IEEE Spectrum column: peer review is still keeping pace with AI-written papers
An IEEE Spectrum editorial column opens with a personal note on trust. After a scroll through social media, its author says about half the feed was AI slop that sharp-eyed users would call out as machine-generated: sometimes to the embarrassment of the person who spread it unknowingly, sometimes to the delight of a poster who shared it on purpose because it was amusing or useful for sowing confusion. The author recalls the GPT-3 days, when chatbots confidently returned plausible but false answers and "hallucination" entered the vocabulary. Six years and several chatbot versions later, the author still cannot trust the answers, even though Claude or Gemini often do useful work. The maxim is to trust but verify.
The column then applies that maxim at scale to the scientific record. It points to a piece by Benjamin Skuse, a freelance writer and doctor of applied mathematics, titled "I Used AI to Fake a Scientific Paper." Skuse tried to get his own AI-generated papers past IEEE's defenses, and the article tells the story of the human gatekeepers trying to beat back the rising tide of slop.
According to features and AI beat editor Eliza Strickland, as relayed by the column author, Skuse was surprised at how effortless it was to produce a paper that was more or less plausible and more or less science. For one experiment, Strickland said, Skuse basically hit "run experiment" and "write paper," did nothing further to the result, and launched it into the maw of the publishing engine.
The finding reported in the column: the combination of IEEE's human-review processes and its research-integrity checker programs is currently keeping pace, in what amounts to a cat-and-mouse game, against people using free tools. The caveat is that more advanced LLMs no longer make the same dumb mistakes the free versions do. That leaves publishers in an arms race between increasingly plausible slop and the systems built to detect it. The column does not say how many papers Skuse generated or submitted, which AI tools he used, or whether any paper was accepted or published, and it cites no detection rates.
The column also covers Wikipedia, which it calls the global depository of human-generated knowledge. For this month's Five Questions, Strickland interviewed Ellia Lebleu, a founding member of WikiProject AI Cleanup, a volunteer effort to detect and delete AI-generated material from the platform. Lebleu's team faces the same problem as commercial publishers, but with fewer people and fewer tools. Strickland's view: Wikipedia is known as a repository of human-generated information and wants to hold on to that, but who knows for how long.
The author closes with two scenarios. If people cannot assume that the scientific record is trustworthy or that basic facts are accurate, understanding of what is real unspools into uncertainty, which undermines trust in the institutions that curate human knowledge and promote science and technology. Or, if AI ends up doing the science and writing the papers, it may not matter whether humans trust the outputs. Strickland puts it this way: it feels crazy that AI will be writing for other AIs at some point, and humans will get left out of the loop.
Key facts
- Freelance writer and applied mathematics PhD Benjamin Skuse tried to get his own AI-generated papers past IEEE's defenses; his article is titled "I Used AI to Fake a Scientific Paper."
- Per the column, IEEE's human review and research-integrity checker programs are currently keeping pace in a cat-and-mouse game against people using free tools.
- More advanced LLMs no longer make the same dumb mistakes as free versions, which leaves publishers in an arms race between increasingly plausible slop and detection systems.
- Skuse was surprised how effortless it was to generate a more or less plausible paper; for one experiment he ran the experiment, wrote the paper and submitted it without editing.
- Wikipedia's volunteer WikiProject AI Cleanup, interviewed by Eliza Strickland via Lebleu, faces the same problem as commercial publishers with fewer people and fewer tools.
Why it matters
The column frames AI slop as a trust problem that reaches from social feeds to the scientific record. Its point is that chatbots still cannot be trusted without checking, and that checking gets harder at the scale of peer-reviewed publishing. A first-hand test by a writer who deliberately submitted AI-generated papers to IEEE gives a rare look at whether the gatekeepers hold. For now the answer is yes against free tools, but the author expects better LLMs to tighten the contest.
Who it affects
Publishers of peer-reviewed scholarship, and the reviewers and integrity staff who act as gatekeepers. Researchers and readers who rely on the scientific record are affected too. So are volunteer communities such as WikiProject AI Cleanup on Wikipedia, which the column says works with fewer people and fewer tools than commercial publishers. The author also points to a wider loss of trust in the institutions that curate human knowledge.
How to use it
The column is a pointer to two pieces rather than a tool. One is Skuse's article, "I Used AI to Fake a Scientific Paper. Here's What Happened." The other is this month's Five Questions interview with Ellia Lebleu. It also lists a related piece on how Wikipedia's editors keep the encyclopedia human-written, and one titled "AI Boosts Research Careers but Flattens Scientific Discovery." The practical habit the author offers is the maxim to trust but verify, applied to chatbot output.
How solid is it
This is an editorial column, not a study. The central finding about IEEE's defenses is relayed second-hand: the column author reports what Skuse discovered and what Strickland told them. It gives no detection rates and no count of papers submitted, and it does not say which tools or models were used. The social media observation of "about half" the feed being slop is the author's impression, not a measurement.
Risks and caveats
The reassurance is explicitly temporary: the column says defenses are "currently" keeping pace, and only against free tools. Advanced LLMs no longer make the same obvious errors, so detection may face a harder contest. The column does not say whether Skuse tried advanced models or how detection would fare against them. Wikipedia's volunteer effort has fewer resources than commercial publishers. The author also floats a bleaker possibility, that AI writing for other AIs could leave humans out of the loop.
“He basically hit ‘run experiment’ and ‘write paper.’ And then he didn’t do a damn thing to it.”
— Eliza Strickland, features and AI beat editor, about Benjamin Skuse