Soumitra Dutta says agentic AI will test academic publishing

Academic publishing has been slowly crumbling for years․ Journals are overwhelmed․ There are not enough trained reviewers․ The architecture of scientific trust is at‚ or beyond‚ capacity in many fields․ Nobody has solved the problem․ And now agentic AI is about to make the problem dramatically worse or‚ depending on how institutions respond‚ dramatically better․

Soumitra Dutta‚ former dean of the Saïd Business School of the University of Oxford and cofounder of the Portulans Institute‚ in a recent post on X‚ wrote about how agentic AI can impact research. Dutta has a PhD in computer science from UC Berkeley from 1990‚ specializing in AI․ He is a co-founder of NexiVerify‚ an AI verification company that produces cryptographic receipts for AI outputs‚ and Caasaa‚ an AI-native software development firm that transforms business processes. Dutta is also the Executive Director of Academic Initiatives and Visiting Professor at SRM Group of Institutions․

Agentic science is autonomous AI systems that are able to autonomously ask questions‚ design an experiment‚ analyze the results‚ update their experiments based on these results‚ and repeat such experiments independently in an iterative fashion․

In other words‚ research that took years can now be done in weeks․ This has changed the economics of science and the politics of science․ "Trust will depend on how research is produced and acknowledged‚" Dutta says․ When a system can produce research at an industrial scale‚ we must seriously ask: how will the reader know what to trust‚ and who is responsible for what gets published?

With the existing pressures facing the peer review system‚ algorithmically generated papers that are complete but devoid of content could threaten to overwhelm the system completely․ This is the underappreciated crisis behind the opportunity Dutta describes․ In institutions‚ for the sake of speed‚ confounding the logical production of knowledge with the measure of the outcomes has been a regular occurrence․

Soumitra Dutta prescribes that good researchers should "Double down on theory and judgment - these become more valuable as production is automated".

The mechanical jobs are the easiest ones to automate․ What is not automatable‚ at least at this point‚ is the capacity to ask whether a question is worth asking - to really think through a problem intellectually before running any experiment․