Current views on generative AI

This post contains my current views on generative artificial intelligence, and Large Language Models in particular. The context is mostly academia, which is about research and teaching.

Personal context

Generative AI is slowly creeping into my professional workflow, not because I am using it myself (I don't, although I guess that I will, at some point), but because everyone around me is.

My students use ChatGPT and other tools like, I believe, NotebookLM and Perplexity Comet. My RSS news feed (that's how old I am) recently had an article on Claude, and I use use Google applications, so I keep getting passively-aggressively asked to use Gemini, which I might do one day via Scholar Labs.

My workplace, which is a university, has taken a very basic stance on generative AI: unless stated otherwise, students are to follow LSE Position 1 (no use of generative AI in graded work), which I suppose goes both ways (no use of generative AI in grading, either).

I do not know of any equivalent position on generative AI in research. It seems like everyone wants to discuss the topic and play around with whatever is available for free online, but no one wants to make hard decisions about it yet, possibly due to upcoming EU-level regulations.

Risks for teaching and learning

From a teaching perspective, generative AI is only useful to me if it helps students going through the following process:

  1. Learn
  2. Draft
  3. Revise
  4. Submit
  5. Defend

Part of what I teach is code, and code is the topic of this blog. As it happens, generative AI is already very good with code, and I am confident that it can be put to good use to go through Steps 1--3 of the process above.

There are, however, at least four reasons why I am currently taking ‘LSE Position 1’ on using generative AI in graded work that relies on code:

  1. Many students are using AI to bypass the learning process, rather than enhance it. This creates security risks, and violates academic ethics in the same way that hiring an external party would. This comes on top of other breaches of students ethics, such as plagiarism.
  2. The two issues mentioned in the previous point cannot be defended against at my level, at least not with my current resources. I can spot security risks, but I cannot reliably detect AI-generated code, which is neither watermarked or scannable through anti-plagiarism tools.
  3. The software that I use in class is mostly open-source, and reproducibility is part of the core principles that I teach in class. As far as I understand, and unless proven otherwise, the kind of generative AI technology used by my students does not enforce these principles.
  4. To make things worse, most generative AI also violates intellectual property, rather than reconfigure it around the ‘copyleft’ and ‘creative commons’ principles that many of us have spent years defending and advocating within fields such as academic publishing.

I have not been exposed to any argument that makes any attempt at solving the ethical, logistical, moral and eventually legal issues that I have outlined above. Until I do, I will treat generative AI as a form of doping, and will keep banning it.

The analogy above with doping is not an innocent one. There is, in my view, a very real rhetorical arc that goes from generative AI to the Enhanced Games. Higher education does not approve of students taking Adderall, and neither do I.

Risks for scientific research

From a research perspective, generative AI is only useful to me if it helps me going through the following process:

  1. Compile existing evidence
  2. Collect meaningful data
  3. Produce meaningful measures
  4. Formulate correct interpretations
  5. Enhance existing knowledge

There is no doubt that generative AI can help with every step above, especially perhaps at the level of data collection and, in the case of ‘big data’ or whatever people call it today, classification. I am also very interested in what it can contribute with regards to compiling scientific studies, in the same way that it is already helping with mathematical problems.

The risks that I have heard about so far when it comes to generative AI and social science research (which is what I do) are the following:

  1. Generative AI can poison the evidence base (Bail 2024) through the mass production of low-quality academic output, or by compromising data such as online surveys (Westwood 2025, Westwood and Frederick 2026). This is already happening.
  2. Generative AI does not yet produce reliable data annotations for the kind of data that I am interested in (Yang et al. 2025), and even if its coding reliability improves, it will require additional effort to mitigate related issues (Baumann et al. 2025).
  3. Relatedly, generative AI cannot improve organically if it maintains its human bias towards evidence produced in the Global North (Ramirez-Ruiz and Senninger 2025), mostly by ‘WEIRD’ individuals (Atari et al. 2023). This will be hard and slow to solve.
  4. Last but not least, generative AI will be used to erode scientific authority at the profit of those who are interested in attacking the contribution that scientific (and higher education) institutions make to society. This is of course far from a trivial issue.

The issues listed are all real, hard to solve, and are controversial insofar as some people have a vested interest in seeing them not addressed, at least not in the short term.

None of these issues will stop me from installing and trying out ellmer one day. However, I do expect this to happen within a scientific environment that will have acknowledged each issue in one way or another, and formulated guidelines to address them.

Are we there yet?

This post was inspired by the /ai ‘manifesto’, which I discovered thanks to Andrew Heiss. I obtained some of the cited references through Jessica Hullman's ‘New course on generative AI for behavioral science’ blog post.

  • First published on March 14th, 2026