What does research say about the effects of AI on software work?

 

In recent years, a great deal of research has been conducted on the effects of AI on work, and the topic has attracted broad interest in companies as well. However, the discussion has strongly focused on technology and productivity. Less attention has been paid to how AI changes the experience of work, learning, responsibility, and coping. Research results always look at past data, so caution is needed in their interpretation, especially when the subject of examination is a very rapidly developing technology.

 

In an experimental study by Peng and colleagues (2023), developers solved a clearly defined programming task either with or without the help of GitHub Copilot. Those who used the assistant completed the task faster, which for the first time showed in a controlled setting that generative AI can increase short-term efficiency.

 

Later field experiments have, however, complicated this picture. Cui, Noy, and Brynjolfsson et al. (2025) conducted several randomized experiments in real company environments. These showed average productivity growth, but also large variation between individuals and tasks. For some, AI was significant support; for others, its effect remained small. This suggests that the benefits of AI are not universal, but depend on the nature of the work, the context, and existing practices.

 

The contradiction between perceived and measured efficiency

 

An interesting perspective on the benefits brought by AI is offered by a study conducted by METR Research. METR (Model Evaluation and Threat Research) is an independent research organization that focuses on evaluating the real impacts and limitations of AI systems. Its approach differs from many earlier studies in that it seeks to model as realistic and demanding work situations as possible.

 

In METR’s 2025 study, experienced open-source developers worked in a codebase familiar to them and solved real development tasks either with or without an AI assistant. According to the results, developers who used the AI assistant completed their tasks on average more slowly. Despite this, they subjectively experienced the work as progressing faster and evaluated the AI as useful.

 

This contradiction is an interesting finding. METR’s analysis shows that a significant amount of time was spent reviewing suggestions, modifying them, and fixing errors. The work may have felt smoother, even though actual time use increased.

 

Learning and meaningfulness of work in the age of AI

 

In software development, expertise is often built slowly. Mistakes, failures, and reflecting on solutions with the team are central sources of learning. Research suggests that AI can shorten this process, but at the same time it may bypass some of the stages where deep understanding emerges. When using AI, an illusion may arise that something has been learned because the AI has processed the matter on your behalf.

 

Observational studies of developers’ use of ChatGPT show that AI is particularly used for ideation and outlining alternatives. This can support thinking, but at the same time a risk of excessive trust emerges. When a suggestion seems credible, the assumptions behind it may not be examined critically enough. This changes the nature of the work and may weaken the experience of control and meaningfulness.

 

Decision-making and retaining responsibility with humans

 

Although AI participates in producing solutions, responsibility does not transfer to the machine. In principle, this is clear to many, but in practice the boundaries of responsibility may become blurred. Research highlights the risk of automation bias, where machine suggestions are followed more readily than solutions made by humans, especially under time pressure.

 

Human-centered use of AI requires that developers have time and permission to evaluate suggestions critically. Decision-making must not become a formal approval, but must remain genuinely human activity. This is both an ethical and a practical issue, because the consequences of errors are real. Based on Stack Overflow’s 2025 survey, developers approach AI-generated code critically for precisely this reason. AI-generated answers were often perceived as “almost correct,” making human review and correction an essential part of the development process.

 

Psychological resilience and safety in the age of AI

 

As AI becomes part of daily software work, workload does not necessarily decrease, but may change its form. Psychological resilience in such an environment means the ability to operate with continuous cognitive load and unfinished solutions without well-being or quality of work deteriorating. Research suggests that high mental load weakens concentration and creativity, unless recovery and reasonable expectations are taken into account in work structures. If the nature of work changes significantly, sufficient time should be given to learn new ways of working and to create shared rules within the work community.

 

Psychological safety in a work community means a work climate where bringing up uncertainty, mistakes, and unfinished ideas is allowed. In AI-assisted work, this is especially important because machine suggestions can be credible but incorrect. If team members do not dare to question or ask for help, the risk of accepting poor solutions increases.

 

Research shows that in psychologically safe teams, learning, detecting errors, and joint evaluation work better. This supports both the quality of work and individual coping. Human-centered use of AI requires an environment where people can use their judgment openly and without social pressure.

 

The responsibility of leadership and organizations

 

Large datasets such as the DORA reports show that the use of AI alone does not predict better performance. What is decisive is the processes and metrics to which the technology is connected. It is the responsibility of leaders to ensure that assumed efficiency does not turn into unnoticed workload.

 

Studies like METR’s show that subjective experience and measured productivity can diverge from each other. If targets are tightened based on an assumed efficiency leap, the risk is increased cognitive load and weakened coping. At the same time, learning may suffer if AI replaces thinking instead of supporting it. It is important for organizations to understand what kind of expertise is critical to their operations and to ensure that the development of this expertise does not erode in AI-assisted work.

 

A human-centered direction means conscious leadership. It means protecting learning, realistic expectations, and clear rules for the distribution of responsibility. When these matters are taken seriously, AI can function as a work partner that strengthens the human role instead of narrowing it.

 

Finally

 

Research shows that the future of software work is not automatic. AI can increase efficiency, but it can also slow down, burden, and change the meaning of work in ways that are not always noticed in time. The adoption of AI, like any other technology, should be seen as a human and organizational issue, not only a technical solution. When humans remain at the center of decision-making, learning, and responsibility, the benefits of AI can be realized without compromising the meaningfulness and sustainability of work.

Connect with Katariina in LinkedIn.

References

• Becker, J., Ringer, T., Sandler, A., Yang, D., Kalliamvakou, E., & McMillan, C. (2025). Measuring the Impact of Early-2025 AI Tools on Experienced Open-Source Developer Productivity. METR Research. 
• Cui, K. Z., Noy, S., Zhang, W., & Brynjolfsson, E. (2025). The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. MIT Sloan School of Management, Working Paper. 
• Edmondson, A. C. (2023). Psychological Safety Comes of Age: Observed Themes in Workplace Research. Annual Review of Organizational Psychology and Organizational Behavior. 
• Kahn, W. A. (1990). Psychological Conditions of Personal Engagement and Disengagement at Work. Academy of Management Journal. 
• Määttä, S. (2024). Cognitive aspects of generative artificial intelligence in software engineering: A scoping review. Thesis. University of Oulu. OuluRepo.
• Peng, S., Kalliamvakou, E., Cihon, P., Demirer, M., & Fast, E. (2023). The Impact of AI on Developer Productivity: Evidence from GitHub Copilot. 
• Stack Overflow. (2025). Developer Survey 2025. https://survey.stackoverflow.co/2025/
• Vaithilingam, P., Zhang, T., Glassman, E. L., & Bansal, G. (2024). An Observational Study of How Developers Use ChatGPT in Software Engineering Tasks. Proceedings of the ACM on Software Engineering, Volume 1, Issue FSE.
• Tiiliharju, H.–M. (2023). Psykologinen turvallisuus ICT-alalla. Thesis. OAMK. Theseus.  
• Kimonen, K. (2021). Psykologinen turvallisuus tietotyön menestystekijänä. Thesis. LUT-University. LUTPub.