Can’t Retract = Can’t Govern

When I asked the question, “Can you retract from an LLM?” late last year, I incidentally raised an entire set of questions we need to consider about whether we adequately interrogated the technology to ensure it could adhere to scholarly norms.

• If you want a reminder of how important norms are, look around you, and also listen to our interview with Skylar Hughes.

Retraction is the most prominent of our norms these days because it’s been sensationalized by a variety of forces — from persistent reporting by Retraction Watch to mass retraction events. But there are other, more or less common norms used to govern the scientific record which are also not clearly supported by LLMs:

• Corrections and corrigenda
• Expressions of Concern
• Citation
• Addenda
• Withdrawal

Why would we adopt an information processing technology that doesn’t let us do our work? It’s a question with an answer I’m not sure we’d be proud to utter.

Let’s walk the perimeter to see whether LLMs are actually compatible with what we do:

• Corrections and corrigenda. A correction notice or corrigenda would just be another set of tokens in the system, competing with the uncorrected version.
  • Are we even writing corrections in a manner that would have any effect?
    • “In Fig. 4a of this Analysis, owing to an error during the production process, the year in the header of the right column was ‘2016’ rather than ‘2010’. In addition, in the HTML version of the Analysis, Table 1 was formatted incorrectly. These errors have been corrected online.”
  • VERDICT: Likely Incompatible
• Expressions of Concern. These occur at the article level, but there are no articles surfaced by an LLM except via citation and concatenation, and any citations are often nothing more than brand freckles on new text agglomerations. Written with overweening confidence, the products of LLMs repel interrogation and engagement with the literature, burying articles. Expressions of Concern have no place here.
  • VERDICT: Incompatible
• Addenda. Significant, peer-reviewed additions to the interpretation of the original publication, addenda can affect the scientific accuracy of published information, the reputation of the authors, or the reputation of the journal. Addenda do not contradict the original publication but are added if the authors inadvertently omitted significant information. Knowing they were added later is sometimes important for the community involved.
  • VERDICT: Likely Incompatible
• Citation. There are standards for citing LLMs as sources, but when you get beyond the mechanical constructions of these, the epistemic unreality hits you. If I cite a website, chances are it will be there later, or at least will have been scooped up in the Wayback Machine. Chances are high I can find it in some vestige of its original, cited form. Not so with an LLM, which never presents anything in exactly the same way, from moment to moment, much less months or years later.
  • VERDICT: Incompatible

Why would we adopt technologies so out of alignment with our norms and needs? It thwarts sleuths, editors, authors, publishers, and academics at multiple levels, confirming that LLMs and AI are about power and not knowledge.

• Also, technologies that defend the status quo and destroy institutions?

Why are we playing with these things at all? They make the scientific literature ungovernable by humans, and subvert norms established over decades and used to signal important caveats, point to supporting documents, and enable authors and publishers to manage claims after publication.

What are we getting in exchange for our institution being destroyed?