The AI Slop-Vigilant Zeitgeist

In 2026 Detectors, Humanisers, and the Stigma of AI-Assisted Writing are all having their year

As of mid-2026, the strongest AI text detector out there seemed to be Pangram, which demonstrated near-zero false positive rates on long-form text. The tool survives most commercial "humaniser" tools in independent testing, but the detection market as a whole remains unreliable, mainly on short, edited, or paraphrased text and on the writing outputs of non-native English speakers. I've typically been a huge sceptic of detector tools, and I rigorously subjected Pangram to a stress test of my own. Now, the 3-part sample set of 'my writing' and 'my-Claude's writing' (my writing with Claude's elegant polish) and 'Pure Claude' was something I could rigorously control. Because in this current landscape of stealth AI use by absolutely every shame-bound human, we can only really know within ourselves the truth and integrity of our authorship claims. And that's precisely the problem I want to thrash out here. The shame, the risk of reputational, academic, and legal harm, and the fear, which is completely valid I might add, of Slopper name-calling.

So the detection field has split into tiers. Pangram now, at the time of writing, sits at the top with independently corroborated false positive rates near zero, while older perplexity-and-burstiness tools and most open-source detectors remain unreliable enough that they cannot responsibly be used as proof; vendor accuracy claims in the high-90s consistently overstate real-world performance.

As it currently stands, the arms race is genuine but asymmetric: 'humanisers' do reliably defeat weaker detectors (ZeroGPT, Winston AI, Writer.com) yet largely fail against Pangram and post-August-2025 Turnitin; however, simple adversarial edits can cut average detector accuracy in about half, so the picture depends heavily on which detector is used.

The publishing and journalism front

The "Shy Girl" novel cancellation (Hachette Book Group, March 2026) is probably the first big commercial novel pulled by a major publisher specifically over AI suspicion. The New York Times reported Pangram's assessment that the novel was 78 per cent AI-generated or assisted; a finding that author Mia Ballard denied. Specifically, Ballard denied 'personally' using AI, attributing it instead to an editor she had hired, and described severe reputational and mental-health harm.

Per a study led by University of Maryland's Mohit Iyyer with Pangram Labs (arXiv preprint 2510.18774, October 2025), an examination of 186,000 articles from 1,500 newspapers in 2025 found "9.1% of articles are labeled by Pangram as either AI-Generated (5.2%) or Mixed (3.9%)". Usage was only 1.7 per cent at papers over 100,000 circulation versus 9.3 per cent at smaller outlets, with Boone News Media highest at 20.9 per cent, and AI use rarely disclosed.

Pangram Labs screened all 19,490 papers and 75,800 peer reviews for ICLR 2026 (prompted by a bounty from Carnegie Mellon's Graham Neubig) and found "21%, or 15,899 reviews, were fully AI-generated", with over half showing some AI involvement, as reported in Nature by Miryam Naddaf in 2025.

So, just checking in: if you're not yet convinced there's some emerging moral panic over this, wait, there's more.

Per Nature (September 2025), using Pangram's tool, the American Association for Cancer Research found "23% of abstracts in manuscripts and 5% of peer-review reports submitted to its journals in 2024 contained text that was probably generated by large language models"; fewer than 25 per cent of authors disclosed AI use despite a mandated requirement to do so.

So how does all this 'detection' work?

If you guessed there's something circular going on here, and we're talking about an AI trained to detect AI, you'd be on the money, well done. You've been paying attention.

Perplexity and burstiness

Perplexity measures how "surprised" a language model is by a sequence of words; it quantifies predictability. Because LLMs are trained to predict the most probable next token (or piece of a word, or punctuation), their output tends to have low perplexity, meaning it is statistically predictable and "unsurprising". GPTZero historically used a perplexity threshold around 20, treating text below it as more, well, machine-like. A sentence such as "Hi there, I am an AI assistant" has very low perplexity because each word is the highly probable continuation.

Burstiness is slightly different. It's about the rhythm of writing. It measures how much that perplexity (and sentence length and structure) varies across a document. Human writing tends to be "bursty", because we apparently mix up short, punchy sentences with longer, more elaborate ones, generating uneven rhythm. This is something which GPTZero attributes philosophically to short-term memory dissuading us from writing similar things twice. In statistician lingo, AI writing tends to regress toward a consistent mean sentence length, producing low burstiness; GPTZero describes language models as writing with a consistent "AI-print".

The two metrics, perplexity and burstiness, are typically combined into a composite score, because either alone produces too many errors. But combine them and we get, well, um, fewer errors?

The limitations are well documented. Perplexity and burstiness are probabilistic signals, not proof. Highly formal or technical human writing (legal contracts, medical reports, technical manuals) can show low burstiness and trigger false positives. Now that's a problem for polished writing; I don't fancy being the one to tell contract law firms and research scientists they need to amp up their signature burstiness if they want to continue to be taken seriously.

But fortunately, modern commercial detectors have moved beyond these two metrics: GPTZero now describes a seven-component, multilayered system in which perplexity and burstiness form only the first statistical layer; and Pangram uses a trained transformer classifier operating on overlapping windows of text rather than relying on perplexity alone, allowing it to flag localised AI segments inside otherwise human manuscripts. Pangram's technical report explicitly notes that perplexity-based methods such as DetectGPT, Sniffer and DetectLLM "fail on human-written documents that happen to be in the training set of LLMs, such as the Declaration of Independence". That explains why the early-2023 wave of perplexity tools generated silly headlines about the US Constitution being flagged as AI.

So, are word-frequency and phrase signals reliable?

Word-frequency and phrase signals

The evidence suggests phrase-based and pure statistical signals are useful as corroboration but (pay attention please) not as proof. There is a genuine, measurable phenomenon. The peer-reviewed study by Kobak, González-Márquez, Horvát and Lause, "Delving into LLM-assisted writing in biomedical publications through excess vocabulary" (Science Advances, 2025, volume 11, number 27, eadt3813; DOI 10.1126/sciadv.adt3813), analysed more than 15 million PubMed abstracts from 2010 to 2024 and found an abrupt rise in certain "style words" after ChatGPT, estimating that "at least 13.5% of 2024 abstracts were processed with LLMs". The study identified 900 excess words and made them available as data (Zenodo DOI 10.5281/zenodo.14898666). Words such as "delve", "intricacies" and "unwavering" showed striking frequency increases.

But the same signals produce false positives and bias. The Stanford study by Liang, Yuksekgonul, Mao, Wu and Zou, "GPT detectors are biased against non-native English writers" (arXiv 2304.02819, published in Patterns, 2023), found that detectors classified more than half of non-native English (TOEFL) essays as AI-generated while being near-perfect on native US 8th-grade essays. The unanimously misclassified TOEFL essays had much lower perplexity, suggesting these detectors "penalize authors with limited linguistic expressions". The paper warned against detectors that are "unreliable and easily gamed". This is the unwavering academic critique (see what I did there). Low perplexity is a biased indicator that systematically disadvantages authors with restricted lexical range, including ESL writers and some neurodivergent writers. Not to get too lost in the intricacies further, it's worth highlighting a confounding note: an August 2024 ETS study on roughly 2,000 GRE samples did not find bias in detectors it trained itself, though it used a simplified setting and did not test the commercial detectors used in practice.

The phrase-spotting approach also degrades over time. As Carnegie Mellon's Daphne Ippolito has noted, AI "tells" are only "transiently useful" because companies constantly revise training recipes, so the telltale vocabulary shifts. So this isn't a trivial detail; it is an expensive and onerous factor for the detection industry to stay ahead of their game, because every frontier model update degrades their product.

And what's more, the cues are contaminating human writing: a Max Planck Institute for Human Development analysis reportedly found podcasters and YouTubers increasingly using AI's favoured words such as "delve" even in spontaneous, unscripted speech.

The detection-evasion (humaniser) landscape

What humanisers do is they attempt to raise perplexity and burstiness and strip "AI tells" while preserving meaning. And if you've ever used one, they also tend to ruin your writing, and that's not on the tin.

Their actual effectiveness is highly dependent on the target detector and is best understood in tiers. Against weak detectors, leading humanisers achieve high bypass rates of 90 per cent and above. Against Originality.ai and updated Turnitin they are inconsistent or failing, and against Pangram (AT THE MOMENT!) they largely fail.

The single most important independent academic finding is from the Chicago Booth study: Pangram's false negative rate "is robust to the use of current 'humanizers'", remaining low even when AI passages were modified using StealthGPT, whereas Originality.ai and GPTZero were "susceptible to 'humanizers'". This corroborates Pangram's own humaniser benchmark, which detected Undetectable AI 90.3 per cent of the time, StealthGPT 95.6 per cent, GPTinf 99.2 per cent and Quillbot 100 per cent. Noteworthy is the heavy source-quality caveat: much "independent" humaniser testing comes from competing humaniser vendors with conflicts of interest, and no truly neutral peer-reviewed head-to-head bypass study exists for WriteHuman, Humbot or Phrasly specifically against Pangram.

A critical caveat in the other direction comes from the peer-reviewed Perkins et al. study, "Simple techniques to bypass GenAI text detectors: implications for inclusive education" (International Journal of Educational Technology in Higher Education, 2024, 21:53; DOI 10.1186/s41239-024-00487-w). It found six major detectors had a baseline accuracy of 39.5 per cent that dropped a mean of 17.4 percentage points under simple adversarial techniques, and concluded the tools "cannot currently be recommended for determining whether violations of academic integrity have occurred". The 39.5 per cent baseline reflects a broad cross-section of detectors and conditions, not the best single detector on long text, which is why this figure and Pangram's near-perfect figures are not contradictory; they describe different parts of the market.

The (huge) burden of proof problem

A recurring concern across legal and academic sources is asymmetry: an accused person must prove a negative (that they did not use AI), usually with no transparent access to how a proprietary detector reached its score. The peer-reviewed article "Heads we win, tails you lose: AI detectors in education" (Journal of Higher Education Policy and Management, 2026; DOI 10.1080/1360080X.2026.2622146) argues that AI detector results, linguistic markers, style changes, student silence, and confessions under pressure individually and collectively "will never reach the level of 'more likely than not'", the balance-of-probabilities standard. Turnitin itself states its score is not definitive proof of cheating, and OpenAI shut down its own classifier in July 2023 citing low accuracy, later stating detectors were not "reliable enough given that educators could be making judgments about students with potentially lasting consequences".

Real cases illustrate the human cost. The New York Times ("A New Headache for Honest Students: Proving They Didn't Use A.I.", 17 May 2025) documented Leigh Burrell, flagged by Turnitin despite Google Docs history showing two days of drafting; she restored her grade only after submitting a 15-page PDF of time-stamped screenshots. A University of Minnesota PhD student was expelled and filed a due-process lawsuit; a French-born Yale EMBA student sued, alleging a coerced confession and discrimination as a non-native English speaker. And for a local facepalm, Australian Catholic University recorded nearly 6,000 alleged academic misconduct cases in 2024, about 90 per cent AI-related, with a substantial share dismissed after investigation (thanks ABC reporting). The University of Pittsburgh stopped supporting AI detection tools, citing false positives and "potential legal sanctions". Students now engage in "self-surveillance", recording their screens for hours.

Well, this is going really well, isn't it?

In journalism and publishing, the same asymmetry applies, with the added twist that accusations frequently originate from crowds rather than tools. The "Shy Girl" suspicion began with Goodreads and Reddit readers and a YouTube video ("I'm pretty sure this book is ai slop") that reached over 1.5 million views before Pangram's analysis was reported. The New York Times "Modern Love" column "I Was Deemed Unfit to Be a Mother" by Kate Gilgan was challenged after AI researcher Tuhin Chakrabarty estimated via Pangram that over 60 per cent might be AI-generated; Gilgan said she used AI for editorial guidance but not to produce the content. Notably, the Maryland and Pangram newspaper study explicitly cautioned that "all findings should be interpreted as detector outputs rather than definitive authorship attributions" and declined to attribute misconduct to individuals, a discipline that public "shaming" campaigns rarely observe.

The cultural moment: encouraged and shamed

AI-assisted writing is promoted by employers and tool vendors while being stigmatised socially. The em dash became a viral "AI tell" in 2025, prompting writers and brands to purge it, and amusingly, prompting hundreds of thousands of frontier model subscribers to access the personalisation panel on their settings page and enter a strongly worded statement that forbids em-dash use (just in case they miss one in the human oversight parse).

The structural device "It's not just X, it's Y" and the rule of three have likewise become suspect, even though LLMs did not invent them. A Duke University study by Jessica Reif, Richard Larrick and Jack Soll, "Evidence of a social evaluation penalty for using AI" (PNAS, 2025), ran four experiments with about 4,400 participants and found workers who use AI are viewed as "lazier, less competent, and less diligent than people who get help from other sources", with the penalty disappearing when the evaluating manager also uses AI. The "delve" panic also carries a documented racial dimension, because "delve" is common in Nigerian English and many data annotators who shape model output are based in Africa, so treating it as an "AI tell" can function as a swipe at a population of writers. Visual artists have faced parallel treatment; DC Comics dropped covers by artist Francesco Mattina after AI-use accusations.

But all the technical benchmarks and linguistic markers aside, the defining story of 2026 worth telling is cultural, not technical; a moral panic if you like. AI-assisted writing is simultaneously encouraged and shamed. Now this is both important and interesting to me, as my day job has me running professional development workshops on how to use Copilot. I dedicate considerable workshop time basically training humans to train Microsoft's new shiny toy... ahem... tool, to write stuff for them. But hold that up side by side with the preponderant anxiety over AI detection, owing to how it has produced book cancellations, shameful sloppy journalism controversies, expulsions, lawsuits, and a viral panic over the em dash. All while the burden of proof has effectively shifted onto the accused to prove a negative. While people like me are running workshops, and big tech people are trotting out ever shinier, more talented toys (when they grow up we can call them tools ok!), the outcome is a lot of desperate arm-waving, doubling down on 'no, I didn't, I wrote it, it's mine ...promise'. So for now, the Zeitgeist has kind of backed itself into a corner.