What is distributed and what is measured
The programme gives away, at no cost, ChatGPT with higher usage limits and larger context windows, GPT-5.6 Sol and GPT-5.6 Sol Pro tuned for long-running tasks, an expanded deep research feature able to gather information from hundreds of sites or from scientific journals only, Codex and ChatGPT Work. Access starts with 10,000 researchers at a limited number of universities, the remaining 90,000 are added over time, and each invited scientist can add four colleagues from the same institution free of charge. It is described as part of a 250 million dollar initiative supporting scientific projects.[1]
No outcome measure is named in the announcement: not publication counts, not reproducibility rates, not time to a result, not error frequency in drafted analyses. Without a defined outcome and a comparison group, what the programme can report is uptake; whether scientific discovery accelerated does not follow from this design.[1]
Selection compromises the measurement in advance
The invitation chain recruits through existing collaboration networks: it starts at a limited number of universities and each invitee adds four colleagues. The participating population therefore differs systematically from the 100,000 researchers being targeted. If users are later compared with non-users, the tool's effect and the characteristics of labs that adopt early become entangled. The competing possibility: OpenAI may be running a separate evaluation of its own design that the announcement does not describe.[1]
This is a fixable design problem. If a pre-announced comparison for a single scientific task, with a defined outcome measure and a comparison group, is published by 30 April 2027, the acceleration claim becomes testable. The measure need not be grand: time from a question to a checkable analysis, plus the error rate in that analysis, would do.[1]
The difference AlphaFold shows
The result that made AI credible in science came from a single body of work whose method was published and independently used by others. The team that produced AlphaFold has now been dispersed: Google DeepMind broke up the team dedicated to the project without closing it, most authors were reassigned to Gemini model work, enzyme design, fusion and genomics, three went to Anthropic, and about 25 percent of the full-time authors left the company. What produced a checkable scientific result was a specified method; broad distribution of access has not yet produced evidence of the same kind.[1], [2]