Academics Need to Wake Up on AI, Part III

  • Author/Source: Alexander Kustov (University of Notre Dame), Popular by Design (Substack)

  • Original: https://www.popularbydesign.org/p/academics-need-to-wake-up-on-ai-part-4c6

  • Key Ideas

  • Most "AI slop" criticism is misdirected --- the academic literature has produced human slop at scale for decades. The replication crisis, p-hacking, HARKing, citation-without-reading, and ghostwritten senior-author papers all predate AI.
  • The "stochastic parrot" critique is empirically false for modern multi-modal models, and ironically describes much of human academic discourse better than it describes today's LLMs.
  • The "one-drop rule" framing of AI contamination (any trace of AI = morally tainted output) is incoherent: the same logic would condemn spell-check, Google search, and research assistants. Disclosure-everywhere norms are game-theoretically unstable because honest users bear costs while secret users free-ride.
  • Not using current AI tools in research and writing is malpractice. The rigor lives in the thinking and verification, not in keystroke provenance.
  • LLMs do produce new knowledge in the practical sense (e.g., Claude Mythos surfacing thousands of previously unknown vulnerabilities, including a 27-year-old bug). Most social science theory is already recombination across domains, which is structurally identical to what LLMs do.
  • Mental models of "the AI user" are stuck in 2023: the typical case is no longer a student cheating with GPT-3.5 but a researcher running iterative agentic workflows.
  • Disclosure is owed only when non-disclosure misleads the audience about what they paid for. Live concerts and personal-voice memoirs deserve disclosure; research and journalism are about correctness and accountability, not provenance.
  • Teaching and research require different policies: Kustov plans to ban devices in his substantive classes (so students build foundations) while using AI extensively in his own research.

  • The pronoun "I" still carries an implicit promise: factual claims don't care who typed them, but personal conviction does. The conviction must be the author's, even if the words were directed rather than typed.

  • Summary

Part III closes Kustov's trilogy by moving from diagnosis to prescription. The proximate trigger is the 2026 ISA Annual Convention in Columbus, where Kustov sat through tenured-faculty presentations he describes as below C-level work --- and notes the irony that he was receiving death threats over his AI series at exactly that moment. From there he extends ten more theses (numbered 21--30 in the trilogy). The core moves: redefining "slop" to include the human variant academia has long tolerated; rejecting purity-based AI bans as superstition dressed as ethics; reframing LLMs as cultural technologies (per Gopnik et al.) that lower the cost of doing the kind of large, data-intensive work previously available only to Chetty-scale labs.

The piece is also a structural experiment. Part I was AI-written with no human editing. Part II was 100% human voice. Part III was iteratively co-written with Claude Code from Kustov's notes and conversations with peers --- the model he's actually advocating. He pledges the opposite of the no-AI pledges circulating in academia: he will use the best LLMs and best human collaborators available so that anything bearing his name reflects his best judgment.

The most actionable arguments for working researchers: (1) automated reproducibility checks at journal submission are cheap and overdue; (2) the bar for what gets published should rise because the bar for what can be produced has already risen; (3) "I'll be quiet about my AI use" is a vote for hypocrisy as professional norm.

  • Relevance to Economics Research

Direct. The piece names economists explicitly --- Chetty, Acemoglu, Andrews, Cunningham, Panjwani, Oster --- and frames AI as a leveling force for junior scholars who can now attempt projects previously gated by team size. The argument that Codex-equipped junior researchers can attempt Chetty-scale questions speaks to the empirical-asset-pricing-with-an-agent shift this wiki has been tracking. The pessimism about AI in teaching (vs. optimism for research) is a sharper position than Part II took and aligns with Svoronos's argument that subject expertise and AI skills should be taught as distinct units. Kustov's claim that automated reproducibility checking is technically feasible today maps onto the Kohler et al. agentic-reproduction paper also in this batch.