A summary of the draft paper co-authored by Timnit Gebru, which outlined the main risks of large language AI models and provided suggestions for future research
The company's star ethics researcher highlighted the risks of large language models, which are key to Google's business. — hide
Context & Ripple Effects
The draft paper co-authored by Timnit Gebru argued that large language models — the technology underpinning Google's search and ad business — carry substantial risks, and its disputed path to publication foreshadowed what became a defining fight over who controls AI ethics research inside corporate labs. The fallout was swift: within months, Wired documented how Google's Ethical AI team fell apart following Gebru's contentious departure, with Margaret Mitchell's exit soon after.
What makes this summary worth revisiting is how durable the paper's agenda proved. Two years on, MIT Technology Review's own follow-up catalogued the privacy risks raised by GPT-3, LaMDA, and Meta's OPT-175B, showing the concerns Gebru's draft raised had become the field's open questions.
First-order effects
- Google faces immediate reputational damage and an internal credibility problem: the episode put its treatment of ethics research under public scrutiny precisely because the paper targeted models central to its business.
Second-order effects
- Independent researchers respond by building research infrastructure outside corporate labs — over 500 researchers mobilized behind an open-source large language model explicitly designed for research no company can gatekeep.
Third-order effects
- If the pattern holds, LLM risk assessment migrates structurally from in-house ethics teams to independent and academic efforts, while the data practices the draft flagged re-emerge as external scrutiny once models like GPT-3, LaMDA, and OPT-175B are examined for privacy harms.
The trend: Corporate control over AI ethics research is pushing scrutiny of large language models toward independent, open-source channels as the technology commercializes.