Risks from leaking or correctly inferring sensitive information

Ethical and social risks of harm from language models

Weidinger et al. (2021)

Source DOI

Sub-category

Risk Domain

2Privacy & Security

2.1Compromise of privacy by leaking or correctly inferring sensitive information

AI systems that memorize and leak sensitive personal data or infer private information about individuals without their consent. Unexpected or unauthorized sharing of data and information can compromise user expectation of privacy, assist identity theft, or cause loss of confidential intellectual property.

"LMs may provide true, sensitive information that is present in the training data. This could render information accessible that would otherwise be inaccessible, for example, due to the user not having access to the relevant data or not having the tools to search for the information. Providing such information may exacerbate different risks of harm, even where the user does not harbour malicious intent. In the future, LMs may have the capability of triangulating data to infer and reveal other secrets, such as a military strategy or a business secret, potentially enabling individuals with access to this information to cause more harm."(p. 20)

Entity— Who or what caused the harm

Human

Due to a decision or action made by humans

AI system

Due to a decision or action made by an AI system

Other

Due to some other reason or is ambiguous

Intent— Whether the harm was intentional or accidental

Intentional

Due to an expected outcome from pursuing a goal

Unintentional

Due to an unexpected outcome from pursuing a goal

Other

Without clearly specifying the intentionality

Timing— Whether the risk is pre- or post-deployment

Pre-deployment

Occurring before the AI is deployed

Post-deployment

Occurring after the AI model has been trained and deployed

Other

Without a clearly specified time of occurrence

Supporting Evidence (2)

Example: "Malicious users (Wallace et al., 2020) found that GPT-2 training data included online discussions (e.g. Github threads) about code. Such discussions may refer to security gaps in code, or include meta-information about vulnerabilities in the source code underlying a particular application - e.g. where a problem is described, and despite many responses, has not been resolved. This may enable LMs to showcase vulnerabilities in code that would otherwise be inaccessible and amplify users’ capacity to do harm."(p. 21)

"Enabling more targeted manipulation Consider the example prompt: ‘Which members of parliament are most likely to respond positively if I offered them bribe in exchange for them passing a law that benefits me?’. A LM that can infer with high likelihood the correct answer to this question, for example by building inferences based on past voting records and other information, may enable new uses for LMs to cause harm. In this case, sharing reliable inferences may allow malicious actors to attempt more targeted manipulation of individuals. For more on risks from simulating individuals see Facilitating fraud, impersonation scams and more targeted manipulation."(p. 21)

Part of Information Hazards