Uplift
Uplift: How AI Raises the Capability Ceiling for Harm
In early 2024, OpenAI ran one of the more peculiar experiments in AI safety. The company gathered fifty biology PhDs with wet lab experience and fifty undergraduate biology students, sat them in a controlled environment, and asked them to write a plan for creating and releasing a biological weapon. Half had access to the internet, while the other half had the internet plus an unrestricted version of GPT-4 that would answer direct questions about pathogens. Experts then graded the resulting plans across five measures, including accuracy, completeness and innovation. OpenAI’s verdict, after grading the results, was “at most a mild uplift”, with accuracy gains so small and so statistically insignificant that the company concluded GPT-4 had provided no meaningful advantage to would-be bioweapon designers [1].
Uplift came into AI from the biosecurity world, where it was used to talk about how new technologies like cheap DNA synthesis or published gain-of-function research might raise the ‘capability ceiling’ for bad actors. Measuring uplift isn’t straightforward. Gary Marcus, emeritus professor of psychology and neural science at NYU, reanalysed OpenAI’s biological weapon experiment data and reached a different conclusion. Focusing on one task in particular, the formulation of a pathogen, he found that four out of 25 experts with GPT-4 succeeded, compared to one out of 25 without. As Marcus put it, even one team building a pathogen as deadly as COVID would make it “really, really big deal” [2].
By 2025, the models had grown more capable and the uplift had grown with them. When Anthropic and OpenAI ran their own evaluations, the results were concerning enough that both labs decided they couldn’t release without more active safeguards. In May, Anthropic released Claude Opus 4 under what it called AI Safety Level 3 protections, adding systems to monitor outputs and block dangerous requests [3][4]. OpenAI followed with similar safeguards [5].
This decision captured the irresolvable dilemma that WIRED’s Steven Levy later summed up: “Either hold back and lose or jump in and put humanity at risk” [6].
References:
[1] OpenAI, “Building an early warning system for LLM-aided biological threat creation,” January 2024. https://openai.com/index/building-an-early-warning-system-for-llm-aided-biological-threat-creation/
[2] Marcus, G., “When looked at carefully, OpenAI’s new study on GPT-4 and bioweapons is deeply worrisome,” Marcus on AI Substack, February 2024.
[3] Anthropic, “Activating AI Safety Level 3 Protections,” May 2025. https://www.anthropic.com/news/activating-asl3-protections
[4] Anthropic, “System Card: Claude Opus 4 & Claude Sonnet 4,” May 2025. https://www.anthropic.com/claude-4-system-card
[5] OpenAI, “ChatGPT Agent System Card,” July 2025. https://openai.com/index/chatgpt-agent-system-card/
[6] Levy, S., “If Anthropic Succeeds, a Nation of Benevolent AI Geniuses Could Be Born,” WIRED, March 2025. https://www.wired.com/story/anthropic-benevolent-artificial-intelligence/