Generative AI may help science move faster. It can find drug leads, design proteins, and help sort huge sets of lab data. But the same tools may also create false clues. In some cases, those clues could look like real biology.
Generative AI learns from past data. It then uses what it learns to make new data or ideas. This can be useful when scientists face data sets that are too large to study by hand.
Researchers now use AI in many parts of biology. It can help design proteins. It can model cells. It can fill gaps in lab data. It can also make synthetic data for tests.
Yet AI can make what are called hallucinations. A hallucination is an output that looks sound but is not true. In biology, this could be a false link between two genes. It could also be a false sign that a drug works.
That risk may lead to real harm.
A false AI result could cause a useful drug to be dropped. It could also push a team toward a drug that does not work. In a worse case, it could make a false disease process seem real.
Thomas Burger, a computational biologist at Grenoble Alpes University in France, studied this risk in 10 uses of generative AI. His work was published in the journal Patterns.
Burger says the level of risk can vary a lot. The key point is how the AI result is used.
If AI gives a test idea, the risk is lower. Scientists can take the idea to a lab. They can then check if the result is real.
Drug and protein screening is one example. AI can check many possible drugs or proteins in a short time. It can then give scientists a smaller list to test.
An AI error can still cost time and money. A good drug may be missed. A weak drug may get more study. But a lab test can show if the drug really works.
The risk grows when AI made data is used as proof.
Synthetic data can be useful. It can fill missing data. It can help protect patient privacy. It can also make some studies cheaper. In some cases, it may help cut the need for animal tests.
But fake details in such data can cause a serious problem.
If AI adds a feature that was never in the real data, a team may see a false effect. That false effect may then shape a study or a claim.
This is not always easy to spot. The problem may not look like a clear AI error. Instead, a real data set may slowly change as it moves through a long AI based process.
That point is key to Burger’s warning. Scientists often start with raw signals from complex tools. Those signals must be cleaned and studied before they can show a clear biological process.
AI may help with each step. But a false change at one step may affect all later steps.
The final result may still look normal. Yet the path to that result may contain an AI made error.
This creates a new challenge for science. Researchers need ways to check AI made results at each stage. Real lab tests remain vital when a finding could change care, drugs, or basic science.
AI can speed up research. It can also help scientists explore ideas that may be hard to test by hand. But its output must not be treated as proof on its own.
The issue is not that AI is always wrong. The issue is that a wrong result can fit the data and the story too well. That makes it harder to spot. A human may trust the result because it seems to match other signs.
For that reason, careful checks matter. Teams may need to compare AI results with raw data, repeat key tests, and use lab work before making a strong claim.
The main lesson is simple. AI can suggest a discovery. It cannot, by itself, show that the discovery is real.
