AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Guarding the Gates of Knowledge

Collected Nov 1, 2026

IEEE Spectrum editor in chief Harry Goldstein uses a recent editor's column to connect two efforts against machine-generated content in knowledge sources people rely on. The first is a Spectrum feature by freelance writer and applied mathematics PhD Benjamin Skuse, "I Used AI to Fake a Scientific Paper," in which Skuse tried to get his own AI-generated papers through IEEE's publishing defenses. The second is a Five Questions interview by features and AI beat editor Eliza Strickland with Ellia Lebleu, a founding member of WikiProject AI Cleanup, a volunteer effort to detect and delete AI-generated material on Wikipedia.

Goldstein frames the problem through his own experience. Scrolling social media, he found roughly half his feed filled with what he calls AI slop, some of it shared unwittingly and some posted on purpose "because they found it amusing or useful for sowing confusion." He traces the trust problem back to GPT-3, when chatbots would answer confidently with plausible but false content, the period when "hallucination" became the everyday term for that failure. Six years and several chatbot versions later, he writes that he still cannot trust the answers he gets, even though Claude or Gemini often do useful work for him. His working rule is trust but verify.

Applied to scholarly publishing, that rule describes the problem Skuse set out to test. According to Strickland, Skuse was surprised at how effortless it was to produce a paper that was more or less plausible and more or less science. In one of his experiments, she said, he essentially hit "run experiment" and "write paper," left the output untouched, and sent it into what she called the maw of the publishing engine.

What Skuse found, per Goldstein's account, is that IEEE's combination of human peer review and research-integrity checking software is, for now, keeping pace with people armed with free tools. Goldstein characterizes this as a cat-and-mouse game. The harder problem is that more capable LLMs no longer make the obvious mistakes the free versions do, which he says leaves publishers in an arms race between increasingly convincing slop and the detection systems built to catch it.

For a developer, the mechanics behind that race are worth spelling out. Publishers run manuscripts through integrity tools that look for signals such as duplicated or manipulated figures, citation patterns that point to nonexistent or irrelevant work, tortured phrasing, and statistical oddities in reported results. Free or lower-tier models tend to leave detectable fingerprints of exactly this kind. Frontier models are better at producing fluent prose, coherent citations and plausible-looking methods sections, which erodes those signals. A likely trade-off is that detection will lean more on process and provenance, such as requiring raw data, code and author verification, than on text classifiers, which are prone to false positives against careful human writers and can be gamed by paraphrasing. This is my reading of where the technical pressure points sit, not something the column reports.

Wikipedia faces a version of the same contest with a very different resource base. WikiProject AI Cleanup is a volunteer group, and Goldstein presents it as confronting the same flood of machine-written material as commercial publishers but with fewer people and fewer tools. Strickland's point is that Wikipedia's identity rests on being a repository of human-generated information and that its editors want to keep it that way, while acknowledging uncertainty about how long they can.

Goldstein widens the argument at the end. If neither the scientific record nor basic facts about the world can be assumed trustworthy, he writes, our sense of what is real unspools into uncertainty. He concedes that some may welcome that condition, but argues it ultimately undermines trust in the institutions whose job is to curate human knowledge and to advance science and technology for humanity's benefit. He closes with Strickland's speculation that in a world where AI does the science and writes the papers, human trust in the outputs may stop mattering, and with her remark that it feels crazy to imagine AI eventually writing for other AIs, with people left out of the loop.

For product and engineering teams, the practical takeaway concerns provenance. Systems that ingest scientific literature or encyclopedia content now need to treat source credibility as a first-class signal rather than an assumption, and pipelines that summarize or cite such material inherit whatever contamination they ingest. My inference is that demand will grow for verifiable data trails, author and process attestation, and human review checkpoints, because detector accuracy alone is unlikely to settle the question. Teams building research assistants, search or knowledge products should plan for a steady adversarial cycle rather than a one-time fix.

Why it matters: Anyone who builds on top of scholarly literature or reference content is affected, because the integrity of that corpus is now actively contested. The reported finding that current IEEE review plus integrity tools still catch free-tool fakes suggests defenses work today; the shift to stronger models suggests that advantage may narrow, which in my reading pushes teams toward provenance and human verification over text detection. Wikipedia's volunteer effort illustrates that the burden falls hardest on open, low-resource projects, a gap product teams may end up filling or inheriting.

Read at IEEE Spectrum · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

A recent scroll through social media got me thinking about trust. About half my feed was filled with AI slop that astute users would call out as being machine-generated, often to the embarrassment of the rube who unknowingly spread it and sometimes to the delight of the poster who shared it deliberately because they found it amusing or useful for sowing confusion. Remember the good old days of GPT-3 when we could ask the chatbot anything and it would almost instantly and very confidently spit back a plausible but f