Scientific oversight as a requirement for responsible AI

Lead author Claudi Bockting, professor of Clinical Psychology of Psychiatry at Amsterdam UMC and co-director of the Centre for Urban Mental Health believes that, ‘AI tools could flood the internet with misinformation and ‘deep fakes’ that can be indistinguishable from real individuals. This could over time, erode trust between people, in politicians, institutions, and science. Independent scientists must take the lead in testing, proving, and improving the safety and security of generative AI. However, most scientists don’t have access to the facilities or public funding to develop or evaluate generative AI tools.’

• Accountability: Advocating for a human-augmented approach, the consortium believes that while generative AI can assist in low-risk tasks, essential endeavors like scientific manuscript preparation or peer reviews should retain human oversight.
  
  
• Transparency: Clear disclosure of generative AI use is imperative. This will allow the broader scientific community to assess the implications of generative AI on research quality and decision-making. Furthermore, the consortium urges AI tool developers to be transparent about their methodologies, enabling comprehensive evaluations.
  
  
• Independent Oversight: Given the vast financial implications of the generative AI sector, relying solely on self-regulation is not feasible. External, independent objective audits are crucial to ensure ethical and high-quality use of AI tools. The proposed scientific body must have sufficient computing power to run full-scale models, and enough information on source cases, to judge how AI-tools were trained, even before they are released. Effective guidelines will require international funding and broad legal endorsement and will only work in collaboration with tech industry leaders while at the same time safeguarding its independence. The authors underscore the urgent need for their proposed scientific body, which can also address any emergent or unresolved issues in the domain.