So What? How AI Detectors Work  🔗 **Source:** [So What? How AI Detectors Work](https://www.christopherspenn.com/2026/07/so-what-how-ai-detectors-work/) > [!note]+ Document Notes > tags: [ai-detection, false-positives, equity, academic-integrity, disclosure, writing, classifiers, neurodiversity] > author: Katie Robbert, Christopher S. Penn, John Wall > source: christopherspenn.com / Trust Insights "So What?" > reference: https://www.christopherspenn.com/2026/07/so-what-how-ai-detectors-work/ > date created: 2026-08-03 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ So What? How AI Detectors Work > <!-- deep-read:sections start --> ## 🧭 Table of Contents - 🔗 References · 📝 Summary · 📖 Main Content · 📚 References · 🗣️ Paraphrase · 📰 Newsletter - ✅ Action Items · 🧪 Hypotheses · 📊 Methodology · ❓ Questions · 🔮 Future Research · 💡 Implications - 🧠 Learning Styles · 📊 Accessibility · 🌍 DEIB · ✅ Logical Validity · 🛡️ DARVO · 🎯 Conclusion · 🧱 Refactor Appendix ## 🔗 References - Article: https://www.christopherspenn.com/2026/07/so-what-how-ai-detectors-work/ - Companion video: https://www.youtube.com/watch?v=_5DanA1Zo9M (2026-07-31, 46m05s) - Hosts: Katie Robbert, Christopher S. Penn, John Wall — Trust Insights *So What?* live show - Referenced: Substack's adoption of Pangram; a Trust Insights paper on Zenodo (December) on Llama 3.2 bias; the BBQ bias benchmark; Trust Insights RAFT responsible-AI framework ## 📝 Summary **The occasion.** A reader accused the Trust Insights newsletter of being AI slop and the team of lying about their AI use. The timing coincided with Substack adopting **Pangram**, an AI detector. So they built one, on a laptop, and demonstrated it. **How detectors actually work.** Almost all current detectors run a language model that scores the **probability of each next word**. Long runs of high-probability words read as machine-generated. Penn's example: "I pledge allegiance to the ___" is near-certain to be *flag* and near-zero to be *rutabaga*. **Which produces the central defect.** Plenty of humans write in high-probability sequences — technical writers, students, anyone describing something plainly. "The sun is bright. The sun is yellow." That is human writing that scores as machine writing, because the detector measures predictability, not authorship. > [!quote] > They do not detect for quality. They do not detect for something being educational or entertaining or helpful. They only detect the likelihood of word order within a corpus of text. **Who this lands on.** Penn is specific: non-English speakers, neurodivergent writers "who write in ways that are not normal aka boring," and anyone trained on the five-paragraph essay. And the training corpora skew — toward the wealthy, the ethnic majority, and the more educated — so people outside those groups get flagged more. **The rule they have held for three years:** > [!quote] > you must never ever ever use an AI detector for anything that's punitive because they still have false accusation rates. Zero is the only acceptable rate when the consequence is expulsion, discipline, or job loss. **The clarifying finding.** A colleague's posts kept flagging as AI. She dictates, then has Claude clean them up. The mechanism: **AI editing rearranges words into a statistically more probable order.** Penn's illustration is Yoda — "For 800 years have I trained Jedi" versus "I have trained Jedi for 800 years." Identical words, ordinary order, and only the second flags. So detectors detect *editing*, not authorship. **The perverse consequence.** Katie's rough unedited draft scored as more clearly human than her polished version. Evading detection therefore means publishing worse writing — which nobody wants to read. **And they are defeatable.** Penn notes that benchmarking against 40-odd text metrics and instructing a model to hit them "basically invalidates the AI detector," because it mathematically defeats the surprisal measure. So the control fails against anyone sophisticated and catches the naive. **Their answer is disclosure, not detection** — Trust Insights' RAFT framework, with transparency as the core: if you used AI, say so, and say how. Penn's deluxe book editions ship the raw dictation recordings as provenance. **Bottom line.** Detectors measure statistical predictability and are routinely read as measuring honesty. They systematically misfire on the least-advantaged writers, are trivially evaded by the sophisticated, and are being wired into punitive decisions anyway. ## 📖 Main Content ### 🗒️ Notes **The trigger.** A strong reader response to the newsletter accusing it of being AI-written and the team of lying about their AI use. They cannot confirm a detector was used, but the timing lines up with Substack rolling out Pangram. Katie names an irony that runs underneath the whole episode: these models were trained on human writing — without consent — and are now used to tell humans their writing is machine-made. **John Wall's position on AI writing.** He does not much care. Most AI writing is slop and therefore uninteresting; AI is genuinely useful for gathering and structuring what is out there — "basically the promise of search that we got stolen from us." And he notes a market shift: *made without AI* is becoming a credibility and luxury signal. Katie's observation on that: 100% human-written was simply how it was for the entire history of writing, and is now a label you can apply and charge for. **The mechanism (the core of the episode).** Detectors run a custom-tuned model measuring next-word probability. Runs of expected words look machine-made. The failure follows immediately. Uncreative writing is legitimately high-probability. Technical writers have limited ways to say a thing. Students describing the sun will produce the same sentences. > [!quote] > for the majority of people who are like, I need to describe the sun. Let me say the things that I know about it. It's going to be a very uncreative, very basic pattern matching. But they wrote it, but now it's going to get flagged as AI. **Known model tells.** Negative parallelism ("It's not this, it's this") — common in both ChatGPT and Claude. Triadic rhythm, triplets — "Claude in particular loves triadic rhythm." Passive-voice openers that restate rather than assert, common in LinkedIn comments. And the structural one, which Penn calls a Claudism: **the pattern is the tell.** At profile level rather than post level, you see uniform sentence shape — short first line, long second, short third — repeating across every post. Katie's counterpoint is sharper than the shape analysis: she identified the sample as AI because it was *meaningless*. "The AI audit only works if it knows whose eyes are reading the message" is a sentence-shaped object that does not say anything. **Detector generations.** Early detectors used static features — sentence length, noun counts. Abandoned: false accusation rates too high. Current ones use language models. Lower, and not zero. > [!quote] > zero is the only acceptable rate when using for anything punitive **Who gets hurt.** Non-English speakers. Neurodivergent writers. Anyone writing to a template — and Penn singles out the five-paragraph essay, taught for decades, which is formulaic by design and flags accordingly. Katie's aside: it also does not teach you to write. **Building one.** Three components: a base model that returns log-probabilities; a server fast enough to run it, since this is compute-intensive; and a training corpus demonstrating what is and is not AI. Penn used **Skyfall** (BeaverAI, an upsample of Mistral, tuned for writing) and ran it on a MacBook. For the corpus he took posts from people he knows do not use AI — Ann Handley, Jay Acunzo, his own — then had Claude, ChatGPT, Meta, Perplexity, Qwen, Gemini and Copilot write synthetic posts on the same topics, paired one-to-one. He notes the opacity problem: Pangram, Originality, GPTZero, Winston — none disclose their base model or training corpus. Which is why results diverge wildly between them. **The Becca case, and the real mechanism.** A friend's posts kept flagging. She dictates, then has Claude tidy them. > [!quote] > when an AI tool edits, it rearranges the words back into a statistically probable pattern The Yoda demonstration: *"For 800 years have I trained Jedi. My own counsel will I keep on who is to be trained."* An AI rewrite yields *"I have trained Jedi for 800 years. I will keep my own counsel on who is to be trained."* Same words. Probable order. Only the second flags. **The demonstration.** The tool colours each word green (surprising) or red (expected) and reports a document score. Katie's newsletter scored *probably not AI*. A Qwen-written post on empathy showed large red blobs — what you look for is clustered runs, not isolated words. Then the finding that produces the conundrum: Katie's **unedited rough draft** scored as more clearly human than her published version. > [!quote] > if you were just write hot garbage and put it in, it will say it's definitely not AI and everyone else is like, this sucks So the incentive is to publish worse writing to prove you wrote it. **Katie's distinction, which is the most practically useful thing in the episode.** When John was her human editor, he fixed grammar and coherence — he did not rearrange her sentences. AI editing tools do, because they are trying to help. That is the difference, and it is controllable. Penn's mitigation: > [!quote] > You must operate only at the sentence level. You may rearrange the order of my sentences, but you may not reconstruct or edit the sentences themselves. And Katie's stricter version, which she used with Gemini: do not change anything, give me a list of suggested changes and I will decide. **Training-data bias.** Whose writing taught the detector what "human" looks like? His example: Korean-American prep-school students writing entirely unassisted, in imperfect English. That is 100% human and would likely not be represented in the corpus. > [!quote] > What typically models are trained on is they're trained on the easiest information available. And that tends to skew wealthy... more educated. And so people outside those communities, they get flagged more. **Katie's 5P dilemma.** She created the 5P framework and has written enough about it for two books. Using AI to assemble her own existing writing feels wrong to her, and would likely flag — but she wrote all of it. Where is the line? Penn's two answers. First, legitimate: break her corpus into paragraphs, tag them by category, and have AI select from her own sentences — preserving her exact word order, particularly for pre-2023 writing. Second, and he flags it as tricky territory: benchmark against 40-odd text metrics and instruct a model to hit them. > [!quote] > you basically invalidate the AI detector can no longer detect that it's AI because you've got a mechanism in place that mathematically defeats surprisal measures **Why anyone uses detectors.** Three cases: genuine deception (deepfakes, proof-of-life fraud); academic integrity, which he calls dangerous; and social status — "people who maybe need to feel better about themselves by yelling at other people who they think are using AI." **Legitimate objections to AI, stated seriously.** Ethically compromised companies; resource intensity in electricity and water; training data that was not ethically sourced; and a company shredding books as it scans them. Penn's position: if your objection is moral, then whether AI touched the artifact matters, and that is coherent. **Disclosure as the alternative.** Trust Insights' RAFT framework, with transparency central: disclose that you used AI and how. Penn's deluxe editions include the raw highway dictation recordings as provenance and lineage — sold as an upsell. > [!quote] > where I and a lot of other people have a problem is if you say something is yours and it isn't **Katie's reframe, and the episode's best exchange.** Hand someone a report, ask what is in it, and they do not know. She insists this predates AI — people on autopilot pulling numbers into a deck they cannot explain. > [!quote] > That's not an AI problem. AI is exacerbating the problem because it's taking out the step where the human actually has to pull the numbers and put the deck together. Which Penn compresses: > [!quote] > AI didn't create the problem of people not using their brains. AI amplified the problem of people not using their brains. And the organizational cause she names: demanding 10x productivity produces exactly the behaviour being complained about. **Practical mitigation.** Give a model the linguistic vocabulary for what to avoid: negative parallelism, anaphora abuse, triadic or tricolonic rhythm, uniform staccato, rhetorical Q&A ("the result? devastating" — "nobody asked you"), and sentences that do not advance meaning. Penn reports that instructing a model to cut any sentence not meaningfully advancing the narrative removed about 30% of the text. ## 📚 References - **Primary** — Penn's own detector built on **Skyfall** (BeaverAI, Mistral upsample, writing-tuned) - **Primary** — Trust Insights paper on Zenodo (December) on Llama 3.2 bias; the **BBQ** benchmark (Bias Benchmark for QA), on which Meta's models reportedly failed ~50% of the time - **Named detectors** — Pangram (adopted by Substack), Originality, GPTZero, Winston — none disclosing base model or corpus - **Corpus sources** — public posts by Ann Handley, Jay Acunzo, Penn; synthetic counterparts from Claude, ChatGPT, Meta, Perplexity, Qwen, Gemini, Copilot - **Framework** — Trust Insights RAFT (responsible AI framework) - **Not provided** — no accuracy figures for the built detector; no citation for the training-corpus-skew claim; the Llama/BBQ result is described rather than shown ## 🗣️ Paraphrase Last week a reader told us our newsletter was AI slop and that we were lying about how we use AI. We can't prove they used a detector, but Substack just rolled out Pangram, so the timing fits. So let's show you how these things actually work — including one we built. Almost all detectors work by measuring how predictable each next word is. If I say "I pledge allegiance to the," the next word is almost certainly "flag" and almost certainly not "rutabaga." Long stretches of expected words look like a machine wrote them. Here's the problem. Lots of people write in predictable sequences. Technical writers do, because there are only so many ways to say a thing. Students do. If you ask someone to describe the sun, you get "the sun is bright, the sun is yellow." A human wrote that, and it flags as AI. There are real tells. Negative parallelism — "it's not this, it's this." Triplets; Claude loves triplets. Passive openers that restate instead of saying anything. But the biggest one is the pattern across a whole profile: every post the same shape, short line, long line, short line. Katie's take was different: she knew it was AI because it was nonsense. Sentence-shaped, and meaning nothing. Older detectors used sentence length and word counts. Everyone abandoned that because it accused too many innocent people. Today's use language models, which is better and still not zero. And zero is the only acceptable number if you're going to expel someone or fire them. We've been saying that for three years: never use these for anything punitive. And they especially penalize non-native English speakers, neurodivergent people who write in ways that aren't "normal," and anyone taught the five-paragraph essay — which is formulaic by design. To build one you need a model that returns log-probabilities, a fast server, and a corpus showing what is and isn't AI. I used Skyfall, a Mistral upsample tuned for writing, on my MacBook. For the corpus I took posts from people I know don't use AI — Ann Handley, Jay Acunzo, me — and had seven different models write synthetic versions of the same topics. Then my friend Becca's posts kept flagging, and she does write them. Turns out she dictates and has Claude tidy them up. That's the whole thing: when AI edits, it rearranges your words into a more probable order. Yoda says "For 800 years have I trained Jedi." AI turns that into "I have trained Jedi for 800 years." Same words. Only the second flags. So we ran Katie's newsletter — probably not AI. Then her rough unedited draft — even more clearly not AI. Which is the trap: the worse your writing, the more human it looks. You could publish your raw draft to prove you wrote it, and nobody could read it. Katie pointed out the real difference: when John edited her, he fixed grammar and made sure sentences were coherent. He didn't rearrange them. AI tools do, because they're trying to help. So tell them not to: you may reorder my sentences, you may not rewrite them. Now, whose writing taught the detector what human looks like? If you took Korean-American students writing entirely unassisted in imperfect English — that's 100% human, and it's almost certainly not in there. These things get trained on whatever data is easiest, which skews wealthy, majority, educated. Everyone else gets flagged more. Katie asked about her 5P book: she's written enough about it for two books, so is it wrong to have AI assemble her own words? Two answers. Legitimately, tag her paragraphs and have AI pick from her own sentences, keeping her word order. Or — and this gets tricky — benchmark against forty-odd text metrics and tell a model to hit them, which mathematically defeats the detector entirely. Which is the point. These things measure a statistic. They don't measure quality, or honesty, or whether something is useful. If you object to AI on moral grounds — the companies, the water and power, the stolen training data, the company shredding books as it scans them — then yes, it matters whether AI touched it, and that's coherent. Our answer is disclosure: say you used AI, and say how. My deluxe editions ship the raw recordings of me dictating while driving, so you can see the lineage. The actual problem is claiming something is yours when it isn't. Which, as Katie says, isn't new — hand someone a report and ask what's in it and they don't know. AI didn't create that. It amplified it, by removing the step where you had to build the thing yourself. ## 📰 Newsletter **AI detectors measure predictability. Everyone reads them as measuring honesty.** Someone told Trust Insights their newsletter was AI slop and they were lying about it. Substack had just rolled out Pangram. So Chris Penn built his own detector on a MacBook and showed exactly what these things do. They score how predictable each next word is. "I pledge allegiance to the ___" — near-certain *flag*, near-zero *rutabaga*. Long runs of expected words read as machine-written. Which is the whole problem, because **plenty of humans write predictably.** Technical writers. Students. Anyone describing something plainly. "The sun is bright. The sun is yellow." Human, and it flags. > They do not detect for quality... They only detect the likelihood of word order within a corpus of text. The people it misfires on are named specifically: **non-native English speakers, neurodivergent writers, and anyone trained on the five-paragraph essay** — which is formulaic by design because school made it that way. And the corpora that taught these things what "human" looks like skew wealthy, majority, and educated. Everyone outside that gets flagged more. Hence the rule they've held for three years: **never use one for anything punitive.** Zero false accusations is the only acceptable rate when the outcome is expulsion or a job. Then the finding that reframes everything. A friend's posts kept flagging — and she does write them. She dictates, then has Claude tidy them up. > when an AI tool edits, it rearranges the words back into a statistically probable pattern The Yoda demo nails it. *"For 800 years have I trained Jedi"* → *"I have trained Jedi for 800 years."* **Identical words. Only the reordered one flags.** So detectors detect *editing*, not authorship. Which produces a genuinely perverse result: Katie Robbert's **unedited rough draft** scored as more clearly human than her polished version. Publishing worse writing is how you prove you wrote it. Her distinction is the practical fix. Her human editor fixed grammar and coherence — he **didn't rearrange her sentences.** AI tools do, because they're being helpful. So forbid it: > You must operate only at the sentence level. You may rearrange the order of my sentences, but you may not reconstruct or edit the sentences themselves. And the structural kicker: Penn notes you can benchmark against ~40 text metrics and instruct a model to hit them, which "basically invalidates the AI detector." **So it fails against anyone sophisticated and catches the naive.** A control that selects against the innocent is worse than no control. Their answer is disclosure rather than detection — say you used AI, say how. Penn's deluxe editions ship the raw recordings of him dictating while driving, as provenance. Best exchange in 46 minutes, though, is Katie's. Hand someone a report, ask what's in it, they don't know: > That's not an AI problem... AI is exacerbating the problem because it's taking out the step where the human actually has to pull the numbers. Penn's compression: **"AI didn't create the problem of people not using their brains. AI amplified the problem of people not using their brains."** And the organizational cause she names, which nobody wants to hear: demanding 10x output produces precisely the behaviour being complained about. ## ✅ Action Items **DevRel** - The sentence-level editing constraint is immediately reusable by anyone who publishes and uses AI assistance. It preserves voice and, incidentally, avoids the detector problem. - The vocabulary of tells (negative parallelism, triadic rhythm, anaphora abuse, rhetorical Q&A) is a practical editing checklist independent of detection. **OSPO** - If any contribution or review process is considering AI detection, the never-punitive rule and the demographic misfire pattern should be on the table before adoption. **Community** - Codes of conduct and contribution policies should specify **disclosure** rather than detection. Detection is unreliable, evadable, and discriminatory; disclosure is checkable and enforceable socially. **WordPress subteams** - **Community / Training** — Directly relevant if AI-authored contributions become a policy question. The equity argument matters in a project with a large non-native-English contributor base. - **Polyglots** — **Highest-relevance team here.** Non-native English writing is systematically flagged by these tools. Any project-wide AI-detection policy would fall hardest on translators and international contributors. - **Documentation / Marketing** — The disclosure framing (say you used AI, say how) is a publishable standard. - **Core / Core Performance / Core AI / Core Program Team / Mobile / Playground / Support / Design / Photos / TV / Accessibility / Meta / CLI / Hosting / Tide / Openverse / Themes / Plugins** — Not applicable to this source. **The WPOCC** - Worth a stated position: disclosure over detection, with the equity reasoning attached. **FAIR** — Not applicable to this source. **Personal / vault** - The sentence-level editing rule is directly applicable to anything published from this vault. - The provenance-as-upsell idea (shipping raw recordings) is an interesting model for demonstrating lineage without argument. ## 🧪 Hypotheses 1. **False positive rates are higher for non-native English writers** — directly testable with matched corpora, and the single most consequential untested claim here. 2. **AI editing alone, without AI authorship, is sufficient to trigger detection.** The Becca case is one instance; testable at scale by running human-written text through editing passes. 3. **Rough drafts score as more human than polished drafts by the same author** — the perverse-incentive claim, easily tested. 4. **Detectors can be defeated by metric-targeted generation**, meaning detection selects against the unsophisticated. 5. **Detector training corpora underrepresent non-standard English**, testable if any vendor disclosed their corpus — which none do. 6. **Restricting AI editing to sentence-level reordering preserves detector scores** relative to full rewriting. 7. **Different detectors disagree substantially on the same text**, which would follow from undisclosed and differing base models. ## 📊 Methodology **What they did.** Built a working detector and demonstrated it live on real texts, including their own published and unpublished writing. For a marketing livestream this is unusually strong — the mechanism is shown rather than described, and the demonstration includes a case that complicates their own position (Katie's newsletter, which used AI assistance, scoring as probably-not-AI). **Genuine strengths.** The paired corpus design is sound: same topics, human and synthetic versions, one-to-one. The Becca case is real diagnostic work — an anomaly investigated until the mechanism was found, rather than dismissed. And they demonstrate on their own writing rather than on a convenient example. **Limits.** No accuracy figures are reported for the built detector — no precision, recall, or false-positive rate, which is precisely the metric the episode's central argument is about. The "known non-AI users" corpus rests on trust that those authors do not use AI. Sample sizes are small. And the demographic-skew claim, which carries most of the ethical weight, is reasoned from how training corpora are generally assembled rather than measured. **Conflict of interest.** The episode originates in a personal grievance — being accused of AI authorship — which gives the hosts a stake in concluding that detectors are unreliable. They are transparent about the trigger, and the technical argument stands independently, but the motivation is worth holding. ## ❓ Questions - What is the actual false-positive rate of any commercial detector on non-native English? Everything ethical here depends on a number nobody has published. - If detectors are defeatable by metric targeting, what exactly is the remaining use case beyond intimidation? - Where is the line on assembling your own prior writing with AI? Katie asks directly and the answer is procedural rather than principled. - How much do commercial detectors disagree with one another on identical text? - Does the sentence-level constraint actually preserve scores, or merely reduce the effect? - If Substack is running Pangram at platform scale, what happens to writers misflagged — is there appeal, and does anyone see the score? - Is there a detection approach that measures *provenance* rather than *probability*? Disclosure and lineage are proposed; cryptographic or workflow-based provenance is not discussed. ## 🔮 Future Research - **Demographic false-positive audit** of commercial detectors — the most important missing work in this space. - **Editing versus authorship as detection triggers**, isolating the two. - **Detector agreement studies** across vendors on identical corpora. - **Provenance systems** as an alternative to statistical detection — signed drafts, edit lineage, dictation artifacts. - **The polish/detectability tradeoff** — quantifying how much quality a writer sacrifices to appear human. - **Effects of platform-scale detection** on writer behaviour and on which voices get suppressed. ## 💡 Implications **For policy — the load-bearing implication.** A detector with a non-zero false accusation rate must never drive a punitive decision, and current detectors all have one. Worse, the errors are not randomly distributed: they concentrate on non-native speakers, neurodivergent writers, and the formulaically trained. An institution adopting one is adopting a discriminatory instrument whether or not it intends to. **For practice.** If you use AI assistance and want to keep your voice, constrain it to sentence-level reordering. It preserves your word order, which preserves both your voice and your detector score — and the first reason is the better one. **For anyone being accused.** Understanding the mechanism is the defence. The tool measures predictability, and predictability is a property of clear plain writing as much as of machine writing. That is a fact about the instrument, not about you. **For the equity argument specifically.** This should be the centre of gravity in any institutional conversation, and it usually is not. The people most likely to be falsely accused are the least able to contest it — students, second-language writers, people already navigating institutions that doubt them. The tool inverts the burden of proof and places it on exactly those least equipped to carry it. **For the deeper problem.** Katie's reframe is the durable one. The genuine issue is people presenting work they cannot explain, and that predates AI by decades. Detection targets a proxy; comprehension checks target the thing. Ask someone what is in their report. ## 🧠 Learning Styles Assessment - **Visual** — **Strong, and the episode's distinguishing feature.** The colour-coded word-probability visualization makes an abstract statistical concept immediately legible; red blobs versus interspersed green is a genuinely good explanatory artifact. - **Aural** — Strong. Three-person conversation, works largely as audio, though the demonstrations lose substance. - **Verbal** — Very strong. The Yoda example and the pledge-of-allegiance example are efficient teaching devices, and the linguistic vocabulary is given precisely. - **Kinesthetic** — Moderate to strong. The sentence-level prompt constraint is copy-pasteable and immediately usable. - **Logical** — Strong. Mechanism → predicted failure mode → demonstrated failure case → mitigation. The Becca anomaly is followed to its cause rather than waved away. - **Social** — Very strong. Three participants who genuinely disagree — Katie's "it's nonsense" reading versus Chris's structural analysis is a real difference, and productive. - **Solitary** — Moderate. 46 minutes with digressions. - **Naturalistic** — Not applicable to this source. **Reading level estimate.** Roughly Grade 12–14. Conversational but concept-dense: log-probabilities, surprisal, negative parallelism, anaphora, tricolon. Transcript quality is poor on names (Robbert, Acunzo, Pangram, Nemotron elsewhere) and mangles "sycophancy," "rutabaga," and "Zenodo." ## 📊 Accessibility Assessment Mapped to WCAG 2.2 AA. 1. **Audio description (SC 1.2.5, AA) — fails, and materially.** The core evidence is a colour-coded visualization described deictically ("you can see there are lots and lots of red," "there's a blob here"). A blind listener cannot access the episode's central demonstration. 2. **Use of colour (SC 1.4.1, A) — fails as presented.** The entire visualization encodes its meaning in red versus green — the most common colour-vision deficiency pairing — with no secondary encoding described. This is a tool the presenter built, so it is a fixable design choice, not an inherited constraint. 3. **Captions (SC 1.2.2, A) — at risk.** AI-generated captions mangle the proper nouns that carry the actionable content. 4. **Contrast / text in images (SC 1.4.3, AA) — unverified.** Small on-screen text, repeatedly zoomed during the demo. 5. **Sensory characteristics (SC 1.3.3, A) — fails.** Heavy positional and colour reference throughout. 6. **Reading level (SC 3.1.5, AAA — aspirational).** Dense unglossed linguistic terminology. **Summary.** The red/green encoding is the notable failure, and it is ironic given the episode's own argument about tools that disadvantage particular groups. Adding a shape, underline or numeric annotation alongside colour would cost little. ## 🌍 DEIB Assessment - **This is the episode's strongest dimension by a distance, and it is substantive rather than gestural.** The equity critique is not an aside — it is the central argument against punitive use. Penn names non-English speakers, neurodivergent writers, and the formulaically educated as the populations these tools misfire on, and explains *why* through training-corpus composition rather than asserting bias. - **Cultural breadth.** The Korean-American prep-school example is well chosen: unambiguously human writing that a detector would likely flag, illustrating the mechanism rather than merely gesturing at it. - **Neurodiversity.** Explicitly named — writers "who write in ways that are not normal aka boring" — and the phrasing does real work by identifying normativity as the problem rather than the writer. - **Class and access.** The claim that corpora skew wealthy, majority and educated because that data is easiest to obtain is a structural argument about how datasets get built. - **Inclusive language.** Good throughout. Katie challenges Chris substantively and is treated as the senior voice on the framework question, which is her own. - **Representation.** Three hosts; two men and one woman, all apparently American. The affected populations are discussed and not present. - **Psychological safety.** Strong. Katie discloses that she used AI assistance on the newsletter that was attacked, and shares her rough draft on air. Modelling that disclosure under accusation is the behaviour the episode advocates. - **The unexamined bit.** The episode notes that "no AI used" is becoming a luxury signal and does not follow it — that market rewards those who can afford slower production, which is its own access problem. ## ✅ Logical Validity **Strong reasoning** - **The mechanism-to-failure derivation is airtight.** If detectors measure predictability, then predictable human writing must flag. That is deductive, not empirical, and it is the argument's spine. - **The Becca case is genuine diagnostic work** — an anomaly pursued to a cause that reframes the whole category from detecting authorship to detecting editing. - **The Yoda example is an excellent minimal pair** — identical words, different order, different result. It isolates the variable cleanly. - **They demonstrate on their own writing**, including a piece that used AI assistance and a rough draft that undercuts the polish they sell. - **Legitimate anti-AI positions are steelmanned**, not dismissed: ethics, resource use, training-data provenance, book destruction. Penn explicitly grants that a moral objection makes the question matter. - **Katie's reframe is the strongest analytical move** — relocating the problem from AI to comprehension, and identifying the 10x-productivity demand as the organizational cause. - **Uncertainty is marked** on the trigger: they cannot confirm the accuser used a detector and say so. **Fallacies** 1. **Motivated reasoning risk, disclosed but unmitigated.** *Location: framing.* The episode exists because they were accused. That gives them a stake in the conclusion. They are transparent about it, and no attempt is made to steelman the pro-detection position beyond three brief use cases. 2. **No accuracy figures for their own detector.** *Location: demonstration.* The entire argument concerns false-positive rates, and they report none for the tool they built. The demonstration is qualitative where the argument is quantitative. 3. **Anecdotal generalization.** *Location: Becca case, Katie's drafts.* Two cases carry the editing-triggers-detection claim. 4. **Unsourced structural claim.** *Location: training-corpus skew.* The demographic argument — which carries the ethical weight — is reasoned from how corpora are typically built, not measured. Plausible and unevidenced. 5. **Trust-based corpus construction.** *Location: detector build.* "People I know don't use AI" is unverifiable, and it is the ground truth the whole classifier rests on. 6. **Ad hominem, mild.** *Location: use-cases segment.* Attributing detector use to people who "need to feel better about themselves" characterizes motives rather than addressing the practice — even if some instances fit. **Tone of voice.** Conversational, occasionally aggrieved given the trigger, and notably self-implicating — Katie volunteers her AI use and shows her rough draft. Commercial content is present and light (Slack community, podcast, courses) and confined to asides. The register is practitioner-to-practitioner with genuine disagreement between hosts, which is rarer and more valuable than agreement. ## 🛡️ DARVO Analysis - **Deny** — **Present in a limited and legitimate form.** The episode responds to an accusation (that they lied about AI use) by disputing it. That is an ordinary defence against a specific charge, not DARVO's deflective denial — and they concede the substantive point by disclosing exactly how AI was used. - **Attack** — **Present, mildly.** The accuser is characterized indirectly via the "people who need to feel better about themselves by yelling at others" line. This is the element closest to DARVO, and it is a single aside in a 46-minute episode whose substance is technical. - **Reverse victim and offender** — **Largely absent, with a caveat.** The framing does position the accused as the wronged party, but the episode does not rest there: it grants that moral objections to AI are coherent, discloses its own AI use, and concedes the underlying concern about people claiming work they cannot explain. **Assessment.** Not DARVO. The structure is present in outline — accused party denies, characterizes accuser unflatteringly — but the deflective function is missing, because the substantive claim is conceded and disclosed rather than dodged. Worth recording precisely because the surface pattern resembles it while the mechanism does not. ## 🎯 Conclusion **Bottom line.** AI detectors measure how predictable a text is. They are read as measuring whether someone cheated. Those are different things, and the gap between them lands hardest on people who write plainly, write in a second language, write differently for neurological reasons, or were taught a formula in school. The clarifying finding is that detection triggers on **editing** rather than authorship. The Yoda pair proves it: identical words, ordinary order, only the reordered version flags. Which means a person who dictates their own thoughts and has a tool tidy them gets flagged, and a person who generates text and then roughens it does not. That inversion is the whole indictment. Penn notes you can defeat these tools by benchmarking against text metrics — so the control fails against anyone sophisticated and catches the naive. A control that selects against the innocent is worse than no control, and it is being wired into punitive decisions at platform scale. Their answer — disclosure over detection, with lineage as evidence — is the right shape, because it is checkable and does not require an unreliable instrument. What is missing is numbers. The episode's argument is entirely about false-positive rates and reports none, including for the detector they built. The demographic-skew claim carries most of the ethical weight and is reasoned rather than measured. Both are testable, and someone should test them. And Katie's reframe is the part that will outlast the tooling: the real problem is people presenting work they cannot explain, which long predates AI. Detection chases a proxy. Asking someone what is in their report addresses the thing. **Next steps.** 1. Adopt the sentence-level editing constraint for anything published — it preserves voice, and the detector benefit is a side effect. 2. If any policy conversation touches AI detection, lead with the never-punitive rule and the demographic misfire, not with accuracy. 3. Prefer disclosure standards over detection tooling, and specify *how* AI was used rather than whether. 4. Ask people what is in their work. That check survives every change in the tooling. ## 🧱 Refactor Appendix ### 📚 Literature Split Pad - Purpose: stage Konik CEE units that become per-source Literature notes. - Split level: **H3**. ### [H1] Detectors measure predictability and are read as measuring honesty > [!quote] > They do not detect for quality... They only detect the likelihood of word order within a corpus of text. **Where:** mechanism and 5P segments. **Why it matters:** the entire category rests on a category error, and naming it is the defence for anyone accused. `#xref` `#articleseed` ### [H1] AI editing rearranges words into probable order — that is what gets detected > [!quote] > when an AI tool edits, it rearranges the words back into a statistically probable pattern **Where:** Becca case; the Yoda minimal pair. **Why it matters:** reframes the category from authorship detection to editing detection, and yields a concrete mitigation. `#xref` ### [H1] The misfire is not random — it concentrates on the least advantaged writers **Where:** who-gets-penalized and training-corpus segments. **Why it matters:** non-native speakers, neurodivergent writers, and the formulaically taught, with a mechanism (corpora built from the easiest data, skewing wealthy and majority) rather than an assertion. `#bmf` `#share` ### [H2] The control is evadable, so it selects against the naive > [!quote] > you basically invalidate the AI detector... because you've got a mechanism in place that mathematically defeats surprisal measures **Where:** 40-metrics segment. **Why it matters:** a defence that experts defeat and innocents fail is worse than none. `#xref` ### [H2] Evading detection requires publishing worse writing **Where:** rough-draft demonstration. **Why it matters:** a perverse incentive created by the tool — proving authorship costs readability. `#addendum` ### [H3] Constrain AI editing to sentence-level reordering > [!quote] > You must operate only at the sentence level. You may rearrange the order of my sentences, but you may not reconstruct or edit the sentences themselves. **Where:** mitigation segment. **Why it matters:** preserves voice; the detector benefit is incidental. Immediately usable. `#xref` ### [H3] Disclosure and lineage beat detection **Where:** RAFT segment; the deluxe-edition dictation recordings. **Why it matters:** provenance is checkable where statistical detection is not. `#xref` ### 🔖 Atomic Split Pad #### [Z1] A control that the sophisticated evade and the naive fail is worse than no control - Scope/condition: holds where evasion is cheap for the informed and the penalty falls on the accused; breaks where evasion is costly or detectable in itself. - Pointer: 40-metrics segment plus the demographic misfire. - **CREATE candidate.** Check against [[3. Resources/Permanent Notes/A Defense Experts Fail Is Not a Defense for Users]] — closest, and possibly the *inverse* framing (that note: experts defeat it, users bear the risk; this note: experts defeat it and the innocent are punished by it). Verify whether it absorbs this before creating. #### [Z2] Editing toward the probable is what detection detects, not authorship - Pointer: Becca case; Yoda pair. - **CREATE candidate**, lower confidence. Nearest: [[3. Resources/Permanent Notes/Shared AI Tools Homogenize Output While Feeling Original to Each User]] — same homogenization mechanism, different consequence. Check whether that note covers it. #### [Z3] AI amplified the problem of people not using their brains - Pointer: Katie's report exchange; Penn's compression. - **CONNECT, do not create.** Captured by [[3. Resources/Permanent Notes/AI Amplifies Existing Bottlenecks Rather Than Creating Capability]]. This is a strong second instance and belongs in that note's Evidence, along with the 10x-productivity organizational cause. #### [Z4] Classifier bias follows corpus convenience - Pointer: training-corpus-skew segment. - **CONNECT** to [[3. Resources/Permanent Notes/Sampling a Fraction of a Source Produces False Authority Over the Whole]] and [[3. Resources/Permanent Notes/AI Quality Gaps Are Policy Problems Not Capability Problems]]. #### [Z5] Provenance beats inference for establishing authorship - Pointer: RAFT disclosure; dictation recordings as lineage. - **CONNECT** to [[3. Resources/Permanent Notes/An Audit Trail Is Trustworthy Only When It Is a By-Product of the Work]] — raw dictation is exactly a by-product record, and to [[3. Resources/Permanent Notes/A Revision Is Legitimate in Proportion to Its Auditability]]. ### ✏️ Fleeting Queue - [F1] Adopt the sentence-level editing constraint in publishing prompts. - [F2] Find any published false-positive rates for commercial detectors on non-native English — the whole ethical argument needs this number. - [F3] Test the rough-draft-scores-more-human claim on own writing. - [F4] Check whether [[3. Resources/Permanent Notes/A Defense Experts Fail Is Not a Defense for Users]] already covers the evadable-control claim. - [F5] The red/green-only visualization is a colour-vision accessibility failure in a tool arguing about disadvantaged groups — worth noting if ever demoed. - [F6] Chase the Trust Insights Zenodo paper on Llama 3.2 / BBQ bias. - [F7] "No AI used" as a luxury signal rewards those who can afford slower production — unexamined access problem worth its own note. <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/So What How AI Detectors Work.md|So What? How AI Detectors Work by Katie Robbert, Christopher S. Penn, John Wall]] ## Highlights - PENN: AI detectors especially penalize non-English speakers. It penalizes people who are neurodivergent, who write in ways that are not normal, aka boring. It penalizes highly templated text. Anyone who remembers school remembers the joy of the five paragraph essay. You say what the paper's about, you have three paragraphs of exposition, you have the conclusion which restates the key point. At that point you may as well just have ChatGPT write it for you, and a lot of students are. But that formulaic template is going to get flagged as AI because it is so strict. ROBBERT: That also doesn't teach you how to write. ([View Highlight](https://read.readwise.io/read/01kz3y2s3c6hatadwtj0gq0r3b)) - Note: The misfire is not random — it concentrates on non-native speakers, neurodivergent writers, and the formulaically taught. Combined with the never-punitive rule, this is the load-bearing policy argument. Highest relevance for Polyglots / any project with a large non-native-English contributor base. - Tags: equity, false-positives, atom/literature/h1 - PENN: But the other way, and this is where you start getting into very tricky territory, is there are 40-some-odd different text metrics that you can use to take a known sample of your exact writing style when it's known good, and you give that to an AI tool as a benchmark and you say you must meet these scores, rewrite until you do. And what happens when you do that is you basically invalidate the AI detector. It can no longer detect that it's AI because you've got a mechanism in place that mathematically defeats surprisal measures. ROBBERT: All to say that AI detectors are a big red flag. ([View Highlight](https://read.readwise.io/read/01kz3y35hsawsd9q7x0z2fs5c7)) - Note: PERMANENT CANDIDATE — a control that the sophisticated evade and the naive fail is worse than no control. Detection selects against the innocent. Check whether "A Defense Experts Fail Is Not a Defense for Users" already covers this or is the inverse framing. - Tags: atom/atomic/z1, evasion, security-controls - ROBBERT: But that's not a new problem. AI didn't create that problem. A lot of people in general are just sort of on autopilot. They'll follow the process, they'll put the numbers together, they create the dashboard, they create the deck. And you say, okay great, tell me about it. Well, I don't know what's in it, I just pulled the numbers and put it in. That's not an AI problem. AI is exacerbating the problem because it's taking out the step where the human actually has to pull the numbers and put the deck together. So basically you're paying someone to push a button and still not know what's in it. The problem is the education around the information, the data. I want people who can critically think about the process that they're doing. PENN: And so when it comes to the use of AI detectors, the detector then becomes a part of the problem rather than part of the solution. Part of the solution is you're not encouraging people to think. You're not asking them to think critically, which is basically metacognition — think about your thinking. You're instead saying I want to see 10x productivity out of you, when they're like, I'm going to use AI for everything because you want 10x more stuff coming out of my desk. ([View Highlight](https://read.readwise.io/read/01kz3y3kyrst34bb6ccd87g2cb)) - Note: Robbert's reframe is the durable part: the real problem is people presenting work they can't explain, which predates AI by decades. Detection chases a proxy; asking what's in the report addresses the thing. Plus the organizational cause — demanding 10x output produces the behaviour being complained about. CONNECT: AI Amplifies Existing Bottlenecks. - Tags: comprehension-debt, metacognition, atom/atomic/z3 ## Links - [[3. Resources/Permanent Notes/A Detector the Sophisticated Evade and the Naive Fail Is Worse Than No Detector]] — evidence-for: style probability produces evasion and punitive false-positive risk. - [[3. Resources/Permanent Notes/A Safety Control That Cannot See Severity Becomes a Hazard]] — applies-in: classifier uncertainty cannot support high-consequence punishment.