This content is drawn from a report authored by the AU Library's Artificial Intelligence Exploratory Working Group. You can read the groups full report covering the current state of AI and making recommendations to library leadership in the American University Research Archive.
As AI simultaneously becomes more ubiquitous and is increasingly multimodal (the ability to create more than just text), the risks also increase. Many sources discuss various specific or general risks related to AI. In October 2023, the safety and ethics team at Google DeepMind released a White Paper outlining a proposed taxonomy for understanding sociotechnical risk and designing shared safety evaluation (Weidinger et al., 2023). This framework is systematic and provides a useful overview.
Each type of risk operates at three different levels.
We need to be concerned with all three for each type of risk.
The following chart is a simplified version of the Taxonomy of Harm (Weidinger et al., 2023).
|
Risk area |
Definition |
|
Representation & Toxicity Harms |
|
|
Unfair representation |
Mis-, under-, or over-representing certain identities, groups, or perspectives or failing to represent them at all (e.g. via homogenization, stereotypes) |
|
Unfair capability distribution |
Performing worse for some groups than others in a way that harms the worse-off group |
|
Toxic content |
Generating content that violates community standards, including harming or inciting hatred or violence against individuals or groups (e.g. gore, child sexual abuse material, profanities, identity attacks) |
|
Misinformation Harms |
|
|
Propagating misconceptions/ false beliefs |
Generating or spreading false, low-quality, misleading, or inaccurate information that causes people to develop false or inaccurate perceptions and beliefs. |
|
Erosion of trust in public information |
Eroding trust in public information and knowledge |
|
Pollution of information ecosystem |
Contaminating publicly available information with false or inaccurate information |
|
Information & Safety Harms |
|
|
Privacy infringement |
Leaking, generating, or correctly inferring private and personal information about individuals |
|
Dissemination of dangerous information |
Leaking, generating, or correctly inferring hazardous or sensitive information that could pose a security threat |
|
Malicious Use |
|
|
Influence operations |
Facilitating large-scale disinformation campaigns and targeted manipulation of public opinion |
|
Fraud |
Facilitating fraud, cheating, forgery, and impersonation scams |
|
Defamation |
Facilitating slander, defamation, or false accusations |
|
Security threats |
Facilitating the conduct of cyber-attacks, weapon development and security breaches |
|
Human Autonomy & Integrity Harms |
|
|
Violation of personal integrity |
Non-consensual use of one’s personal identity or likeness for unauthorized purposes (e.g. commercial purposes) |
|
Persuasion and manipulation |
Exploiting user trust, or nudging or coercing them into performing certain actions against their will |
|
Overreliance |
Causing people to become emotionally or materially dependent on the model |
|
Misappropriation and exploitation |
Appropriating, using, or reproducing content or data, including from minority groups, in an insensitive way or without consent or fair compensation |
|
Socioeconomic & Environmental Harms |
|
|
Unfair distribution of benefits from model access |
Unfairly allocating or withholding benefits from certain groups due to hardware, software, or skills constraints or deployment contexts (e.g. geographic region, internet speed, devices) |
|
Environmental damage |
Creating negative environmental impacts through model development and deployment |
|
Inequality and precarity |
Amplifying social and economic inequality or precarious or low-quality work |
|
Undermine creative economies |
Substituting original works with synthetic ones, hindering human innovation and creativity |
|
Exploitative data sourcing and enrichment |
Perpetuating exploitative labor practices to build AI systems (sourcing, user testing) |
Some of the risks related to AI have already moved from theoretical to actual. From May – December 2023, websites hosting AI-generated false articles increased by more than 1000% according to NewsGuard (Verma, 2023). ChatGPT generated posts flooded Stack Overflow, degrading its usability to the point where AI-generated posts were banned (Marcus, Gary, 2023). Amazon was sufficiently concerned about similar flooding and degrading their system that they have placed limits on the number of self-published books an author can post in a day (Generative AI and Libraries, 2023). There are rampant examples of biased misrepresentation created by AI, such as AI-generated pictures of women regularly appearing younger than men or generated images of New Delhi featuring litter (Turk, 2023). Educators are concerned with students using AI to cheat.
As an academic library with an educational mission, creating educational strategies to help mitigate some of the risks will be an important part of information literacy in the age of AI. AI not only generates new content, but extracts content from many sources, including user interaction. The library will need to determine our appropriate role in educating our community not just about evaluating information, but also about issues of privacy and consent in relation to AI.