LLMs are accelerating academic overproduction, OECD expert warns
Artificial intelligence is advancing research, but it's also pouring fuel on a long-standing problem: too many papers and too many grant applications. Alistair Nolan, who analyses AI in science at the OECD, warns that large language models (LLMs) make it easier to publish more, not necessarily better. If incentives stay the same, AI will amplify quantity over quality and strain the entire system.
He points to familiar behaviors-salami slicing, incremental studies, and a publish-to-progress culture-that already produce an unmanageable volume. Give researchers a tool that speeds up writing, and output will spike. That includes grant proposals, where volume alone can bury funders.
Evidence the flood has started
Funders are reporting a surge in applications likely boosted by LLMs. Denmark's science minister recently said public research funders are being "run over" by proposals. Success rates in Horizon Europe have dropped sharply this year, with some consultants blaming AI-assisted submissions.
Conferences are seeing submission spikes, too. Meanwhile, analyses suggest LLM fingerprints in scholarly prose, with unusual surges in terms like "underscore," "delve," and "pivot." Tools are lowering language barriers for researchers who don't work in English natively-that's a positive use-but the net effect still points to more volume.
The worst-case loop
Nolan's concern isn't just more content. It's the feedback loop where academia turns to AI to manage the mess it helped create-LLMs summarizing literature, screening grants, and triaging papers. That's risky territory given hallucinations, bias, and the tendency to favor textual conformity over novelty.
European guidance echoes this caution. The European Commission has warned against using LLMs to assess proposals, and broader ethics guidelines for trustworthy AI stress human oversight and accountability. If we let AI define "fit," bold ideas get filtered out.
What's actually needed: incentive reform, not just tools
AI isn't a fix for misaligned incentives. We need to change what we reward. Nolan argues for experimentation in how we assess, publish, and fund research-less conservatism, more trials of new models.
Immediate steps for research leaders and PIs
- Publish fewer, better papers. Consolidate related results instead of slicing them across multiple outlets.
- Use LLMs for editing and structure, not idea inflation. Always disclose AI assistance and keep a record of prompts and outputs.
- Prioritize synthesis. Write review papers and living overviews that connect fields and reduce duplication.
- Shift lab KPIs from paper count to originality, data/code sharing, and real-world or cross-field impact.
For departments and committees
- Update hiring and tenure criteria. De-emphasize raw counts; value novelty, reproducibility, and open practices.
- Adopt narrative CVs that highlight contributions, leadership, and team science.
- Reward high-quality null results, replications, and robust datasets.
For funders
- Introduce two-stage calls (short pre-proposals, then full proposals by invite) to cut wasted effort and review load.
- Set submission caps per PI or per institution per call to prevent volume spikes.
- Use partial lotteries among proposals that clear quality thresholds to save time and reduce bias.
- Require AI-use disclosures in proposals and forbid LLMs for evaluative judgment. Automate only clerical checks with human oversight.
- Pilot random audits for AI misuse and overreliance on templated text.
For journals and conferences
- Limit serial publication of near-duplicate studies; encourage integrated manuscripts.
- Offer Registered Reports and fast lanes for high-quality replications and datasets.
- Require structured methods, data/code availability, and AI-use statements at submission.
- Desk-triage with humans, not LLMs. Use AI only for format checks and plagiarism flags, then review with expert eyes.
Guardrails that keep novelty alive
- Make transparency default: disclose if, where, and how AI assisted writing or analysis.
- Protect ideas that don't "read" like previous winners. Screen for sameness bias in any automated step.
- Invest in human synthesis. Fund expert panels and living reviews that distill the literature and guide priorities.
Metrics that matter
- Track disruption/novelty indices, not just citation counts.
- Monitor reviewer workload, time-to-decision, and resubmission loops across venues.
- Measure topic diversity in funded work and acceptance decisions.
- Audit the share of AI-written or AI-heavy text and its relation to outcomes.
Bottom line
LLMs lower the cost of writing. If we keep paying for volume, we'll get more of it-and bury signal under noise. The fix is governance: change what counts, how we review, and where we spend attention.
Use AI to reduce friction, not standards. Keep humans in charge of novelty, judgment, and risk-taking. And redesign incentives so fewer, stronger contributions win.
Resources
- OECD AI principles
- EU ethics guidelines for trustworthy AI
- Prompt engineering practices for responsible AI use
Your membership also unlocks:
