase in the AI’s level of autonomy. To manage risks involving AI as the trigger, appropriate levels of human oversight can be built-in."
-
"A risk may be triggered by a human, where the AI serves merely as a tool, or by the AI acting autonomously with no human intervention, or it may involve a combination of both, with the human delegating some parts of decision-making to the AI. For risks where AI is the entity, these risks are exacerbated by an increase in the AI’s level of autonomy. To manage risks involving AI as the trigger, appropriate levels of human oversight can be built-in."
-
"A risk may be triggered by a human, where the AI serves merely as a tool, or by the AI acting autonomously with no human intervention, or it may involve a combination of both, with the human delegating some parts of decision-making to the AI. For risks where AI is the entity, these risks are exacerbated by an increase in the AI’s level of autonomy. To manage risks involving AI as the trigger, appropriate levels of human oversight can be built-in."
-
62.06.00 · Risk Category
—
-
"For GPAIs or foundation models, risks emerge during training, prior to being repurposed and deployed in more specific AI systems or applications. Risk assessments can be conducted before deployment, and monitoring of AI models can occur as required throughout the deployment phase. In certain cases, version updates or model recalls may be warranted post-deployment."
-
"For GPAIs or foundation models, risks emerge during training, prior to being repurposed and deployed in more specific AI systems or applications. Risk assessments can be conducted before deployment, and monitoring of AI models can occur as required throughout the deployment phase. In certain cases, version updates or model recalls may be warranted post-deployment."
-
62.07.00 · Risk Category
"For “system and operational harms,” the AI systems interact with other systems and industries, where a failure in an AI system could lead to failures of a wider scope."
-
62.09.00 · Risk Category
"These are in contrast with “societal harms,” which are less direct but have more far-reaching effects on segments of society"
-
—
-
62.10.00 · Risk Category
"Finally, “legal and rights-related harms” concern either harms from illegal activities or harms from violations of human rights."
-
62.10.03 · Risk Sub-Category
Direct Harm Domains (legal and rights-related harms)
Criminal activities
—
-
62.18.02 · Risk Sub-Category
Model Evaluations (Interpretability/Explainability)
Misunderstanding or overestimating the results and scope of interpretability techniques
"The results of explainability techniques are not free of bias and require careful interpretation. Users might develop a false sense of security or reliability if the resulting explanations align with their initial beliefs, leading to confirmation bias and an overestimation of abilities of these techniques [24]."
-
62.19.00 · Risk Category
"This section catalogs the risk sources related to GPAI failure modes or attacks targeting GPAIs. Many of these apply mainly to LLM-based GPAIs, which share some common failure modes such as jailbreaks and trojans. These vulnerabilities often extend beyond GPAIs and fall into the broader field of adversarial machine learning. However, additional vulnerabilities may arise with the introduction of new modalities, longer context windows, or different encodings."
-
62.29.00 · Risk Category
-
-
62.30.00 · Risk Category
-
-
62.31.00#1 · Risk Category
-
-
62.31.00#2 · Risk Category
-
-
—
-
—
-
—
-
"Existing laws could include providing data subject rights such as opt-out, right to access, and right to be forgotten."
Informational only, not legal advice. Verify every claim against the linked official sources and consult qualified counsel before acting.