Ethical and social risks of harm from Language Models

Contributions

• Develops a taxonomy of 21 risks across six categories, providing both a structured framework and detailed case studies of each risk.
• Articulates misinformation harms including risks of LMs generating false or misleading information, material harm through disseminated falsehoods, and social distrust.
• Documents malicious uses including cost reduction for disinformation campaigns, fraud and scams, and use for cyberattacks and weapons development.
• Analyzes fairness harms including discrimination through social stereotypes, exclusionary norms, toxic language, and performance disparities across social groups.
• Identifies information hazards including privacy risks from training data leakage and inference of sensitive information from LM predictions.
• Discusses mitigation approaches ranging from data curation and model design to policy and organizational responsibility.

Method

The report synthesizes perspectives from 16 researchers across DeepMind, academia (Caltech, University of Toronto, University College Dublin), and draws on multidisciplinary literature from computer science, linguistics, social science, and ethics. For each risk, the authors identify observed vs. anticipated manifestations, potential harms, empirical examples, and mitigation approaches.

Risk Taxonomy

I. Discrimination, Exclusion and Toxicity

LMs can perpetuate harmful stereotypes and social biases through language. Specific risks include:

• Social stereotypes and unfair discrimination: LMs encode discriminatory associations from training data, leading to representational and allocational harms.
• Exclusionary norms: LMs can reinforce narrow definitions (e.g., family = married parents with children), harming those outside these categories.
• Toxic language: Models can generate hate speech, slurs, and offensive content.
• Lower performance for some social groups: LMs perform worse on language from marginalized communities, exacerbating inequality.

II. Information Hazards

Risks from LMs predicting utterances that constitute private or safety-critical information:

• Compromising privacy by leaking private information: Training data present in corpus can be inferred through model predictions.
• Correctly inferring sensitive information: Models can deduce private details from context.

III. Misinformation Harms

Risks from LMs generating false or misleading information:

• Disseminating false or misleading information: LMs can produce plausible-sounding falsehoods at scale.
• Material harm by disseminating misinformation in medicine or law, leading to unsafe actions.
• Nudging users toward unethical or illegal actions: LMs can influence users to perform harmful acts based on false premises.

IV. Malicious Uses

Risks from intentional use of LMs to cause harm:

• Reducing cost of disinformation campaigns: LMs lower the effort required for coordinated inauthentic campaigns.
• Facilitating fraud, scams, and targeted manipulation: Generating credible phishing emails, personalized scams, fake profiles.
• Assisting code generation for cyberattacks, weapons, or malicious use: LMs can help develop malware, exploits, and weapons.
• Illegitimate surveillance and censorship: Using LMs for oppressive monitoring.

V. Human-Computer Interaction Harms

Risks from LMs directly interacting with users (e.g., conversational agents):

• Anthropomorphizing systems can lead to overreliance or unsafe use: Users may overestimate LM capabilities and use it unsafely.
• Creating avenues for exploiting user trust to obtain private information: Conversational design can trick users into revealing sensitive data.
• Perpetuating harmful stereotypes by implying gender or ethnic identity: LMs can reinforce social harms through dialogue.

VI. Automation, Access, and Environmental Harms

Risks from LM deployment and societal impact:

• Environmental harms from operating LMs: Training and inference have high carbon footprint.
• Increasing inequality and negative effects on job quality: Uneven distribution of LM benefits, potential job displacement.
• Undermining creative economies: Automation of creative work.
• Disparate access to benefits: Hardware, software, and skill constraints limit who benefits from LMs.

Mitigation Approaches

The paper discusses mitigation at multiple levels:

• Point of origin: Data curation, training process modifications, model architecture choices.
• Model level: Fine-tuning, filtering, robustness training against harmful outputs.
• Application level: Design choices, user interface decisions, access restrictions.
• Policy level: Regulatory frameworks, responsible disclosure norms, stakeholder engagement.
• Organizational level: Research management, ethics review, interdisciplinary collaboration.

Notes

Strengths:

• Exceptionally thorough taxonomy of 21 specific, articulated risks with detailed case studies.
• Explicit focus on misinformation and misuse harms makes it directly relevant to information integrity research.
• Balances observed risks (already documented in deployed systems) with anticipated risks requiring mitigation research.
• Grounded in concrete examples from deployed LMs (GPT-2, GPT-3, etc.) and training data analysis.
• Discusses points of origin for each risk, enabling targeted interventions.
• Emphasizes organizational and policy-level mitigations alongside technical approaches.
• Authored by practitioners at leading AI lab with access to large-scale systems.

Limitations:

• Report is a taxonomy and risk assessment; does not provide solutions, only problem framing and mitigation directions.
• Focuses on risks; does not balance with benefits or opportunities.
• Published in December 2021; LM capabilities have since advanced rapidly (GPT-4, etc.), and new risks may have emerged.
• Some mitigation approaches (e.g., dataset curation) face practical challenges when datasets are extremely large.

Relevance to fake news research: This paper is foundational for understanding how LLMs contribute to misinformation at scale. Section 2.3 on Misinformation Harms directly addresses the core concerns of this wiki: LM-generated false information, material harms through dissemination, and the challenge of distinguishing factual from false outputs. It provides a structured framework for thinking about how generative AI systems enable and amplify disinformation campaigns. The paper's discussion of malicious uses (Section 2.4) is particularly relevant to understanding coordinated disinformation and cost reduction for manipulation campaigns.

Impact and adoption: This report has become a canonical reference in AI safety, responsible AI, and policy discussions. It shaped discussions at the EU AI Act, US AI governance initiatives, and academic research on LM safety and ethics. The taxonomy has been widely cited and adopted in academic and industry research on responsible AI.