Generative AI (we are not talking about machine learning research applications here) is taking over science. It is being forced into our workflows by big tech companies, and sometimes also by our employers. Not taking action implies accepting this new status quo. Why resist? 

• Ecological issues: The energy and water use by IA have a huge impact on our planet.
• Copyright issues: Most genAI are trained on proprietary texts, images, and code without explicit consent.  
• Biases: AI is trained with biased data, and hence, its responses are not neutral and exacerbate existing biases.
• Hallucinations: Not only are hallucinations common, but AI is overconfident and designed to please the user. These two aspects create the perfect situation to develop trust in wrong information. In addition, regarding code, it creates overly complex R code (even when correct), and debugging complex code often takes longer than writing it in the first place. Most dangerous, when it lacks information, it fabricates it (e.g., X-ray interpretation https://arxiv.org/abs/2603.21687).
• Reproducibility: When delegating some tasks to the genAI (e.g., automatically fixing data formats), we lose reproducibility.  
• Desincentivize learning: People learning to do a task with AI have lower levels of retention: https://arxiv.org/pdf/2506.08872v1. 
• Switch our role from creatives to producers. This is my highest worry. When we stop training our skills, we lose them. AI will create uniformity in how we write, what we cite, or how we do things. It will speed up a culture of publish or perish, forcing us to do many jobs, none of which we are proficient in. Expectations on productivity will rise, likely at the cost of quality (or innovation). If the trend continues, it might even create a huge knowledge debt by training a full generation that can produce many outputs with a professional look, but who are not able to think deeply and critically. The big problems need experts, not producers.  

At the lab, we are aware of these problems. We are also aware that AI can speed up many tasks when used properly. We are not that naïve. Hence, as a compromise, we propose to ensure AI is not our default tool. Our first option is not using AI, but to start with pen and pencil (if needed). Especially for new tasks, we devote our time to learning how to do them. This does not imply that if we get stuck, we can not ask genAI. genAI can be used for learning purposes (being aware of its impact and limitations). 

This implies taking a few actions that require some effort:

• Not to embed AI in all our workflows by default: Using a search engine that does not prompt AI by default (e.g., DuckDuckGo), not installing Copilot or the like when coding. This might seem a silly step, but humans are lazy, and adding barriers to our laziness is needed. Reading, writting and coding are an integral part of our job, and those require practice. If we stop practicing, we lose the skills. 
• We do not prioritize speed. Earning a PhD is learning how to do science. How to write code, how to write an introduction. This implies doing it yourself and not delegating your tasks. 
• When the benefits of AI for a particular task are clear, we select the right tool (e.g. Google Scholar Labs for references) and use it with care. This often implies cautious prompts that are as specific as possible, and not broad task delegation. e.g., explain to me why this fails > fix this line of code > write me code that does X. 
• Be upfront. Discuss early the expected genAI usage for a given project with all co-authors, and ensure everyone is comfortable with the plan. Disclose substantial AI use in all our outputs. If we use it, we explain it. This is transparency. Exceptions can be small grammatical edits by e.g., Grammarly, and small technical questions, but any paragraph, code, or image generated should be credited.
• We take responsibility for checking all AI outputs. It is our duty to fact-check AI work, from ensuring the code is functional, the references exist and say what the AI claims, or the suggested text corrections do not change the intended meaning. In a nutshell, we take full responsabiliyty of the tools we use. 

Thanks to Tim Poisot for triggering this discussion by sharing his lab rules: https://poisotlab.io/generative-ai/