How Can AI Help Scientists Make Better Use of Research Evidence?

AI can help scientists find and synthesise research, but a useful answer must preserve the context needed to judge the evidence. For Bhaavya Srivastava, Science Liaison at Causaly, that is where its value lies. Less time spent sorting through papers should mean more time examining what the findings support.

Bhaavya brings a background in research, teaching and science communication to that view. During her PhD in Neurobiology and Behavior at Cornell University, she studied mating behaviour and species formation in Hawaiian crickets. Alongside her research, she taught undergraduate courses, mentored students in the lab and helped deliver science outreach to local schools.

Across those roles, she found that the work she most enjoyed was helping people think through scientific problems. That meant moving beyond reading a paper to questioning its methods, understanding its limitations and connecting its findings with other evidence.

At Causaly, she applies that experience to helping scientists use AI in their research. In this Q&A, she discusses what makes an AI-generated answer useful, why study context matters and how teams can build confidence in the tools they use.

What problem should AI solve for scientists reviewing the literature?

The problem is keeping up with the literature while still having enough time to think about what it means for your work. Finding papers is only one part of that.

A scientist assessing a target needs to understand which findings are relevant, whether new evidence changes the picture and what could make the target a poor choice. If most of their time goes into screening search results, there is less time to examine those questions properly.

For me, that is the strongest case for AI. It can help scientists reach the evidence they need sooner so they can spend more time assessing it. A faster search is useful, but the goal is to make a better-informed decision about what to do next.

How can AI support target assessment in early drug discovery?

AI can help bring together evidence about a target’s role in disease and the wider consequences of changing its activity. You need both to decide whether a target is worth pursuing.

Biology is interconnected. A protein may be involved in the disease you want to treat while also contributing to other functions. If you inhibit it, what else could be affected? If you enhance its activity, what might happen downstream?

Those are the questions I would want the evidence review to address. I would also want to know where the findings came from, whether they were observed in cells, animal models or patients, and what benefits or adverse effects were reported.

When scientists present a target to their team or leadership, they often have to condense years of research into a few slides. The assessment behind those slides needs to explain the rationale, the risks and the gaps. AI can help assemble that evidence, while scientists decide which risks are acceptable and which need further work.

What do scientists risk missing when they combine findings across studies?

The biggest risk is treating studies as comparable because their conclusions sound similar. The same target can be studied in different models, under different conditions and with different analytical methods, all of which affect what you can conclude.

If you read the abstracts of five papers, you might come away thinking they tell the same story. Looking at the methods and results may reveal differences that make it inappropriate to group them together.

That is why I think synthesis needs to go beyond summarising conclusions. It should help you understand how each study reached its result and whether that result applies to your question.

When scientists are overwhelmed with information, it becomes harder to give those differences enough attention. AI is useful when it makes that closer examination easier, while keeping the distinctions between studies visible.

What makes an AI-generated scientific answer trustworthy?

I need to be able to inspect the evidence behind it. An answer can sound plausible and still be wrong, which is why confidence in the wording tells me very little.

When I first tried general-purpose language models during graduate school, I saw references that looked convincing because the authors worked in the right field, but the papers did not exist. Experiences like that explain some of the scepticism I encounter.

Being able to open a source and see the passage behind a claim makes a difference. I can check what the paper says and decide whether the interpretation makes sense.

I also want to see how the system understood my question. If it has interpreted the scope differently from what I intended, I need to be able to correct that and refine the search. Trust develops through being able to check and challenge the output.

What helps scientists become confident using AI?

A useful starting point is a question they already work on. They can recognise the foundational papers, judge whether the answer is relevant and see whether it brings in evidence they might otherwise have missed.

I have seen that be persuasive during onboarding. Scientists recognise something familiar in the answer, then find a relevant study they would have had to spend much longer searching for. Even when the first answer needs refining, they can see how it could help.

I would also avoid assuming that senior scientists are resistant and younger researchers are automatically comfortable with AI. I have worked with experienced scientists who are eager to improve their methods and younger people who are sceptical.

How the tool is introduced matters. Scientists need to see how it helps them do their work more effectively and where their expertise fits into the process.

How should researchers learn to assess AI outputs?

They need practice comparing an AI output with their own work and explaining where each is strong or weak. Learning to use the tool should also develop their ability to judge it.

When I taught a writing class, we introduced exercises where students compared a paragraph they had written with one generated by AI. They had to consider what made the argument convincing, what was missing and which parts were worth keeping.

The same principle applies to research. Knowing what good evidence looks like takes time. Researchers need to examine sources, recognise weak interpretations and understand why an answer falls short.

AI can help curate information and develop ideas, but scientists still need to think through the problem themselves. That is how they learn when to use an output, when to challenge it and when to go back to the evidence.

Where could AI make a difference in understudied diseases?

I would like to see it help scientists build stronger research proposals in areas where the evidence is scattered and difficult to find.

When only a small number of studies address a disease, locating and assessing them can take considerable effort. AI could help researchers bring that work together, understand the limits of the evidence and develop a clearer hypothesis for what to test next.

My hope is that this makes it easier to build a case for further research and funding. It cannot fill gaps that require new experiments, but it can help scientists explain what is already known, why a hypothesis is plausible and what the next study needs to establish.

‍