Last time, we stared into the abyss of the Cartesian demon reborn as a 405B-parameter LLM, capable of conjuring entire scientific corpora that fool even the jaded. You read a paper claiming waffles accelerate baldness? It might be a real study from a mid-tier journal, p-hacked into oblivion by desperate postdocs. Or it might be a hallucinated artifact, spat out by Chat or Claude in the time it takes to blow your nose. The philosophical impact hits hard: how do you know which is which? And why bother, when the next one might be equally untrustworthy?
But here's the rub: this isn't just theoretical dread anymore. It's happening, at scale, in 2025. The demon isn't lurking in some philosopher's thought experiment; it's embedded in the journals, the databases, and even the chatbots academics use to navigate it all. And the fallout? Retractions pouring in like confetti at a funeral, whistleblowers ghosted by editors, and AI tools — meant to fix the mess — compounding it instead. Welcome to the sequel: where scepticism isn't a luxury, it's survival.
The Retraction Tsunami: From Trickle to TorrentRemember when a retraction was a rare scandal, whispered about in faculty lounges? In 2023, we hit 10,000 globally. By 2024, it was pushing 15,000. Now, in late 2025, we're on track for 20,000, with AI-fuelled fraud leading the charge. It's not just volume; it's velocity. Journals that once dragged their feet for years are yanking papers in days — when they act at all.
Take Neurosurgical Review, a Springer Nature stalwart. In February, they began retracting dozens of AI-generated commentaries and letters, swamped by synthetic slop from authors at places like Saveetha University in India (already a retraction magnet with 80+ pulls in 2024 alone). The fakes? Polished gibberish on neurosurgery topics, complete with bogus citations and stats that scanned clean on initial checks. One author griped that retractions for "procedural oversights" like undisclosed LLM use were overkill — misconduct, he said, means fabricated data, not just lazier writing. Fair point, but when the whole piece is 80% prompt-engineered, where's the line?
Or consider Frontiers in Cell and Developmental Biology, a heavy hitter with 114,000+ citations in its portfolio. Last August, they retracted a peer-reviewed paper riddled with AI hallucinations: diagrams of "protemns" (oops, proteins), garbled acronyms like "zxpens," and — wait for it — a rat with cartoonishly exaggerated genitalia grafted onto a biology figure. The authors admitted using AI for visuals but skipped verification. Reviewers? Missed it. Editors? Initially greenlit it. Now the journal's scrambling: how many more AI-illustrated bombs lurk in their back catalogue, shaping drug trials or gene therapies?
And it's not isolated. A Lancet sub-journal just pulled a COVID metformin paper two years after the authors flagged issues themselves. Whistleblowers, meanwhile, are playing whack-a-mole with 300+ suspect papers from Japanese docs Yoshihiro Sato and Jun Iwamoto — ethical lapses galore, with 121 retractions so far. Journals ghosted half their complaints, deferring to glacial institutional probes that often vanish into the ether.
This isn't cleanup; it's collapse. The "publish or perish" grind — millions of papers yearly, special issues ballooning like MDPI's 3,000+ in one journal alone (at £2,600 a pop) — has birthed paper mills churning fakes for hire. Hindawi (Wiley's arm) axed 8,000+ in 2023; 2025's just warming up. Sleuths like Guillaume Cabanac are begging publishers to retract 3,000+ more they've flagged. Science doesn't "auto-correct," he says—it needs humans with pitchforks.
The Demon Doubles Down: AI as Both Culprit and CopHere's the cruel irony: academics are turning to AI to fight AI. Tools like Review Feedback Agents (tested on 20,000+ conference reviews) promise to spot fraud faster than overworked peers. But plot twist — they are bad at it. An October study quizzed 21 LLMs on retracted papers: on average, they flagged fewer than half correctly, generating inconsistent answers to identical prompts. Worse, these models are trained on retracted junk, regurgitating debunked claims in chat responses. Ask GPT-5o about a pulled COVID study? It might cite the ghost data verbatim, oblivious to the retraction buried in some database it half-remembers. In short, AI LLMs cannot be trusted as many lawyers have found out the hard way with cases just being invented, hallucinated:
Even visuals are doomed. AI microscopy images? Indistinguishable from real ones, per October reports — perfect for faking cell slides in cancer research. And ethics? A June paper argues AI-generated figures with "major errors" (like omitted organs or added mutations) warrant retractions if they mislead — but only if caught early. Good luck with that in a system where peer review is a 30-minute skim.
Social media's the new canary: Xbuzz predicts retractions weeks ahead, as sleuths dissect PDFs live. But even there, the demon lurks — deepfakes of "studies" go viral, eroding trust faster than journals can scrub.
So, where does this leave us? In full Descartes mode: doubting everything empirical because the medium is now as suspect as the message. Consensus science? It's a house of cards built on sand. Policy? Imagine regulators citing a "study" that's just 1,200 tokens of probabilistic prose.
But despair's a dead end. The old priesthood is toast; time for democratic alternatives. Prediction markets to bet on findings before they're published. Blockchain-ledgers for immutable data provenance. Open-source replication bounties, crowdfunded via crypto. And maths—glorious, self-verifying maths—as the anchor. Euclidean proofs don't need LLMs; they just work, bridge after bridge.
We're not post-science; we're pre-something-better. The demon exposed the flaws; now we forge tools that even it can't fake. In the meantime, verify ruthlessly. Demand raw data. And if a paper links breakfast carbs to cranial sparseness? Use your own brain, not unreliable AI to debunk it.