Peer pressure: Will AI slop swamp science?
OpenAI just released Prism, a free LaTeX workspace with GPT-5.2 built in. It drafts text, formats citations, turns whiteboard sketches into diagrams, and lets co-authors work in real time. It's positioned as a time-saver so scientists can "focus on the science."
That pitch lands in a tense moment. Researchers and publishers are already worried about "AI slop" flooding journals-polished papers that read well but don't move a field forward.
What Prism actually is
Prism rides on tech from Crixet, a cloud LaTeX platform OpenAI acquired in late 2025. It's a writing and formatting tool-not a system that runs experiments or proves theorems on its own.
Even so, OpenAI's demos show it auto-finding literature and formatting bibliographies. Kevin Weil, OpenAI's VP for Science, says AI is shifting from curiosity to core workflow, noting weekly traffic to ChatGPT on "hard science" topics. He also admits: the model can invent citations, and researchers must verify references themselves.
Why researchers are uneasy
Lowering the cost of producing polished manuscripts raises a blunt risk: more submissions, same review capacity. Editorial screens often use presentation quality as a proxy for merit. AI breaks that proxy.
We've seen the failure modes. Meta pulled Galactica in 2022 after it produced convincing nonsense. In 2024, Sakana AI's "AI Scientist" drew criticism for papers a journal editor said he'd desk-reject: "very limited novel knowledge."
What the data says
A 2025 paper in Science reported 30-50 percent higher output from AI-assisted authors, but worse peer-review outcomes. Reviewers seemed to detect when intricate prose papered over weak science. As Cornell's Yian Yin put it, this is "a very widespread pattern" that warrants a hard look from funders and gatekeepers.
A separate analysis of 41 million papers (1980-2025) found AI-using researchers publish more and get cited more-while the collective scope of exploration narrows. Yale's Lisa Messeri warned that should trigger "loud alarm bells": the tool can benefit individuals while "destroy[ing] science" as a collective enterprise.
Policies are tightening. Science's H. Holden Thorp wrote that the journal is "less susceptible" to AI slop thanks to human editorial investment, but "no system, human or artificial, can catch everything." The journal allows limited AI for editing and reference gathering-with disclosure-and prohibits AI-generated figures. Cambridge University Press's Mandy Hill has called for "radical change," noting the publishing system is already under strain and AI will add pressure.
Acceleration vs overload
OpenAI highlights cases where GPT-5.2 sped real work: a mathematician solving an optimization problem over a few evenings, a physicist reproducing symmetry calculations that took months. That drifts beyond writing help. The line is blurring.
There are real upsides for researchers who don't write English fluently. But if the submission firehose opens wider, peer review buckles. Weil's goal is "10,000 advances" that compound. The open question: will this yield more knowledge-or just more papers?
UC Berkeley statistician Nikita Zhivotovskiy says GPT-5 helps polish text and catch math typos-useful, practical. The flip side: polished veneer can push weak work through initial screens. Conversational workflows can hide assumptions and spread responsibility too thin.
The anxiety is plain in technical communities: "We're in a post-scarcity society now, except the thing we have an abundance of is garbage-and it's drowning out everything of value."
Practical guardrails for labs, reviewers, and editors
- Set a lab policy. Allow AI for grammar, clarity, and LaTeX formatting. Prohibit AI for claims, proofs, model choices, or interpretation. Write it down and enforce it.
- Disclose precisely. In the paper and cover letter, state where AI was used, with examples. Keep a log of prompts/outputs in your repo or appendix.
- Verify every reference. Use a reference manager with DOI lookup. Spot-check each citation for existence, correctness, and relevance. Treat AI-suggested sources as leads, not facts.
- Tighten methods. Pre-register when possible. Publish data, code, configs, and seeds. Add a provenance statement: which sections, figures, or equations had AI assistance.
- Stress test your claims. Run ablations, alternative specifications, and negative controls. Ask explicitly: what would falsify this result?
- Figures and diagrams. Stick to human-generated figures unless a venue allows AI with disclosure. Archive source files.
- Internal review before submission. One colleague checks proofs/math. Another runs the code from scratch. A third checks citations and prior art overlap.
- For reviewers. Request code/data and an AI-use disclosure. Sample-check citations. Evaluate novelty and validation over "polish." Flag conversational vagueness in methods.
- For editors. Require structured AI-use forms, enforce consequences for false disclosure, and triage with heuristics (citation anomalies, generic phrasing, methods-claims mismatch). Consider caps on simultaneous submissions and invest in editorial staff.
- For funders/institutions. Shift incentives from paper counts to verified contributions, open materials, and replication. Support reviewer time as funded work.
Sensible ways to use Prism today
- Use it to clean prose, fix LaTeX, and standardize formatting. Save the hard thinking for you and your co-authors.
- Never trust auto-generated references. Validate via DOI/Crossref and your library tools.
- Do not outsource core science: problem framing, derivations, statistical design, or interpretation.
- Let it summarize prior work, then read the originals before citing.
- Document every assisted step. Add an "AI assistance" note in your README and paper appendix.
- Run a contamination check: for each equation, claim, and figure, list source (derivation, experiment, or citation) and who verified it.
Useful references
Editorial guidance and ongoing coverage from major venues can help you calibrate your policies and disclosures.
If you're setting lab-wide norms for responsible AI-assisted writing, this curated hub may help you upskill your team's prompts and workflows: Prompt Engineering resources.
Bottom line
Prism will cut the time to produce clean LaTeX and readable drafts. The volume of submissions will rise. Quality control must rise faster.
Treat AI like a calculator for prose: helpful, unforgiving, and your responsibility. Keep the bar high, document everything, and make the science-data, code, and reasoning-so clear a rushed review can still catch the truth.
Your membership also unlocks:
