Skip to main content

16  AI in a scientific workflow

The growth of AI-based systems during recent years is changing science. Evidence supporting this conclusion can be found in your everyday science-related activities, such as a Google search supplying an AI summary of your question, your text-editor suggesting edits to your manuscript, or when you are confronted with a “conclusive” remark based on Chat GPT output during a project meeting. Science is also changing on a larger scale. The creation of agent-based scientific systems, where AI agents is the central component of the scientific workflow, have been shown to lead to novel discoveries in e.g., biomedicine (Ghareeb et al. 2026; Gottweis et al. 2026). Science is only one field affected by the AI boom1, an insight that suggest that we will likely see more, not less, AI in years to come.

Besides discussions in the scientific community regarding potential benefits and disadvantages associated with AI in science, we also need to address AI as part of our own scientific habits. Fatalism can be avoided here as we can reason about using AI just as any other tool used in scientific work. The researcher and writer Cal Newport argues that AI tools should be regarded as any other tool, not as a magic machine2. Regarding AI tools as any set of potential tools we can more critically evaluate their use. Newport writes about selecting, or rather avoiding, network tools such as social media if one wants to develop as a knowledge worker (Newport 2016). Although Newports discussion does not concern AI tools, we can use the same type of analysis applying it to AI tools in science. Newport suggests “the craftsman approach to tools selection” (Newport 2016, 191) which states that we should use tools that provide benefits, substantially greater than any negative impact, supporting “core factors determining success and happiness in your professional and personal life” (Newport 2016, 191). Using a deeper analysis is in contrast to adopting an “any-benefit” approach to tool selection (Newport 2016, 186). This strategy push us to adopt any tool that give some potential benefit, without balancing benefits to negative consequences.

Before analyzing AI in science as a tool we need to establish that AI in science can be a lot of different things. We could use the AI chat interface to discuss with a large language model. We could also use the same model as a coding agent assisting in writing code to construct scientific analyses. Many more specialized systems exists (just Google, e.g., AI science tool), however for our discussion, let’s say that we have access to a large language model-type system in our terminal, capable of editing the same files that we work on in a given repository. This AI tool, a coding agent, could be Claude Code, or an open source alternative, such as Open Code. This setup might be suitable for data analysis, literature searches or building software, all being key components in modern scientific research.

We also need to establish an overall goal for our workflow (for which we are considering AI tools). Let’s say the overall goal with our scientific is to establish new knowledge. However, we want to be able to say that we have produced knowledge about a phenomena that can be communicated to others as well as something that have supported our own learning process. We are therefore interested in developing and understanding skills and strategies used to contribute to our collective knowledge. In formulating our goal we cannot separate our personal learning experience from providing new results to the scientific community. Now, let us analyse the use of AI coding agents as part of a scientific project.

A coding agent, set up using e.g., Claude Code, could provide a scientific environment close to what is described by Gottweis et al., (2026) where in the scientist(s) provide instructions and goals and agents act on those to autonomously iterate over the scientific process (question → data → analysis → conclusion → question → …) asking for human input when new experiments needs to be performed to provide more data. As such, the AI tool can be used to provide new knowledge to the world.

Does this support our goal? The short answer is NO. A (semi-) autonomous agentic workflow does not support personal learning, which was alo part of our goal. Although we could instruct the coding agent to provide evidence for its conclusions, such as reproducible code, mathematical proofs, detailed statistical analyses, and so on, the AI system works on another time scale compared to human learning. Reviewing conclusions and evidence to support personal learning about the process producing the results would likely demand a considerable investment in time and commitment. But why bother? We already have the new knowledge.

The consequences of AI use in learning has started to accumulate, and they point towards a problematic relationship between learning and AI. Observational evidence suggest that AI usage increase homework scores but reduces exam scores (Strömberg et al. 2026). Similarly, in experiments, using an AI chatbot hurts performance in exams although performance during practice increase (Bastani et al. 2025). Recently, Liu et al. (2026) confirmed that unassisted performance is negatively affected by AI usage during learning, however, they also point to AI usage negatively affecting persistance. Study participants that used AI in learning new concepts were more likely to give up during unassisted tests (Liu et al. 2026). The authors extrapolate their findings from short-term interventions to long-term every day usage and warn that AI usage will hurt our ability to learn as mental offloading shifts our understanding of how to learn.

Using the above points we can conclude that broadly adopting AI for mental offloading during scientific may hurt our overall goal with a scientific workflow. It could be argued that any benefit (assist scientific workflow) does not outweigh the negative impact on personal skill-, and knowledge development.

16.1 Critically adopting AI in science with purpose

If we aim to adopt AI in science we need to do so with intent and critical reflection of our overall goals, as shown above. E.g., if development of personal knowledge is part of your overall goal we need to be able to grind to solve problems and not mentally offload to a AI agent. However, not all aspects of a scientific workflow may be part of the learning target. If we want to focus on solving problems that matters (e.g., developing understanding) we could use an AI agent to solve problems that interfere with attacking more important issues. In such a scenario, the list below could be a set of problems for which a coding agent is used:

  • Bookkeeping: Sort and keep track of files used to support a data analysis.
  • Solve isolated problems: Suggest improvements to avoid bottlenecks in coding, re-factor code to improve computations, find bugs, solve issues in version control conflicts, etc.
  • Make written code less complex: Simplify code to improve readability.
  • Find inconsistencies: Review connections between software and conclusions and find inconsistencies.

In another project, where the full workflow is the intended target for personal learning, using an AI agent to address issues with version control or code readability might hurt your goal.

16.1.1 Promoting yourself to an AI agent supervisor

In the Claude Code learning resources we can read that Anthropic recommends researchers to verify statistical results, review code before publication and save conversation logs for records (see Research-Specific Considerations (codewithclaude.net accessed 2026-11-09)). This highlights the complexity of offloading as part of the scientific workflow. We need to maintain a certain level of proficiency to supervise the AI system. If you, instead of having the role of a researcher promote yourself to supervisor of an agent you are also expect to be able to find inconsistencies and coding errors produced by the agent. As the project grow, this may become a large part of your research activities. In my experience, maintaining a human-readable and well organized and documented repository will help maintaining control over any agent output. To accomplish this we should adopt general habits for data analysis that are also part of a non-AI-infused workflow, such as:

  • Keep a structured repository for code and prose.
  • Use comments explaining why, when, how, etc.
  • Use human-level bookkeeping, keep a change-log, list relevant files and describe them.
  • Use version control. Any breaking changes can be reverted using a well maintained version control repository.
  • Revisit code and improve to make it better and more readable.

16.2 What we have not covered…

When considering adopting an AI-centric workflow, one should also consider:

  • Data privacy. Your data might not be suitable for sharing with a coding agent. Data privacy regulations may force you to limit any AI use unless you use suitable solutions. For Norwegian researchers, coding agents are soon part of the TSD suite making it possible to use a coding agent hosted on secure servers 3

  • Environmental impact. The use of AI is predicted to increase, the market expect even more as planned data centers are expected to use electricity proportional to several nuclear power plants in Sweden4. We might worry that our AI usage has consequences on the environment.

  • Societal impact. Above we discussed learning, but AI usage will likely also affect other areas of the human experience. Are we willing to contribute to such changes?


  1. AI boom on wikipedia.↩︎

  2. See the e.g., the blog post “On AI Coding and Its Discontents”.↩︎

  3. At the time of writing this, coding agents are not available through TSD outside the University of Oslo. See here↩︎

  4. Listen to a report about Swedish data centers here (in Swedish).↩︎