AI will save us time, but we still have to think for ourselves

Michaela Liegertová has long been at the intersection of science, education, and new technologies. Originally a cell and molecular biologist, she is one of the prominent voices in the field of meaningful use of artificial intelligence in research and academia. She comes to IBT with the task of helping us better understand AI, setting up its safe and practical use, and showing where it can really save us time or open up new possibilities. 

We talked about what her new role will bring to employees, where to start, and how to approach AI without unnecessary fears or uncritical enthusiasm.

You come to us with the task of helping colleagues use AI more meaningfully. What do you think should be the first practical change they experience? Where can AI save them time the fastest?

My role should be very practical, so that they can come to me with a specific use case and together we can choose a suitable tool, workflow, or conversely, discuss where caution is needed. AI can save the most time where routine is currently a waste of energy: filling out and reformatting documents, preparing reports and grant messages, converting notes into structured text, navigating calls, but also routines close to science itself. In grants, it can be very useful, for example, in deciphering agency requirements, comparing a draft with evaluation criteria, or simulating the critical perspective of an evaluator.

AI shouldn't do science for us, it should give us back time for what it can't do for us: interpretation, decision-making, formulating good questions, and scientific judgment. It makes sense to delegate routine. Not responsibility.

AI can do a lot of things today, but not everything that can be done technically is automatically a good idea. How do you set the right rules for when and how to use AI, but also know where caution is needed? What advice would you give to someone who is not sure whether AI is appropriate for a particular task?

It makes no sense to build rules as a long list of prohibitions, but as a practical way of making decisions. The very first question is: what kind of data are we working with? Is it a public text, an internal document, an unpublished result, personal data, clinical or genetic information? Only then does it make sense to choose a tool. In AI, a simple rule applies: data and risk first, then tool.

The second question is what impact the output will have. When AI helps with the style of a regular email, the risk is different than when it suggests analysis, interprets data, or helps formulate a scientific conclusion. The closer the output is to a scientific claim or institutional decision, the stronger the control needs to be.

The third question is: can the output be verified? This is fundamental for scientific work. AI can be a great working and thinking partner, but its output is not true just because it sounds good. In science, it is necessary to take AI outputs as hypotheses to be verified, as a thought substrate for one's own critical evaluation, not as ready-made conclusions.

Practically speaking: for public and low-risk tasks, we can be quite bold. For internal documents and unpublished results, we already have to think about the environment, license, and tool settings.

And now, since you're working on how to use AI in a smart and thoughtful way, I've allowed myself a little experiment. The following three questions were created with the help of ChatGPT as an example of how AI can help you find a slightly different perspective.

You talk about AI being able to create a kind of “virtual expert team” for scientists. If each of us were to build a small AI team around us for everyday work, what roles do you think should be included?

A virtual expert team is not one chatbot that knows everything. It's more like a small orchestra of tools and roles. One tool can act as a research expert and help map the literature or sources. Another as an editor who will refine the text, translate it or simplify it. Another as a methodological consultant or assistant for working with data and code who will help with analysis, checking tables or reproducing calculations.

A role that should never be missing is that of a critic. Regular chatbots tend to praise and encourage. They are trained to be pleasant and helpful. If the user does not ask for criticism themselves, they often do not receive truly beneficial feedback. And yet in science we need the exact opposite, someone to tell us where the argument is weak, what alternative explanations we are overlooking, what is not sufficiently substantiated, and where we are just telling ourselves a nice story. That is why we need to be able to ask AI specifically about weaknesses, not just suggestions for solutions.

Research projects should include an auditor and a validator. Roles that independently verify whether the methods used are relevant and perform reproducibility tests. And in biological projects, specialized models for sequences, proteins, molecules, single-cell data or image analysis come into play.

The person in this team does not disappear. On the contrary, they are even more important. They must define the question, provide context, decide what to believe, and most importantly, be responsible for the outcome.

You warn of the erosion of expertise: that if AI explains everything to us quickly, we can lose depth of understanding. How do we know in practice the difference between when AI is really helping us grow and when it starts to think for us?

A good test is a simple question: do we understand the problem better after using AI than before? If so, AI has helped increase competence. If the result is just smoother text, a nicer answer, or a more impressive presentation, but lacks the ability to explain why it is correct, AI is starting to think for us.

A lot depends on how we ask. “Do it for me” is a completely different type of usage than “help me understand this,” “explain the procedure to me,” “find weaknesses,” “suggest alternatives,” or “test me out on this.” In research, it is even more sensitive. AI can now produce hypotheses, graphs, code, research and interpretation very quickly. This can be useful, but also treacherous, because smooth output can create an impression of certainty. AI can also help in the opposite way, forcing us to look at a problem from multiple perspectives and pointing out possible blind spots, alternatives or points of view that we would not have thought of on our own. True expertise is recognized by not being satisfied with the first answer: it can verify the output, compare it with alternatives, re-run the analysis, find a weak point and sometimes even discard an attractive conclusion that does not hold up.

So AI doesn't automatically develop expertise. It develops it when it forces you to ask better questions and verify more thoroughly. It weakens it when it's enough for the answer to sound good.

You often talk about transparency and reproducibility. But in everyday work, people are not going to write a methodological appendix to every email or presentation. What do you think “reasonable transparency” should look like when using AI in everyday tasks?

Reasonable transparency means saying enough so that the other person understands how AI influenced the outcome. It's not a confession or a methodological appendix to every email. When AI just helps with the formulation of a regular message or the style of the text, there's no need to make a drama out of it.

But for more important outputs, the role of AI should be traceable and defensible. It makes a difference whether AI just edited the language, or helped with research, wrote the first draft of a section of text, generated code, designed an analysis, interpreted data, or influenced a scientific conclusion. The closer AI gets to the content, decision-making, or scientific claim, the more specific the description of its role should be.

In practice, for common low-risk tasks, common sense is enough. For grants, manuscripts, analyses, institutional outputs or materials that go out, it is good to have a simple record of what tool was used, for what, when and what was verified. For data analysis, this also includes the model version and archived code so that the result can be repeated. In fact, something like a “laboratory diary” of how the AI ​​came to work. For scientific output, it is not enough to write the name of the model, it is more important that the path from the data to the conclusion can be traced: where the data came from, what the AI ​​processed from it and what the human verified. Transparency is a condition for trustworthiness.

And here's another thing. If we create a stigma around AI, people won't stop using it. They'll just stop acknowledging it. That's worse for an institution than open, reasonable, and reasonably set rules.

A few words in conclusion

In practical terms, AI makes sense where it speeds up work without compromising our control over what we claim. A good approach is simple: try it on tasks where the output can be assessed, give it enough context, demand criticism, and for anything that goes beyond or approaches a scientific conclusion, verify the path from data to claim. Almost anyone can produce text or a graph these days. What makes the difference is what comes after: validation, verification, and the willingness to say “I don’t know this yet, I need to check this, and I can defend this.”