Beyond Abstract Ethics
What if the dozens of ethical AI manifestos released by tech giants and governments are actually hindering the scientists they are meant to guide? Since 2015, we have seen over 100 ethical initiatives, yet most rely on abstract goals like "justice" that offer little practical help to a researcher in the lab.
The Core Problem: "Triple-Too"
A new study diagnoses a "Triple-Too" problem plaguing current ethical initiatives:
- Too many competing initiatives.
- Too much philosophical abstraction.
- Too much focus on distant, catastrophic risks instead of everyday utility.
A Practical Framework: User-Centered Realism
Instead of high-level philosophy, the research proposes a "User-Centered Realism" framework. It is built to move Generative AI from theory into practice with five concrete pillars.
The Five Pillars of the Framework:
- Understanding of the AI system's capabilities and limits.
- Respect for Rights (privacy, consent, intellectual property).
- Integrity of the scientific process and data.
- Demonstrable Utility where AI offers a clear advantage.
- Transparency in methods and AI use.
Why This Shift is Critical
This practical shift is essential because for the modern researcher, opting out of AI may now constitute an ethical failure.
The Risk of Inaction
Ignoring AI tools risks suppressing scientific productivity and failing to equip the next generation with essential expertise. The ethical imperative is to use these powerful tools responsibly, not to avoid them.
The Standard for Use
The goal is not a perfect algorithm, but ensuring AI is only deployed when it is superior to human-only alternatives. An example is using an LLM to analyze massive text blocks more efficiently than a team of human researchers.
From "Vibes" to Technical Rigor
The framework targets specific, technical failure points in the AI lifecycle where bias and error creep in.
Catastrophic Forgetting
This occurs when a model loses previously learned knowledge during new training, compromising its reliability and prior learning.
Paradigm Lock-in
Here, an AI system simply parrots the past biases present in its training data. This can effectively chill future scientific innovation by reinforcing outdated paradigms.
Combating Key Challenges
The study advocates for concrete technical solutions to well-known problems like inaccuracy and attribution.
Fighting "Hallucinations"
To combat hallucinations—where models act as "stochastic parrots" generating plausible but false information—the study advocates for techniques like Retrieval-Augmented Generation (RAG) and Federated Learning.
The Murky Line of Credit
The study highlights the new ethical challenge of attribution. For instance, author Rie Kudan admitted approximately 5% of her Akutagawa Prize-winning work was ChatGPT-generated, sparking a global debate on where assistance ends and plagiarism begins.
The Steep Path to an Ethical Lab
Implementing this rigorous framework faces significant practical hurdles.
Rapid Obsolescence
The lightning-fast iteration of Large Language Models (LLMs) means any ethical checklist or documentation standard faces rapid obsolescence, struggling to keep pace with the technology.
The Documentation Burden
Maintaining the "formidable task" of logging every prompt, model seed, and iteration could slow down the very research these tools are designed to accelerate, creating a tension between rigor and progress.
The Practitioner's Burden
Ultimately, the study argues that scientists can no longer wait for a universal rulebook. As policies from major journals like Nature and Science diverge, the burden of "distributed agency" falls on the individual practitioner.
The researcher remains the ultimate truth-seeker, responsible for navigating a landscape where the absolute criteria for "acceptable use" remain a subjective gray area.
Reference: Lin, Z. (2025). Beyond principlism: Practical strategies for ethical AI use in research practices. AI and Ethics, 5, 2719–2731. DOI: 10.1007/s43681-024-00585-5