Radboud faculty doc sets 'do not use' default for GenAI
TL;DR
- A Radboud University faculty guidelines document sets the default at 'do not use' generative AI unless five Dutch research-integrity virtues can all be met.
- Acceptable exceptions are narrow: cleaning up OCR errors in historical sources, generating synthetic data for testing or privacy, and augmenting datasets with metadata.
- If researchers do use it, they must pick 'maximally open, local, non-proprietary' models, never send sensitive data via proprietary APIs, and log prompts and versions.
A Radboud University faculty document tells researchers not to use ChatGPT or Microsoft Copilot unless they can do so honestly, scrupulously, transparently, independently, and responsibly. Those are the five virtues of the Dutch Code for Research Integrity, and the document, written by Mark Dingemanse for the Centre for Language Studies and Radboud Institute for Culture and History, calls this the first principle for any decision to use generative AI.
Its argument is that most current generative AI "undermines these virtues by design." The list of failures is compact: "Input data is legally questionable; output reproduces biases and erases authorship; fine-tuning involves exploitation; access is gated; versioning is opaque; and use taxes the environment."
The productivity pitch gets dismissed in a line. "Productivity does not equal insight, and when kept unchecked it may hinder innovation and creativity."
The exceptions are narrow. Cleaning up OCR errors in scanned historical sources. Generating synthetic data for testing analyses or privacy protection. Augmenting datasets with summaries or metadata. Anything beyond that, and the researcher inherits a checklist: choose "maximally open, local, non-proprietary models"; never send privacy-sensitive data through proprietary APIs; log prompts, model versions and dates; disclose every use.
On academic writing itself, Dingemanse is blunt. Writing "is not merely an output format but a means of thinking, crediting, arguing, and structuring thoughts. Hollowing out these skills carries foundational risks." He also names OpenAI for "exploiting Kenyan workers" in the labour that tunes chatbot output, and notes that passing off generated output as human-authored in scientific writing "counts as misconduct, and failure to disclose has already led to retractions."
This is version 1, dated April 2024, with input from Andreas Liesenfeld, Iris van Rooij, Olivia Guest, Tamar Sharon and other listed colleagues, and is promised to be revised as the landscape shifts.
Shared on Bluesky by 1 AI expert
-
Not uni-level & not full refusal but my faculty has officially endorsed a policy that calls for more slow science, sets out the many ways in which GenAI erodes values of research integrity, and advises "to not use it, un…
View on Bluesky →
Originally reported by osf.io
Read the original article →Original headline: OSF