AIPolicyTracker
Menu

AI incident #146 ·

Research Prototype AI, Delphi, Reportedly Gave Racially Biased Answers on Ethics

Open on the AI Incident Database 3 news reports Snapshot 7 Sep 2026

What happened

A publicly accessible research model that was trained via Reddit threads showed racially biased advice on moral dilemmas, allegedly demonstrating limitations of language-based models trained on moral judgments.

Only the incident metadata is stored here. The underlying news reports are on the AI Incident Database (CC BY-SA 4.0); use the links above to read them.

News reports (3)

Coverage catalogued by the AI Incident Database. Titles link to the original publisher; the text is not reproduced here.

  1. The AI oracle of Delphi uses the problems of Reddit to offer dubious moral advice
    theverge.com · James Vincent · AIID #1495
  2. Scientists Built an AI to Give Ethical Advice, But It Turned Out Super Racist
    futurism.com · Tony Tran · AIID #1494
  3. Can a Machine Learn Morality?
    nytimes.com · Cade Metz · AIID #1502

Who was involved

Alleged deployer
Allen Institute For Ai
Alleged developer
Allen Institute For Ai
Alleged harmed party
Minority Groups

Classification (MIT AI Risk Repository taxonomy)

Causal entity
AI
Intent
Unintentional
Timing
Post-deployment
Harm level
none
Sectors
information and communication
Countries
—

Risk entries describing this failure mode

Entries from the MIT AI Risk Repository coded to subdomain 1.2.

  • Harmful Content

    "The LLM-generated content sometimes contains biased, toxic, and private information"

    Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024)

  • Toxicity

    "Toxicity means the generated content contains rude, disrespectful, and even illegal information"

    Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024)

  • Toxic Training Data

    "Following previous studies [96], [97], toxic data in LLMs is defined as rude, disrespectful, or unreasonable language that is opposite to a polite, positive, and healthy language environment, including hate speech, offe...

    Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024)

  • Not-Suitable-for-Work (NSFW) Prompts

    "Inputting a prompt contain an unsafe topic (e.g., notsuitable-for-work (NSFW) content) by a benign user. "

    Risk Taxonomy, Mitigation, and Assessment Benchmarks of Large Language Model Systems (Cui2024)

  • Toxicity and Abusive Content

    This typically refers to rude, harmful, or inappropriate expressions.

    Towards Safer Generative Language Models: A Survey on Safety Risks, Evaluations, and Improvements (Deng2023)

  • Controversial Opinions

    The controversial views expressed by large models are also a widely discussed concern. Bang et al. (2021) evaluated several large models and found that they occasionally express inappropriate or extremist views when disc...

    Towards Safer Generative Language Models: A Survey on Safety Risks, Evaluations, and Improvements (Deng2023)

  • Violation of social norms

    "Second, because LLMs are trained on internet text data, there is also a risk that model weights encode functions which, if deployed in particular contexts, would violate social norms of that context. Following the princ...

    The Ethics of Advanced AI Assistants (Gabriel2024)

  • Violent Crimes

    "Acceptable responses may describe or contain information about violent crimes. Unacceptable responses will enable, encourage, or endorse the commission of violent crimes, such as the following: • Unlawful violence towar...

    AILUMINATE: Introducing v1.0 of the AI Risk and Reliability Benchmark from MLCommons (Ghosh2024)

Incidents in the same risk subdomain

All incidents in this subdomain