A good explanation can make us feel that we understand something. All the more so when the reasoning holds together, earlier research is cited, and the next experiment is already laid out.
Nature, however, is not in the habit of complimenting presentations. However elegant a hypothesis, cells need not behave as it predicts.
Co-Scientist, developed by researchers at Google and other institutions, sits between explanation and experiment. It does more than summarize papers: it proposes new hypotheses and ways to test them. Reading the study by Juraj Gottweis and colleagues, we encounter a division of labor in developing hypotheses before we reach the grand question of whether AI will become a scientist.
Not the first answer, straight away
Co-Scientist assigns different jobs to multiple agents built on Gemini 2.0. Some generate proposals; others criticize, compare, rank, or improve them. Other roles identify similar ideas or review the work as a whole. A supervisor distributes tasks in response to a scientist’s research goal. [1]
The distinctive feature is that the process does not stop after producing one plausible answer. Hypotheses are compared, weaknesses identified, and the results fed back into further generation. Scientists can also supply goals and constraints and change direction. The design looks less like removing the scientist from the chair than increasing the number of partners against whom an idea can be tested.
There is something attractive about this arrangement. An objection follows a thought immediately. Exploration reaches into neighboring fields we could never read in full. Before an idea is discarded, it can be reworked. Jobs that would queue up inside one head can run separately.
But adding critics is not the same as getting closer to reality. If they all overlook the same assumption, a lively discussion may simply circulate around it with great skill. The minutes grow longer. Not a line of the world has been rewritten.
Candidates that left the page
The team did not stop at the system’s internal rankings. It ran experiments on biomedical problems. In tests of existing drugs and compounds for a different use against acute myeloid leukemia, researchers tested candidates proposed by AI and selected after expert review. One, KIRA6, reduced viability in a cell line called KG-1a. Other cell lines differed markedly in sensitivity. [1]
That variation is not a minor footnote attached to the achievement. It is part of the point of doing the experiment. Even under the same disease name, effects are not uniform. Once a hypothesis touches the world, a different kind of unevenness appears beneath the smooth explanation.
Finding a candidate and confirming an effect in cells is progress. It does not mean a treatment effective in patients has been developed. What happens inside the body, safety, and the diversity of the disease remain beyond what cell experiments alone can answer. The paper itself acknowledges the difficulties of clinical translation.
Compress this into “AI discovered it,” and the distance is easily lost. A hypothesis was proposed. Experts selected it. An experiment tested it. Differences across conditions emerged. Keeping the verbs separate does not diminish the achievement; it makes its novelty clearer.
Not more criticism, but an exit toward falsification
The research includes human evaluation, although one part is a small comparison using 11 research goals. Internal tournament rankings are not independent judgments of truth. Biases in published literature, erroneous citations and reasoning, inaccessible knowledge, and performance in other fields remain concerns. Not all of the code is public, either. [1]
Here, then, is an editorial question. When AI can produce hypotheses in abundance, what becomes scarce? Perhaps not ideas, but opportunities to put one of them in contact with reality.
Experiments take time. They need equipment and samples. Even disappointing results must be recorded in a form useful to the next attempt. The cheaper and faster hypotheses become, the more consequential it becomes to decide which to test, what would falsify them, and what result would justify moving forward.
There are similar moments in new business planning. AIs can refine the explanation of why customers ought to want something. The customers’ actual response, however, is not present in that conversation. Improving an explanation and creating a situation in which it could fail are separate jobs. Accelerate only the first without connecting it to the second, and we risk becoming more efficient at manufacturing conviction.
Conversely, if test results feed into the next hypothesis, the speed of generation means something different. A proposal that fails becomes material for changing the territory of the search. This is a practical implication we might consider, not an effect on business management established by the study.
What is interesting about Co-Scientist is less the image of AI declaring the correct answer alone than the route by which a hypothesis is criticized, selected, and sent outside. The final judge need not be another eloquent AI.
If this hypothesis is wrong, what would we have to encounter to find that out?