AI can review literature, debug code, and detect patterns faster than a human. Yet it does not understand context, cannot be held accountable, and may homogenize the way scholars think and write. Prof. Christopher Hartwell examines where the usefulness of the tool ends and the researcher’s responsibility begins.
This article is an edited version of Prof. Christopher Hartwell’s opening keynote, “Artificial Intelligence Meets Natural Ignorance: Thinking Responsibly about Research in a ChatGPT World,” delivered at the 2026 Responsible Research Summit at Kozminski University. The talk has been shortened and reorganized for publication. Repetition, digressions, and stage-specific references have been removed while preserving its argument, examples, and authorial voice.
AI is an excellent intern. It cannot be left on its own
We live at a curious moment in the history of knowledge. Never before have researchers had access to so much information or to tools capable of processing it so quickly. AI systems can summarize literature, generate hypotheses, write code, translate texts, and even draft entire sections of a paper.
Convincing? ChatGPT wrote that introduction. I asked it to open my keynote and received a polished, elegant passage. A few small changes were enough to make it sound like me. That is precisely the problem.
Artificial intelligence can help us overcome our ignorance about the world and the mechanisms that shape it. It can review literature, write code, and help us clarify an argument. We also know that it can be abused: students submit essays written by language models, reviews are automated, and researchers send AI-generated work to journals.
Malfeasance, however, is only part of the problem. People looking for shortcuts found them before AI. The more serious risk concerns those who act ethically but delegate too much to technology. Over-reliance on AI can cost us our voice. Yet our voice, curiosity, and passion are precisely what make us researchers. I also worry that AI may homogenize science not only within disciplinary silos, but across them.
The most useful way to think about AI is as an intern. A good intern can perform a great deal of valuable work, but cannot be left unsupervised. The task must be defined precisely, the context supplied, and the result checked. That requires a substantial initial investment. Only then does the tool genuinely begin to save time.
What AI is genuinely useful for
I use AI in my own research. One of my master’s students used it to learn R, Stata, and web scraping. He began without those skills and, step by step, built the toolkit he needed for his thesis. To me, this was an exemplary use: the technology did not conduct the research for him. It enabled him to learn new methods.
I have worked in Stata for many years. When code breaks, I ask a model to help locate the error. Once, after hours of unsuccessful debugging, AI noticed that an Excel file had been imported with a one-cell offset. In another project, it helped process photographs of nineteenth-century Russian banking newspapers taken in an archive in Helsinki. Advanced OCR and AI tools extract balance-sheet data from the images. A human still has to collect the material and assure its quality, but no one has to transcribe every table by hand.
AI also makes it easier to translate large collections of documents. I speak Russian, but not well enough to work effortlessly with an extensive set of materials on Soviet nationalities policy from 1931 to 1933. Machine translation allowed me to see the material more quickly in the language in which I think most fluently. And when a journal, after five rounds of review, asked me to convert seven pages of references from APA to Chicago style, ChatGPT completed the mechanical task in moments.
These are not the most spectacular applications. AI supports disease diagnostics, image recognition, time-sensitive data analysis, market surveillance, and web crawling. Students in Lausanne developed an application that identifies mites threatening honeybees from photographs. The common thread is clear: AI is strong at visualization and detection, pattern recognition, combining information from multiple sources, rule-based tasks, and work that must be carried out quickly at scale.
That is why the intern analogy matters. AI is excellent at ordered, repetitive tasks. It does not follow that it understands why a task matters or what a result means for the research project as a whole.
Publication pressure creates room for abuse
The dark side of AI is well known. In one study of undergraduates at the University of California, Berkeley, a quarter of daily AI users admitted to cheating. In one of my own project instructions, I hid a sentence in white font. It asked students to discuss Schumpeter’s dog Ellie, supposedly the inspiration for most of his work. Schumpeter probably did not have a dog called Ellie. Ellie is my dog, and I can assure you that she inspires no economic theories.
A student reading the document could not see the hidden line, but a language model could. Of roughly seventy students, only one ended his essay by telling me about Ellie. It was reassuring that the others passed the test. It does not change the fact that instructors now need new ways to identify unreflective use of these tools.
AI abuse does not stop at students, and the problem also concerns researchers, editors, and reviewers. According to a Nature report, signs of AI-generated text appeared in 14 percent of biomedical abstracts, while low-quality papers exploiting public datasets and automation are increasing. The pressure to “publish or perish” creates powerful incentives to use paper mills and other shortcuts. Algorithms can also be used to manipulate citation metrics on which promotion and evaluations still depend.
An absurd AI-generated illustration can pass editorial and peer review even when its labels are meaningless. An academic can use AI to write an article warning students not to use AI. These examples do not prove that the technology should be rejected. They show how easily procedures fail when pressure for speed and productivity becomes more important than judgment.
We should not panic, however. Sometimes the instructor’s expectations are themselves narrow or biased, and a model’s answer exposes a weakness in the assignment. We need rules and transparency, but we also need an honest assessment of what we actually expect from students, authors, and reviewers.
Research is more than pattern recognition
Research is detective work. We begin with a problem and try to shine light into an area we do not yet understand. We collect data, analyze and interpret it, and present a view of how a particular slice of the world works. We ask about mechanisms and channels, and whether an observed relationship is correlation or causation. We design an experiment or a research strategy and then defend a position grounded in evidence.
This is why we present at conferences, defend dissertations, and submit to peer review. We challenge methods, assumptions, theoretical frames, and claimed contributions because that is how science works. Sometimes the data support our priors and sometimes they contradict them. In either case, knowledge can advance.
AI helps with a substantial part of that work, but the rest reveals its weaknesses. A model recognizes patterns, yet often misses context. A prompt can be revised repeatedly, but at some point it would be quicker to write or analyze the material yourself. A language model is a probabilistic machine, not a database. It produces the answer that is most likely given the available material. That does not make the answer true.
The quality of the output depends on the training data and may reproduce their biases. Elegant language does not guarantee understanding either. AI favors superlatives, short paragraphs, and polished sets of three. The prose may appear professional, but it increasingly sounds the same regardless of the author. If early-career researchers, especially those writing in another language, allow a tool to replace the process of learning how to communicate, we deprive them of the opportunity to develop that skill.
The same limitation appears in visualization. I once needed a map showing the route from Zurich to Astrakhan. The image looked convincing, except that it placed Astrakhan in Kazakhstan rather than Russia. I tried four models and spent an hour attempting to move the city to the correct location. The map was attractive, but it lacked a reliable data layer. It was a useful demonstration of the difference between an output that looks right and one that is right.
Homogenized science loses its voice
When many people ask the same model for a quick solution to a similar problem, they receive similar answers. A simple marketing exercise makes this visible: candidates who use AI to prepare a campaign for Swiss chocolate arrive with polished but nearly identical designs. Models are improving and users are learning to prompt them more effectively, yet the risk of convergence toward the mean remains.
Science already homogenizes researchers to some extent. Every discipline socializes us into a body of literature, a set of methods and tools, and a particular seminar culture. We learn where to publish, how to review, how to teach, and how to gather evidence. This process is imperfect, but it emerges through experience, disagreement, and trial and error. A discipline is an ecology of plans: different ideas about where it should go meet, compete, and correct one another.
AI can apply a pattern from the past, but it does not know why that pattern emerged. It may also recommend the newest technique without understanding whether it suits the question. It can dig, but the researcher must point it in the right direction. Otherwise, the result is simply a field full of holes.
A model also lacks the full tacit knowledge embedded in scholarly practice. A doctoral researcher needs to know not only what has been published, but which questions matter to a community, what particular editors are seeking, and how a debate is developing at conferences. An analysis of past publications may reveal patterns, but it cannot replace a conversation with an editor, supervisor, or colleague.
AI operates on language, but it did not create language. It manipulates maps of knowledge built by people, but it does not explain how the territory first became intelligible. It is effective at discrete tasks. Academic work, however, is not a collection of discrete tasks. It is a process that connects questions, data, meaning, accountability, and dialogue.
A tool must not decide the direction
An IBM training manual from 1979 stated: “A computer can never be held accountable. Therefore, a computer must never make a management decision.” The principle remains relevant. AI should remain a tool for our passions and grand challenges. It must not decide where the research goes.
Overuse of AI also imposes a linguistic norm that did not arise organically within a community of scholars. At the beginning of my keynote, the model offered to make my introduction more humorous. It could have done so, but it would not have been my humor. Every researcher has a distinctive way of asking questions, building arguments, and explaining results. If writing is someone’s strength and contribution to a research team, why outsource it to a machine? AI can be a sounding board for ideas, but it should not take authorship of our thinking.
Responsibility rests especially with experienced researchers. We are guides for undergraduates, doctoral students, postdocs, and early-career colleagues. We need to show them how to use AI responsibly, not merely forbid the tool. Qualitative research and novel forms of data collection may gain in relative value precisely because they demand more than the processing of what already exists.
AI’s capabilities will continue to multiply, and so will its pitfalls. Researchers therefore need to be proactive, cautious, and reflective. We must understand not only what AI can do for us, but what it can do to us.
We must never forget that AI is a tool. It should help us overcome our natural ignorance rather than reinforce it.
