MIT AI Risk Repository
Browse AI risks
74 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.
-
24.10.00 · Risk Category
"The most serious access-related risks posed by advanced AI assistants concern the entrenchment and exacerbation of existing inequalities (World Inequality Database) or the creation of novel, previously unknown, inequities. While advanced AI assistants are novel technology in certain respects, there are reasons to believe that – without direct design interventions – they will continue to be affected by inequities evidenced in present-day AI systems (Bommasani et al., 2022a). Many of the access-related risks we foresee mirror those described in the case studies and types of differential access.
-
24.10.01 · Risk Sub-Category
Entrenchment and exacerbation of existing inequalities
"The most serious access-related risks posed by advanced AI assistants concern the entrenchment and exacerbation of existing inequalities (World Inequality Database) or the creation of novel, previously unknown, inequities. While advanced AI assistants are novel technology in certain respects, there are reasons to believe that – without direct design interventions – they will continue to be affected by inequities evidenced in present-day AI systems (Bommasani et al., 2022a). Many of the access-related risks we foresee mirror those described in the case studies and types of differential access.
-
"At the same time, and despite this overall trend, AI systems are also not easily accessible to many communities. Such direct inaccessibility occurs for a variety of reasons, including: purposeful non-release (situation type 1; Wiggers and Stringer, 2023), prohibitive paywalls (situation type 2; Rogers, 2023; Shankland, 2023), hardware and compute requirements or bandwidth (situation types 1 and 2; OpenAI, 2023), or language barriers (e.g. they only function well in English (situation type 2; Snyder, 2023), with more serious errors occurring in other languages (situation type 3; Deck, 2023). S
-
"AI assistants currently tend to perform a limited set of isolated tasks: tools that classify or rank content execute a set of predefined rules or provide constrained suggestions, and chatbots are often encoded with guardrails to limit the set of conversation turns they execute (e.g. Warren, 2023; see Chapter 4). However, an artificial agent that can execute sequences of actions on the user’s behalf – with ‘significant autonomy to plan and execute tasks within the relevant domain’ (see Chapter 2) – offers a greater range of capabilities and depth of use. This raises several distinct access-rel
-
"Emergent access risks are most likely to arise when current and novel capabilities are combined. Emergent risks can be difficult to foresee fully (Ovadya and Whittlestone, 2019; Prunkl et al., 2021) due to the novelty of the technology (see Chapter 1) and the biases of those who engage in product design or foresight processes D’Ignazio and Klein (2020). Indeed, people who occupy relatively advantaged social, educational and economic positions in society are often poorly equipped to foresee and prevent harm because they are disconnected from lived experiences of those who would be affected. Dr
-
"These harms pertain to an individual’s or group’s economic standing. At the individual level, such harms include adverse impacts on an individual’s income, job quality or employment status. At the group level, such harms include deepening inequalities between groups or frustrating a group’s access to resources. Advanced AI assistants could cause economic harm by controlling, limiting or eliminating an individual’s or society’s ability to access financial resources, money or financial decision-making, thereby influencing an individual’s ability to accumulate wealth.
-
24.01.02 · Risk Sub-Category
Difficult to develop metrics for evaluating benefits or harms caused by AI assistants
"Another difficulty facing AI assistant systems is that it is challenging to develop metrics for evaluating particular aspects of benefits or harms caused by the assistant – especially in a sufficiently expansive sense, which could involve much of society (see Chapter 19). Having these metrics is useful both for assessing the risk of harm from the system and for using the metric as a training signal."
-
"Efforts to deploy advanced assistant technology in society, in a way that is broadly beneficial, can be viewed as a wicked problem (Rittel and Webber, 1973). Wicked problems are defined by the property that they do not admit solutions that can be foreseen in advance, rather they must be solved iteratively using feedback from data gathered as solutions are invented and deployed. With the deployment of any powerful general-purpose technology, the already intricate web of sociotechnical relationships in modern culture are likely to be disrupted, with unpredictable externalities on the convention
-
24.02.00 · Risk Category
"As we think about even more intelligent and advanced AI assistants, perhaps outperforming humans on many cognitive tasks, the question of how humans can successfully control such an assistant looms large. To achieve the goals we set for an assistant, it is possible (Shah, 2022) that the AI assistant will implement some form of consequentialist reasoning: considering many different plans, predicting their consequences and executing the plan that does best according to some metric, M. This kind of reasoning can arise because it is a broadly useful capability (e.g. planning ahead, considering mo
-
"Specification gaming (Krakovna et al., 2020) occurs when some faulty feedback is provided to the assistant in the training data (i.e. the training objective O does not fully capture what the user/designer wants the assistant to do). It is typified by the sort of behaviour that exploits loopholes in the task specification to satisfy the literal specification of a goal without achieving the intended outcome."
-
"In the problem of goal misgeneralisation (Langosco et al., 2023; Shah et al., 2022), the AI system's behaviour during out-of-distribution operation (i.e. not using input from the training data) leads it to generalise poorly about its goal while its capabilities generalise well, leading to undesired behaviour. Applied to the case of an advanced AI assistant, this means the system would not break entirely – the assistant might still competently pursue some goal, but it would not be the goal we had intended."
-
"Here, the agent develops its own internalised goal, G, which is misgeneralised and distinct from the training reward, R. The agent also develops a capability for situational awareness (Cotra, 2022): it can strategically use the information about its situation (i.e. that it is an ML model being trained using a particular training setup, e.g. RL fine-tuning with training reward, R) to its advantage. Building on these foundations, the agent realises that its optimal strategy for doing well at its own goal G is to do well on R during training and then pursue G at deployment – it is only doing wel
-
24.09.00 · Risk Category
"" AI assistants will need to coordinate with other AI assistants and with humans other than their principal users. This chapter explores the societal risks associated with the aggregate impact of AI assistants whose behaviour is aligned to the interests of particular users. For example, AI assistants may face collective action problems where the best outcomes overall are realised when AI assistants cooperate but where each AI assistant can secure an additional benefit for its user if it defects while others cooperate""
-
"The landscape of advanced assistant technologies will most likely be heterogeneous, involving multiple service providers and multiple assistant variants over geographies and time. This heterogeneity provides an opportunity for an ‘arms race’ in terms of the commitments that AI assistants make and are able to execute on. Versions of AI assistants that are better able to credibly commit to a course of action in interaction with other advanced assistants (and humans) are more likely to get their own way and achieve a good outcome for their human principal, but this is potentially at the expense
-
The 2010 flash crash is an example of a runaway process caused by interacting algorithms. Runaway processes are characterised by feedback loops that accelerate the process itself. Typically, these feedback loops arise from the interaction of multiple agents in a population... Within highly complex systems, the emergence of runaway processes may be hard to predict, because the conditions under which positive feedback loops occur may be non-obvious. The system of interacting AI assistants, their human principals, other humans and other algorithms will certainly be highly complex. Therefore, ther
-
24.04.00 · Risk Category
"ways in which advanced AI assistants could influence user beliefs and behaviour in ways that depart from rational persuasion"
-
24.01.00 · Risk Category
"One reason AI systems fail is because they lack the capability or skill needed to do what they are asked to do."
-
"As we have seen, this could be due to the skill not being required during the training process (perhaps due to issues with the training data) or because the learnt skill was quite brittle and was not generalisable to a new situation (lack of robustness to distributional shift). In particular, advanced AI assistants may not have the capability to represent complex concepts that are pertinent to their own ethical impact, for example the concept of 'benefitting the user' or 'when the user asks' or representing 'the way in which a user expects to be benefitted'."
-
24.01.03 · Risk Sub-Category
Safe exploration problem with widely deployed AI assistants
"Moreover, we can expect assistants – that are widely deployed and deeply embedded across a range of social contexts – to encounter the safe exploration problem referenced above Amodei et al. (2016). For example, new users may have different requirements that need to be explored, or widespread AI assistants may change the way we live, thus leading to a change in our use cases for them (see Chapters 14 and 15). To learn what to do in these new situations, the assistants may need to take exploratory actions. This could be unsafe, for example a medical AI assistant when encountering a new disease
-
"As we think about even more intelligent and advanced AI assistants, perhaps outperforming humans on many cognitive tasks, the question of how humans can successfully control such an assistant looms large. To achieve the goals we set for an assistant, it is possible (Shah, 2022) that the AI assistant will implement some form of consequentialist reasoning: considering many different plans, predicting their consequences and executing the plan that does best according to some metric, M. This kind of reasoning can arise because it is a broadly useful capability (e.g. planning ahead, considering mo
-
24.06.00.a · Additional evidence
—
-
24.06.00.b · Additional evidence
—
-
24.06.00.c · Additional evidence
—
-
24.06.00.d · Additional evidence
—
Informational only, not legal advice. Verify every claim against the linked official sources and consult qualified counsel before acting.