What are the risks from AI?

This week we spotlight the twenty-fourth framework of risks from AI included in the AI Risk Repository: Gabriel, I., Manzini, A., Keeling, G., Hendricks, L. A., Rieser, V., Iqbal, H., Tomašev, N., Ktena, I., Kenton, Z., Rodriguez, M., El-Sayed, S., Brown, S., Akbulut, C., Trask, A., Hughes, E., Stevie Bergman, A., Shelby, R., Marchal, N., Griffin, C., … Manyika, J. (2024). The Ethics of Advanced AI Assistants. In arXiv. https://doi.org/10.48550/arXiv.2404.16244

‍

Paper focus‍

This paper presents the ethical and societal risks of advanced AI assistants, defined as artificial agents with natural language interfaces that plan and execute sequences of actions on behalf of a user – across one or more domains – in line with the user’s expectations.

Included risk categories

This paper provides an overview of AI challenges, as well as opportunities, organized into three main areas.

1. Value alignment, safety, and misuse: risks arising from the use of advanced AI assistants, involving topics such as value alignment, well-being, safety, and malicious use.

• AI assistants misaligned with user interests
• AI assistants misaligned with societal interests
• AI assistants imposing values on others
• AI assistants used for malicious purposes
• AI assistants vulnerable to adversarial attacks

2. Human–assistant interaction: risks arising from the relationship between advanced AI assistants and users, involving topics such as influence, anthropomorphism, appropriate relationships, trust, and privacy.

• AI assistants manipulating/influencing users to benefit developers or third parties
• AI assistants hindering users’ self-actualisation
• AI assistants optimised for frictionless relationships
• Users unduly anthropomorphising AI assistants
• Users becoming emotionally dependent on AI assistants
• Users becoming materially dependent on AI assistants
• Users at risk of harm if they have undue trust in AI assistants
• AI assistants infringing upon user privacy

3. Societal-scale impacts: risks associated with the deployment of technology at a societal level, involving topics such as cooperation, misinformation, equity/access, economic impact, and the environment.

• AI assistants encountering coordination problems leading to suboptimal social outcomes
• AI assistants leading to a decline in social connectedness
• AI assistants contributing to the spread of misinformation via excessive personalization
• AI assistants enabling new kinds of disinformation campaigns
• AI assistants causing job loss/worker displacement
• AI assistants deepening societal-level technological inequality
• AI assistants having negative environmental impacts

Key features of the framework and associated paper

• Provides a systematic and comprehensive analysis of ethical and societal issues introduced by advanced AI assistants
• Identifies an “evaluation gap” (in which existing approaches focus on model-level considerations) and makes recommendations for a suite of evaluations that take into account the broader sociotechnical system, including human–AI interaction, multi-agent, and societal effects.

⚠️Disclaimer: This summary highlights a paper included in the MIT AI Risk Repository. We did not author the paper and credit goes to Iason Gabriel, Arianna Manzini, Geoff Keeling, and co-authors. For the full details, please refer to the original publication: https://arxiv.org/abs/2404.16244.

Further engagement 

→ View all the frameworks included in the AI Risk Repository 

→ Sign-up for our project Newsletter