Research at Finland’s Åbo Akademi University has found that newer artificial intelligence skills are appearing alongside established programming and data skills in the occupations examined. Experiments also found uneven benefits from AI-assisted work, with human judgment affecting the quality of the result.
Nurlan Musazade’s doctoral research combines analysis of recruitment advertisements with experiments examining how people solve problems using generative AI, software that produces text, code and other material in response to instructions.
The dissertation, defended on September 25, examines data, analytics and AI-related occupations. Its evidence concerns advertised skills and performance on selected tasks. It does not measure whether skills are changing faster than jobs are being eliminated.
For employers, the research raises a hiring question: what should employees know themselves when software can produce part of their work? The answer affects recruitment, training and how companies assess the output their employees deliver.
Employers want AI connected to their operations
In the recruitment analysis, jobs with AI in their titles were more closely associated with newer large language model tools and skills involving workflows. Large language models, or LLMs, are systems trained on large amounts of data that can generate and process language.
The university’s September 29 announcement describes greater employer interest in people who can apply AI tools, connect them to other systems and adapt them to organizational requirements.
Applying a model inside a business involves more than obtaining an answer from a chatbot. Information has to reach the tool, its output has to fit the task, and somebody has to decide whether that output is acceptable.
Those requirements can create work involving software connections, data preparation and checking results. A company using an existing model still needs people who understand its own records, applications and operating procedures.
Established technical skills remain in the mix
Programming, machine learning and data-storage tools remained central across the data-related occupations examined, although demand for some technologies declined within particular professions. Machine learning involves training computer systems to identify patterns in data.
Python, a programming language widely used for data analysis and AI work, appeared repeatedly in combinations of skills. Musazade identifies it as a central connection between different technical requirements.
The hiring pattern points toward employees adding capabilities to an existing foundation. Knowing a newer AI tool may help a candidate, while the ability to work with data and programming remains relevant to the advertised job.
Recruitment advertisements have limits as evidence. They describe what employers ask for, which may differ from the work eventually assigned. Their geographic and occupational coverage also prevents treating the results as a description of every labor market.
As we reported in our earlier coverage of skills shortages and graduate unemployment, a qualification, a vacancy and an employer’s assessment of job readiness measure different things. Listing more requirements in an advertisement does not establish how well an applicant can perform the work.
Access to ChatGPT produced uneven results
A related peer-reviewed study published in Algorithms in October 2025, by Musazade, József Mezei and Xiaolu Wang, examined management-consulting-style problems involving financial analysis, strategy and data requirements.
The experiment recruited 16 master’s and doctoral students and recent graduates from business administration and information technology disciplines at a university in Finland. Participants were randomly assigned to work with or without access to ChatGPT using GPT-4 and completed one of two consulting cases.
Researchers assessed analytical thinking, creative thinking and systems thinking. Systems thinking concerns how different parts of a problem interact, including effects that may be missed when each part is examined separately.
AI-assisted participants performed better on some measures, including aspects of logical reasoning and problem definition. The pattern varied between tasks. An advantage on one measure in one case did not necessarily appear in the other.
Screen recordings allowed the researchers to examine how participants used ChatGPT. More collaborative users combined its responses with their own reasoning or calculations. Others copied substantial amounts of its output.
Collaborative behavior was associated with stronger performance in several assessed skills. Some participants who copied responses also performed well on individual measures, making a simple division between successful collaborators and unsuccessful copyists inaccurate.
With only 16 recruits, and one non-using participant excluded from the initial group calculations, the experiment is exploratory. It cannot establish a population-wide effect on intelligence, workplace productivity or long-term skill development. The comparisons between interaction styles were observations within the experiment, not a separate randomized test of collaboration.
Time saved still leaves work to check
The researchers reported that GPT users finished their cases around 15 to 20 minutes earlier on average. Among those users, completion time had no strong relationship with average performance.
Faster delivery and better work must be assessed separately. A report can arrive sooner while containing an incorrect calculation, a misunderstood instruction or a conclusion that its evidence does not justify. The researchers identified errors involving numbers, case information and task requirements.
For businesses, the implication is to assess the completed work as well as the time spent producing it. Savings from quicker drafting may be reduced if checking and correcting the output takes longer elsewhere in the organization.
The experiment used GPT-4 and was published before the dissertation’s September 2026 defense. Its results describe those participants, tools and cases, rather than the performance of every model available today.
Training has to include evaluation
Musazade told the university: “It is about how effectively people can collaborate with AI.”
That places attention on how employees frame a problem, supply information and evaluate the answer. Training limited to entering prompts would leave those parts of the job unaddressed.
An employee reviewing an AI-generated financial recommendation, for example, still needs enough knowledge to examine its assumptions and arithmetic. Training can ask employees to explain how they checked a result and which parts required correction.
Employers and educators face different timescales. A business can introduce a new tool while employees are still learning it; a course must prepare students for work beyond the life of any single application. The research gives them grounds to teach AI use alongside programming, subject knowledge and the evaluation of evidence.
A remaining question is whether repeated AI assistance develops those abilities or allows users to avoid practicing them. A short consulting experiment cannot settle that. Answering it requires following people over time and examining the work they can perform when assistance is unavailable.
