“Thinking is meant to be slow and painful”

Text terhi hautamäki images Outi Kainiemi english translation marko saajanaho

The journey of a research idea from imagination to concrete research has become shorter.

This is University of Jyväskylä postdoctoral researcher Risto Turunen’s short response to the question of how his work has changed. As far as he is concerned, language models are not just office tools and are right at the core of research itself. He researches political language and has utilised language models in the study of the history of democracy and visions for the future.

“Initially, digital human sciences were mostly based on counting words. Language models can be used to analytically calculate more interesting phenomena such as emotions or temporalities in texts”, Turunen says.

A machine learning model turns historical research data into machine-readable data, even handwritten archival data. Language models also assist humanists with coding for simple data analysis.

”Language models can be used to analytically calculate more interesting phenomena such as emotions or temporalities in texts.”

In the Political Temporalities (POLTE) project, three researchers categorised two thousand sentences from parliamentary speeches based on whether the sentences referred to the future. Afterwards, the language model categorised six million sentences while mimicking the researchers’ thinking.

In addition to such guided uses, Turunen also uses language models without guidance, with the machine automatically recognising structures such as similarities between parliamentary speeches from different countries. Quantitative methods can find regularities that would otherwise be difficult to identify. Traditional close reading also becomes more effective.

“When, for example, the parliament’s speeches have been converted intto a vector database, you can take the tens of thousands of speeches and find the ones in which democracy is said to be at risk.”

A customised model is often better

Large language models have been prominent for about three years. Universities have also developed their own tools, largely based on commercial large language models. According to Turunen, other solid options exist as well. For example, he uses TurkuNLPs FINBERT, suitable for Finnish-language data, and the open source Poro language models.

Turunen says the gap between open and closed commercial models has closed significantly.

“In many categorisation problems, a model customised just for that problem does better than energy-guzzling large models.”

Researchers are pondering the question of how much AI can be used while still having the researcher remain in the driver’s seat. A purely AI-generated text can be recognised by generic, surface-level ideas which fail to develop further. Brainstorming ideas back and forth is also still done better in the break room than with a language model.

The application of AI methods requires field-specific consideration. In history research, a language model trained on data from the internet age can lead to anachronistic interpretations. When the research can be scaled to cover longer periods of time or more countries at once, increasingly diverse content knowledge is required.

Turunen learned the basics of text mining when working on his doctoral thesis. In postdoc projects, he has learned new skills as part of larger research groups. In his opinion, it would be good to include the basics of machine learning in human sciences education.

“I think they are part of a history researcher’s expertise just like archival work or scientific writing.”

Is AI being used correctly? 

AI methods are often best suited for tasks outside a person’s core knowledge: quickly going through large masses of data and identifying regularities. On the other hand, the emergence of language models raised hopes of more efficient routine work. They assist in text formatting, information searches, creating literature reviews, formatting presentation materials, proofreading, and translation. Ideally, AI would handle boring work to give us time for meaningful tasks.

“It seems we have not been able to adhere to this guideline very well”, University of Helsinki Professor of Computer Science Teemu Roos assesses.

“People often use AI for things that should be our core competencies: thinking, writing, idea evaluation, and peer review.”

Roos has seen hints of AI use by article reviewers, which naturally is not entirely positive. This is a larger issue of the massive increase in the quantity of articles and AI output resembling them, which already sees the human work-based peer review system in an acute crisis.

”People often use AI for things that should be our core competencies: thinking, writing, idea evaluation, and peer review.”

“With more and more input, there are not enough reviewers with time to look at it properly. That hurts carefully prepared articles, whose publication is delayed or, in the worst case, stopped.”

Helsingin Sanomat stated in September 2025 that 29,000 research papers were submitted to the AI industry’s AAAI conference. There are not enough sufficiently competent reviewers in the world for such a number of papers. Even if the papers are not simply “AI slop”, the threshold for producing text has become so low that no one has time to read it all and separate the wheat from the chaff.

One major question is what outsourcing writing, reading, and thinking does to learning. Roos says many of his colleagues have already gone back to exams conducted in exam rooms on paper or offline computers.

No productivity explosion after all

According to Roos, different fields should take ownership of the phenomenon that is not simply technological. In recent years, we have learned that AI is not a monolith – not good or bad, not something that belongs or does not belong to me.

“When the changes to economy and working life brought on by AI are discussed, researchers of these fields are listened to surprisingly little while paying attention to AI researchers like myself.”

Roos thinks researchers could offer a voice of reason in the discourse currently dominated by commercial parties and consultants that benefit from AI.

“You see in the papers how something used to take weeks in the past but now takes half an hour with AI. But when you look at productivity indicators, profits from working hours, or the GDP, the growth is actually very modest.”

Researchers could offer a voice of reason in the discourse currently dominated by commercial parties and consultants that benefit from AI.

This is a similar productivity paradox to the one caused by the introduction of office computers in the past. Productivity did not see an explosion of growth after all. According to Roos, productivity has increased in individual tasks as a result of AI. However, no one works full time on something whose efficiency has increased by orders of magnitude.

“Building and implementing systems, continued development, monitoring, and maintenance including licences and power bills may also cost so much that it might completely offset any benefits.”

Roos still sees “low hanging fruit” – easy opportunities to improve work – in AI use. In addition to language models and AI agents based on them, it is possible to use plenty of other AI that models data statistically.

“It is a shame the analytics-style machine learning has been cast aside so much despite having a lot of momentum over five years ago. This is more familiar to many than you might think. For example, Excel has methods for creating predictive models.”

All fields of science use statistical analysis processes. When these methods are taken further, we start to talk about machine learning.

“It does not sound very sexy but is probably a bigger deal for science than generative AI is”, Roos says.

Reviewing must be reviewed

At higher education institutions, AI has probably been discussed most from the study and review points of view. Often, the focus is on problems, such as how to prevent cheating or harmful use. University Lecturer Heidi Salmento works at the University of Turku’s Unit of University Pedagogy and, alongside her other work, researches AI use in studies. According to Salmento, students do not always use AI in a meaningful way for learning but are able to reflect on the effects of tools on their learning when asked.

“We have talented and trustworthy students who usually think and act responsibly. However, traditional types of tasks may still tempt them to use AI in a disadvantageous way for learning. Sometimes, students take shortcuts in such tasks.”

Ethical questions and the consumption of energy and natural resources are also issues considered by students. Currently, pedagogically meaningful ways to use AI methods are desired. Some teachers practise them with students. For example, one teacher in a technical field has asked their students to teach how they utilise AI. Students and teachers have been considering together what is advantageous AI use and what may hurt learning.

Salmento finds AI tools important, but scientific thinking skills should be polished alongside them. The university has had to think about teaching in a different way while questioning the traditional text-focused review process.

The university has had to think about teaching in a different way while questioning the traditional text-focused review process.

“It has been based on the assumption that text quality reflects developing thought. Text may no longer be sufficient evidence when seemingly correct text can be produced without in-depth thinking.”

Some teachers have adopted oral exams or other ways to measure learning. Text-based tasks emphasise personal, forward thinking more than before. It is crucial for the student to understand their own thinking and that of others, as well as how information is created.”

“Scientific knowledge and scientific thinking skills are emphasised, which is the heart of the academic world anyway.”

Tools change thinking

Professor Minna Ruckenstein from the University of Helsinki’s Centre for Consumer Society Research recently asked students on a course if they wanted to use AI. Some did, others completely refused. This created an interesting discussion about how students themselves see their learning.

“Tools change thinking, doing, and the doer”, Ruckenstein says.

Ruckenstein researches the human and social connections of AI and leads the REPAIR project funded by the Strategic Research Council. The project studies algorithmic systems in different aspects of society.

According to Ruckenstein, it is important to consider the effects of AI on a wider scale than simply at the tool level. The benefits are evident in many things. For example, creating surveys, analyses, transcripts, and translations is efficient. It is important to consider what part of your learning and work is important to you and cannot be outsourced or expedited.

Ruckenstein offers a reminder that learning requires some degree of difficulty.

“Thinking is meant to be slow and painful. Reading texts is sometimes hard. If simplified summaries of those texts are available, you may miss many important things.”

This time requires dialogue

An obvious tangible sign of change in work is excessive AI-generated slop. Ruckenstein says that doctoral researcher applications are coming at such a rate that opening the emails has become a chore. Another change can be observed in teaching. Ruckenstein would no longer make students write reflective papers about singular articles.

“There are not enough teaching resources to work closely with students even though this time specifically requires dialogue and developing things together.”

Ruckenstein has noted increasing cynicism among students. They wonder where they should be headed and whether or not working hard is worth it at this point. The economy is a factor. Students fear they fail to find employment or that AI takes their jobs. Another source of cynicism is the tech-oriented reality which removes people from their roots and recycles information that not connected to anything.

The REPAIR project makes efforts to repair a cynical future by diagnosing the roots of the problem from the young person’s point of view and considering possible solutions.

Ruckenstein says right now is the time to think about what we want to communicate to young adults, what kind of creator of information the university is seen as, and how work should be measured and reviewed. Conducting science is not just pushing out articles but thinking with others, in-depth understanding, and perceiving the world through writing.

“The great thing about this work is the constant discourse over the nature of information, which has been horribly misunderstood in public in recent years. The best researchers understand the limits of their own knowledge and dare change their opinions when new information emerges.”