Artificial intelligence is starting to take on parts of the scientific process that once seemed out of reach – automated analysis of complex data, hypothesis generation, even the drafting of full scientific reports. One system, Kosmos, compresses months of exploratory research into a single overnight run.
What is Kosmos?
Kosmos is an “AI scientist” capable of independently searching the literature, analysing data, forming hypotheses and generating research reports with citations. It does this in roughly 12-hour cycles. Each cycle is said to be equivalent to 4-6 months of human research work.
Originally developed at a nonprofit research group funded by Eric Schmidt, the former CEO of Google, Kosmos is now commercialised through Edison Scientific, with more than 30,000 academic and biotech users. It’s an example of agentic AI, systems that take a goal, break it into steps, run the steps autonomously, and decide what to do next without waiting for human prompts.
To generate insights, Kosmos is given the same starting materials a human research team would need: a dataset, a research question and access to the scientific literature. It then runs hundreds of small analysis tasks, reads hundreds of papers, and writes tens of thousands of lines of code. As it works, it gradually builds an internal “world model”, a structured scientific memory that connects all the findings. Once this is in place Kosmos begins proposing hypotheses and assembling research reports.
How well does it work?
According to the arXIv paper:
- 79% of scientific statements were accurate
- 85% of data-analyses were correct
- 82% of literature review findings were correct
- Only 58% of interpretive statements were accurate
Together, these measures show Kosmos’s strengths and limitations.
What did Kosmos discover?
Cooling protects the brain by shifting cell metabolism
Kosmos was asked to identify how cooling protects brain cells using data from experiments on mouse brain tissue. The AI inferred that cooling shifts cells into a specific “energy-saving mode” that helps them survive stress, supporting the rationale for cooling patients after cardiac arrest.
A universal rule for how nerve cells wire themselves
Scientists asked Kosmos to uncover structural rules governing neuronal connections. Eight large datasets were provided, including synapse (connection) counts, wiring lengths and connection patterns in different species and brain regions. Kosmos rediscovered a known finding: most neurons have only modest numbers of connections, with a minority forming many. This pattern governs how the brain processes information and adapts after injury.
A gene that may protect the heart from scarring
Kosmos was asked to find genetic factors that reduce myocardial fibrosis – the build-up of stiff, scar-like tissue in the heart muscle. It received genetic association data, information about likely causal variants, and data on how those variants influence gene activity. Kosmos picked out certain variants of a particular gene (SOD2) as protective, suggesting a possible therapeutic pathway.
A genetic “dimmer switch” relevant to type 2 diabetes
Scientists asked Kosmos to identify protective mechanisms in type 2 diabetes. It was given datasets showing DNA variants linked to diabetes risk, genomic regions that switch genes on or off (enhancers), and gene activity correlations with disease. Kosmos identified a genetic variant that acts like a dimmer switch controlling how cells handle stress in the endoplasmic reticulum, the cell’s protein-folding and stress-management machinery. The protective form may slow diabetes progression.
Rethinking the sequence of damage in Alzheimer’s disease
Kosmos was asked to determine the sequence of neuronal damage in Alzheimer’s disease. Tau is an abnormal, misfolded protein that accumulates inside neurons. Using data from individual neurons, Kosmos inferred that the structural support around neurons deteriorates only after tau builds up. This suggests the structural decline is a consequence rather than an early cause of disease.
Why some brain regions age faster than others
Kosmos was asked to identify why certain neurons are especially vulnerable during ageing and in Alzheimer’s disease, using datasets from neurons in a memory-related brain region. It found that these neurons lose proteins that keep their membranes chemically balanced. Without them, they expose “eat me” signals detected by microglia – the brain’s immune cells. This mechanism may contribute to early neuronal loss in ageing and Alzheimer’s disease.
Conclusions
These findings are exploratory. None is definitive, and they don’t reveal fundamentally new biology. Several simply rediscover what scientists have already seen in the data; others provide plausible early-stage leads that may guide future experiments.
Nevertheless, Kosmos shows that AI can run analyses, generate hypotheses, cross-check its work and produce fully referenced scientific documents. In doing these tasks, Kosmos behaves like a scientific collaborator though it has limitations and requires human supervision.
Kosmos generates more leads than scientists can test
Kosmos can speedily generate dozens of plausible ideas, but scientists must determine which ones are credible. Validation is the bottleneck and the volume of hypotheses may exceed the capacity to test them.
Total analytical time
Some researchers question whether Kosmos truly performs “six months of work in 12 hours,” noting that validating the AI output may take just as long as a traditional analysis.
One in five statements is wrong, and interpretation is worse
A 15-20% error rate would be unacceptable in clinical but also laboratory settings. The interpretive error rate – over 40% – makes it clear that systems like Kosmos can’t operate without oversight. Reliability rather than speed is the limiting factor.
Not reproducible
The underlying code and model weights have not been released, meaning that independent researchers cannot rerun or verify Kosmos’s results.
The Bottom line
Kosmos is a sign of what’s already here and will become commonplace: biomedical research that includes in the team agentic AI systems able to perform months of exploratory analysis in a day.
Alongside optimism about these advances is a concern that use of systems like Kosmos leads to a future in which researchers don’t fully read the papers they cite or fully write the papers that describe their own studies, potentially weakening the understanding that drives good experimental work.
Kosmos is a tool, not a scientist. Its interpretive weaknesses show why human oversight is essential. It does not understand biology in any meaningful sense; it is pattern-matching at scale.
For clinicians, the message is that AI will accelerate the accumulation of scientific knowledge but that reliability must be scrutinised and standards of evidence upheld.
Understanding how tools like Kosmos work, also their limitations, will become critical parts of evidence-based medicine.
The challenge is to build the literacy, guardrails and judgement needed to use these tools wisely.
My Summary
What Kosmos shows:
- AI can scan literature, analyse datasets and propose hypotheses
- It can replicate and extend scientific findings
- It accelerates exploration but cannot replace experiments
Implications for healthcare:
- Faster biomarker and drug-target discovery
- New tools for understanding disease mechanisms
- New skills needed to evaluate AI-generated science
Be cautious:
- Accuracy is limited
- Interpretations are often wrong
- AI-generated findings require experimental validation
AI Summary (Gemini AI)
