Who Should Stop Unethical A.I.?

ai-ethicspeer-reviewai-safetyresearch-governance

Abstraction: Debate over ethics review mechanisms for AI research publication

Key points:

NeurIPS 2019 Speech2Face paper sparked debate about whether voice-to-face prediction research "should exist," highlighting the lack of ethics checkpoints in CS compared to biology, psychology, and other fields
NeurIPS 2020 implemented broader-impact statements and a second-layer ethics review panel; ~4 technically strong papers were rejected out of ~10,000 submissions; some reviewers flagged dozens of papers
Key categories of concern: weaponizable AI (facial recognition, surveillance), identity-hardening systems, automated weapons, and fake-content generators
Computer science lacks IRB-equivalent oversight; IRBs are explicitly barred from evaluating "long-range effects," and most CS research doesn't involve human subjects by traditional definitions
OpenAI withheld its GPT-2 text synthesis code citing misuse risk; the comment-generation paper from Microsoft/Chinese state lab added harm acknowledgments post-publication
Joe Redmon (YOLO creator, 25k+ citations) quit computer vision research entirely due to military/surveillance applications

Connections: Neurips · Openai · Google · Deepmind · AI Ethics · AI Safety · Bias Fairness