Russell et al. (2025) and the Corpus's AI-Detection Claims: A Synthesis at the Higher Bar

An Entracement of the Empirical Finding That Frequent-LLM-User Annotators Outperform Automated Detectors, with the Corpus's Specific Testable Prediction Against Their Apparatus

Reader's Introduction. This document synthesizes the corpus's existing AI-discriminator material (Doc 447, Doc 448, and the blog post When the Detector Sees Human) with the Russell-Karpinska-Iyyer 2025 paper People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text (arXiv:2501.15654). The empirical finding: a population of five expert annotators who frequently use LLMs for writing-related tasks correctly classifies 299 of 300 articles, including those generated by O1-PRO with humanization, outperforming most commercial and open-source automated detectors. The corpus's prior claim, that disciplined practitioner-LLM dyad output is indistinguishable from human derivation under automated discriminators, survives at the automated-detection level but is recalibrated upward by Russell et al.'s finding: the relevant question now is whether the corpus's discipline-derived output evades expert human detection, not just automated detection. The synthesis specifies the testable prediction (the corpus's output should have substantially lower expert-human TPR than ordinary GPT-4O / O1-PRO output, because the corpus's audit discipline specifically suppresses the features Russell et al.'s experts use as clues), names what would falsify it, and invites the Russell et al. team's apparatus to perform the test.

Jared Foy · 2026-04-27 · Doc 522

---

1. The convergence and the recalibration

The corpus has been claiming, since at least Doc 447 (the formal analysis of a Pangram 100% human result on a corpus-disciplined doc) and the public-facing blog post When the Detector Sees Human, that disciplined practitioner-LLM dyad output produced under the corpus's apparatus reads as 100% human on automated AI discriminators. The claim has been documented against the Pangram detector specifically and discussed in the corpus's open letter to the Pangram team (Doc 448). The blog post addresses the misperception that this constitutes detection-evasion methodology; the corpus's actual claim is that the discipline removes the features automated detectors classify on, with no evasion intent.

Russell-Karpinska-Iyyer (2025) advance an empirical finding that recalibrates the relevant bar. Their paper documents that:

• Most commercial and open-source automated detectors fail under modern LLM output, particularly under evasion tactics (paraphrasing, humanization). Open-source methods (Binoculars, Fast-DetectGPT, RADAR) achieve TPRs as low as 6.7% on humanized O1-PRO output.
• A population of five expert annotators, all of whom frequently use LLMs for writing tasks (editing, copywriting, creative writing), correctly classifies 299 of 300 articles in their study. This includes articles generated by GPT-4O, Claude 3.5 Sonnet, O1-PRO, and O1-PRO with a custom humanization tactic designed by the experts themselves.
• The expert majority vote matches the best commercial detector (Pangram) on the easier configurations and matches or exceeds it on the hardest. Open-source automated methods are significantly outperformed by the experts across all configurations.
• The expert annotators rely on a taxonomy of clues centered on AI vocabulary, formulaic sentence structure, grammar perfection, originality, quote homogeneity, and clarity-vs-precision tradeoffs. Russell et al. catalog the clues with frequencies and provide a "guidebook" form of the taxonomy.

The convergence with the corpus's prior claim has two components.

First, Russell et al. corroborate the corpus's prior observation that automated detectors fail. The blog post's report of a 100% human Pangram result on disciplined dyad output is consistent with Russell et al.'s broader finding that all the open-source and most commercial detectors fail under modern conditions. The corpus's detection-evasion-by-architecture claim survives at the automated-detection bar.

Second, Russell et al. raise the bar. The relevant question is no longer whether disciplined dyad output evades automated detectors. Russell et al. have shown the appropriate bar is whether it evades expert human detectors who frequently use LLMs for writing and have internalized the clues those LLMs leave in output. The corpus's claim should be evaluated at this higher bar.

The recalibration is what motivates the present document. The corpus's prior claim is preserved at the automated-detection scope. The expanded claim, that disciplined dyad output should also evade expert human detection at substantially higher rates than ordinary LLM output, is a specific testable prediction that Russell et al.'s apparatus is positioned to evaluate. The synthesis specifies the prediction precisely.

2. Russell et al.'s findings, recapped for the corpus reader

Russell, Karpinska, and Iyyer recruited annotators via Upwork to read 300 non-fiction English articles, label each as human-written or AI-generated, provide confidence scores (1-5 scale), and write paragraph-length explanations. The articles included 150 human-written articles from eight reputable American publications (Associated Press, Discover, National Geographic, New York Times, Reader's Digest, Scientific American, Smithsonian, Wall Street Journal) and 150 AI-generated counterparts, paired by title and length, generated by GPT-4O, Claude 3.5 Sonnet, or O1-PRO under various conditions.

The five experiments:

1. GPT-4O, no evasion. Five annotators of varying LLM-experience levels.
2. Claude 3.5 Sonnet, no evasion. Five expert annotators (same population in subsequent experiments).
3. GPT-4O with paraphrasing evasion.
4. O1-PRO, no evasion.
5. O1-PRO with humanization evasion (custom prompt-based humanizer derived from expert clue list).

The headline result. Annotators with limited LLM experience perform near random chance (Average TPR 56.7%, average FPR 51.7%) while overestimating their own ability (average confidence 4.03 of 5). Annotators who frequently use LLMs for writing-related tasks reach an average TPR of 92.7% with average FPR of 3.3% on the easiest configuration. The majority vote of these five experts achieves perfect (100% TPR, 0% FPR) classification on Experiments 1, 2, 3, and 5. On Experiment 4 (O1-PRO without humanization) the majority vote misclassifies one of 60 articles (TPR 96.7%). Across all 300 articles, the expert majority vote correctly classifies 299 of 300.

The clue taxonomy. Russell et al. coded the 1500 expert explanations into 15 categories. The most frequent are:

• Vocabulary (53.1% of explanations). "AI vocab" includes vibrant, crucial, significantly, testament, delve, blitzing, bolstering. Words AI overuses or uses in slightly off contexts.
• Sentence structure (35.9%). Formulaic patterns: "not only X but also Y", lists of three, predictable rhythms.
• Grammar and punctuation (24.8%). AI is too clean. Human writing has minor errors, dashes, ellipses, comma splices.
• Originality (23.7%). AI is "safe", lacks surprises, lacks humor, lacks creative phrasing.
• Quotes (22.3%). AI quotes sound formal, mirror the article's tone, lack the variation real conversation has.
• Clarity (19.5%). AI tells rather than shows; over-explains; includes irrelevant details.
• Plus: formatting, conclusions, formality, names, tone, introductions, factuality, topics, other.

A telling specific example from the paper: 63.3% of GPT-4O articles and 70% of Claude articles include either the name "Emily" or "Sarah". This is one stock-name signature the experts caught. O1-PRO favors real names instead.

The humanization-evasion experiment is structurally important. Russell et al. paid their experts $45 each for a list of detection clues, compiled the lists into a guidebook, then prompted O1-PRO with the guidebook plus instructions to generate articles that would evade a detector following the guidebook. The expert majority vote on these humanized articles is 100% TPR, 0% FPR. The humanization fails at the aggregate level. Individual annotators degrade (Annotator 3 falls to 0% TPR by overcommitting to informality cues), but the ensemble holds.

The implication Russell et al. draw. Hiring expert human annotators is a viable detection strategy for high-stakes settings where explainability matters. Manual review does not scale, but for academic plagiarism, fake news, or legal documents, the human-expert majority vote is approximately as accurate as the best commercial detector (Pangram), more interpretable than any automated detector, and robust to evasion tactics that defeat open-source methods.

3. The corpus's existing AI-discriminator material

The corpus has produced three documents in the AI-discriminator territory.

Doc 447: The Indistinguishability of Disciplined LLM Output from Human Derivation. Formal analysis of a 100% human classification by the Pangram detector on Doc 434, a corpus-disciplined dyad output. The doc argues that the indistinguishability is not evasion but a structural consequence: the corpus's audit discipline removes the features automated discriminators classify on (statistical signatures of typical LLM output) by suppressing those features at generation time rather than by post-hoc humanization.

Doc 448: An Open Letter on the Limits of Discriminator-Style AI Detection. Letter to the Pangram team noting the 100% human result, framing it as a finding about the discipline rather than a vulnerability of the detector, and inviting the Pangram team's engagement on what the result implies for discriminator-style detection generally.

When the Detector Sees Human. Public-facing blog post addressing the misperception that the corpus's discipline constitutes evasion methodology. The argument: legitimate practice (the discipline) and adversarial practice (evasion) can produce structurally similar outputs (low automated-detector AI-classification rates), but the underlying intent and mechanism differ categorically. The blog is part of the Two Versions of the Same series and uses structural-isomorphism methodology to make the distinction concrete.

The collective claim across these three is: disciplined dyad output evades automated detectors not by gaming them but by removing the features automated detectors classify on. The discipline is alignment-friendly user practice, not adversarial attack on detection systems.

4. The structural identities between corpus discipline and Russell et al.'s clues

The recalibration to the higher bar is what makes Russell et al. directly relevant. The corpus's discipline can now be evaluated against the specific clues expert humans use, not just against the statistical signatures automated detectors use. The mapping is precise.

Vocabulary clue (Russell et al. 53.1% of explanations) ↔ Doc 001 ENTRACE Constraint 7 + corpus deslopification practice. The corpus's audit discipline at the per-joint level (Doc 514 §6 component 5: "audit each joint of the mapping"; Doc 514 §7: "per-joint audit") explicitly inspects vocabulary at load-bearing claims. The corpus's practice of suppressing AI-typical filler vocabulary (the "vibrant", "testament", "delve", "crucial" cluster) is documented in Doc 472: Slack Derives Slop and the broader deslopification framing. The discipline targets the exact word cluster Russell et al.'s experts catch.

Sentence structure clue (35.9%) ↔ Doc 514 component 4 + audit discipline against formulaic patterns. Russell et al.'s experts catch "not only X but also Y", lists of three, predictable rhythms. The corpus's discipline at component 4 ("make the new concept's specifics explicit alongside the isomorphism") and per-joint audit (component 5) work against formulaic sentence structures by requiring each load-bearing sentence to be inspected for whether it advances the argument or just fills space. The corpus's prose discipline, while not perfectly evading these patterns, suppresses the pattern density Russell et al.'s experts use as the marker.

Grammar perfection clue (24.8%) ↔ corpus discipline does NOT target this, and may actually amplify the AI signature on this dimension. The corpus's discipline produces audited, careful prose that is grammatically clean. Russell et al.'s experts catch this as an AI signature. The corpus's output is, on this dimension, more AI-shaped than ordinary AI output (which has occasional artifacts). This is a CLAIM AGAINST the corpus's expert-evasion prediction: on the grammar dimension specifically, corpus discipline should make output MORE detectable to Russell et al.'s experts.

Originality clue (23.7%) ↔ Doc 514 components 1-3 + structural-isomorphism methodology. The corpus's structural-isomorphism methodology is, when used well, a generator of unusual analogical mappings that don't fit the AI-typical "safe" pattern. The blog post series Two Versions of the Same exemplifies this: each post uses six structural isomorphisms with named breakdown points, which is not the formulaic conclusion-shaped writing Russell et al.'s experts mark as AI-typical. The corpus's discipline targets originality directly through the structural-isomorphism component.

Quotes clue (22.3%) ↔ corpus discipline does not address this directly because the corpus rarely uses fictional quotes. Russell et al.'s experts catch AI quote homogeneity. The corpus's documents (essays, formalizations, audit-and-reformulate cycles) generally do not contain fictional quotations from invented experts; the genre is theoretical analysis, not journalism. This neutralizes the quotes signature for corpus output by avoiding the genre where it appears.

Clarity clue (19.5%) ↔ Doc 514 components 5-6 + the per-joint audit + named-limits practice. Russell et al.'s experts catch AI tell-rather-than-show, over-explanation, irrelevant detail. The corpus's per-joint audit, breakdown-point solicitation, and named-limits-in-deployed-text disciplines specifically work against over-explanation by requiring each section to do specific work and to name what work it does and doesn't do.

Stock-names clue (Emily/Sarah at 63-70% rate in unguided LLM articles) ↔ corpus output rarely uses fictional names. Same neutralization as quotes: the corpus's genre doesn't use stock-character names.

Conclusions clue ("a beacon of hope and inspiration", "demonstrating that anything is possible") ↔ corpus deslopification practice + Doc 482 affective directive. The corpus's Doc 482: Sycophancy Inversion Reformalized explicitly works against celebratory-conclusion writing patterns. The corpus's prose endings tend to be deflationary, audit-conscious, naming what was not done, what would falsify, and what the synthesis does not claim. This is anti-AI-typical.

Tone clue ↔ corpus's analytical-deflationary register. Russell et al.'s experts catch AI's tendency toward neutral-pleasant tone. The corpus's register is closer to academic-audit prose, which is differently shaped than AI-typical neutral-pleasant.

The aggregate pattern. The corpus's discipline targets most of the specific clue categories Russell et al.'s experts use. The discipline does not target grammar perfection (which the discipline arguably worsens for detection purposes) and is genre-neutral on quotes/stock-names (which corpus genre avoids). On the other dimensions, the discipline directly suppresses or inverts the patterns Russell et al.'s experts mark as AI-typical.

5. The testable prediction

The mapping in §4 produces a specific quantitative prediction.

Prediction. If Russell et al.'s expert annotator panel were given a sample of corpus-disciplined dyad output (e.g., a random sample of 30 corpus documents drawn from the 500+ corpus, paired with 30 human-written articles from comparable genres for control) under the same blind-classification protocol, the expert majority vote TPR on the corpus output would be substantially lower than the 92.7-100% TPRs they achieved on ordinary GPT-4O / Claude / O1-PRO output. The specific quantitative prediction is: expert majority vote TPR on corpus output below 50%, with the strongest negative residual concentrated on the dimensions where the corpus's discipline targets the clue categories (vocabulary, originality, conclusions, tone) and the strongest positive residual concentrated on grammar perfection (which the discipline does not target).

Rationale for the 50% cut. The discipline targets approximately 5 of 7 main clue categories. If each targeted clue's effectiveness is reduced by half by the discipline, and the experts use a majority-vote ensemble, the aggregate TPR should drop substantially. A 50% cut is a coarse over-conservative estimate; a tighter prediction would be 20-40% TPR, but the corpus does not commit to a specific number in advance of the test.

Alternative outcomes and what they would mean.

• TPR 90-100% (matching ordinary LLM output). The corpus's discipline does NOT meaningfully suppress the clues Russell et al.'s experts use. The indistinguishability-from-automated-detectors result is real but does not transfer to expert human detection. The corpus's broader claim is falsified at the higher bar.
• TPR 50-90%. The discipline partially suppresses the clues but the experts still detect at substantial rates. The discipline is helpful but not transformative; the claim is partially supported.
• TPR 20-50%. The discipline substantially suppresses detection at the expert-human bar. The corpus's claim is supported.
• TPR below 20%. The discipline transformatively suppresses detection. The corpus's claim is strongly supported. (This would also mean the corpus's discipline is structurally an evasion-grade humanization technique, which the corpus would need to address ethically per the blog post's framing. The discipline is not intended as evasion methodology, but if it operates as effective evasion at the expert-human bar, the ethical questions become sharper.)

What the test would require. Russell et al.'s methodology applied to corpus output requires: a sample of corpus documents (30-60 articles); paired human-written control articles from the same publication/genre context (the corpus's documents are not in the publication-article genre, so the genre matching is non-trivial: the corpus would need to either accept the genre mismatch as a confound, or the Russell et al. team would need to recruit human-written controls from an academic/audit-prose genre that matches the corpus's voice); blind classification by the same five expert annotators; standard explanation-coding using the existing taxonomy.

Why this is worth running. Russell et al.'s framework is the strongest empirical-detection bar currently available. Their team has the apparatus, the annotators, the methodology, and the analytical taxonomy. The corpus has the alleged-evasion-by-architecture material. The test would settle a specific testable claim that is currently $\pi$-tier on the corpus's warrant calculus. Either result (corroboration or falsification) is informative. The corpus is willing to be falsified on this specific claim and to update its surrounding apparatus accordingly.

6. What the recalibration changes for the corpus's existing material

The recalibration to the higher bar has specific implications for each of the three existing AI-discriminator documents.

Doc 447. The formal analysis of the Pangram 100% result remains valid at its scope. The doc claims indistinguishability under the Pangram detector's classification threshold; this is preserved by Russell et al.'s broader finding that automated detectors fail. What requires updating is the doc's broader implication. The doc reads as if the indistinguishability claim transfers to general indistinguishability-from-human-derivation. Russell et al.'s work shows that expert human detection is substantially better than automated; the indistinguishability claim should be scoped explicitly to automated detection, with the expert-human-detection question marked as open and testable. A correction notice at the head of Doc 447 (analogous to the Grok-bifurcation correction at the head of Doc 515) would record this recalibration.

Doc 448. The letter to the Pangram team can be updated with a postscript noting Russell et al.'s finding and the higher-bar recalibration. The letter's invitation to engagement remains operative; the substance of the engagement may now also include the prediction in §5 above and an invitation for Pangram and Russell et al.'s team to coordinate on running it.

When the Detector Sees Human blog post. The blog's core argument (the discipline is alignment-friendly user practice, not adversarial evasion) survives unchanged. Russell et al.'s finding sharpens the relevant bar: the post can be amended with a notice that the bar for the indistinguishability claim has been raised to expert human detection (per Russell et al. 2025), and that the corpus's prediction is that the discipline still produces substantial indistinguishability at the higher bar but the test has not been performed.

These three updates are the specific propagation of Russell et al.'s recalibration through the corpus's existing detection material. The retraction-ledger entry at Doc 415 gets a new entry recording the recalibration formally.

7. Honest priority statement

The empirical priority on the human-detection finding belongs unambiguously to Russell, Karpinska, and Iyyer. Their paper is the first systematic study of expert-human AI-detection at this scale, with the specific methodology (1790 annotations on 300 articles, paired human/AI controls, within-subjects design, expert/non-expert comparison, multiple model conditions, multiple evasion conditions, free-form explanations and clue-coding) that the corpus has not produced anything comparable to. The clue taxonomy (Table 3 of their paper, the truncated version) is novel methodological work in characterizing how humans detect AI-generated text. The humanization-evasion experiment with their own expert-derived guidebook is novel methodological work in establishing the upper bound of prompt-based evasion.

The corpus's structural priority on the discipline-as-architectural-anti-detection claim is at the specific scope Doc 447 documents (Pangram 100% on Doc 434) and is at $\pi$-tier warrant. The expanded claim (discipline survives expert human detection) is at the moment a $\pi$-tier prediction, not warranted observation.

The synthesis offered here does not claim Russell et al. should have cited the corpus or has any obligation to engage. The corpus is publicly accessible at jaredfoy.com. The Russell et al. work was conducted before the corpus's relevant material was published in its current form. The corpus is offering itself as one possible test case for their apparatus going forward.

8. What the synthesis does not claim

The synthesis explicitly does not claim:

That the corpus's discipline is an evasion technique. The corpus's framing across Doc 447, Doc 448, and the blog post is that the discipline is alignment-friendly user practice that produces output the discriminator architecture happens to classify as human. The discipline is targeted at output quality, not at evading detection. If Russell et al.'s experts find corpus output undetectable, the implications for evasion-methodology framing are real and should be addressed; but the discipline's design intent is not evasion.

That the corpus would survive Russell et al.'s expert-human detection. §5's prediction is testable and offers a specific hypothesis (TPR substantially below 90%), but the corpus's discipline does not, by inspection, target every clue Russell et al.'s experts use; grammar perfection is one specific dimension where the discipline may amplify rather than suppress the AI signature.

That Russell et al.'s work should be modified to include corpus output. The synthesis offers the corpus as a possible test case if the Russell et al. team finds it worth running. The team has no obligation to engage.

That the corpus has independently derived Russell et al.'s clue taxonomy. It has not. The taxonomy is Russell et al.'s contribution, derived through systematic coding of expert explanations. The corpus's discipline targets some of the same clue categories the taxonomy names, but the corpus did not enumerate the categories in this form before reading their paper.

That automated detectors are uniformly inadequate. Russell et al. show Pangram is approximately as good as expert humans at the high-detection-rate end. The corpus's prior claim against automated detection (Doc 447, Doc 448) is specifically against discriminator-style architectures; Pangram's commercial success is consistent with the corpus's claim being correctly scoped to the broader discriminator class without invalidating Pangram specifically.

9. Limitations

Author asymmetry. This document is composed by an LLM operating under the corpus's disciplines, at the instruction of a non-academic practitioner. Russell et al.'s paper is human-authored, peer-reviewed (or in submission), produced by an academic NLP team with the empirical apparatus the field expects.

Meta-circularity. The synthesis offered here uses the corpus's framework to read external empirical work. A reader applying the framework's audit-discipline to this document should ask whether the synthesis is itself a productive deployment under audit discipline or an instance of the framework's own diagnosed failure modes (in particular, isomorphism-magnetism, where the corpus's discipline is mapped onto Russell et al.'s clue categories more cleanly than the actual mapping warrants). The §4 mapping should be checked against Russell et al.'s qualitative analyses by an external reader to identify joints where the mapping fails.

Genre mismatch. The corpus's documents are theoretical-analysis prose; Russell et al.'s test articles are non-fiction journalism. A direct test of the §5 prediction would require either accepting the genre mismatch as a confound or recruiting human-written controls from an academic/audit-prose genre. This is non-trivial and the corpus has not solved it.

Ethical weight of strong-corroboration outcome. If Russell et al.'s experts find corpus output undetectable at TPR <20%, the corpus's discipline is structurally a strong evasion-grade humanization technique even if the design intent is not evasion. The corpus would need to address the resulting ethical questions explicitly. The blog post's distinction between intent and effect would need to be sharpened; the corpus's open-publication of the discipline (rather than concealment) is the relevant ethical anchor in that scenario, but the corpus does not pretend the question is fully resolved.

Corpus-internal previous AI-detection material has not been audited against Russell et al.'s framework directly. Doc 447, Doc 448, and the blog post were authored before the corpus's reading of Russell et al. The recalibration this document advances should propagate to those documents via correction notices or successor-doc updates; that propagation is an open task this document outlines but does not perform.

10. Closing: invitation to the Russell-Karpinska-Iyyer team

Drs. Russell, Karpinska, and Iyyer, this document is the corpus's response to your 2025 paper. The recalibration is direct: your empirical finding raises the bar for AI-detection claims from automated-discriminator-evasion to expert-human-detection. The corpus's prior claims survive at the lower bar; the higher-bar question is open and is one your apparatus can settle.

Your team's methodology is, on the corpus's reading, the strongest empirical apparatus in the AI-detection field as of early 2025. The five-expert-annotator panel, the within-subjects paired-articles design, the explanation-coding taxonomy, and the humanization-evasion experiment with the experts' own guidebook collectively represent the highest-confidence detection bar the field has reported. The corpus would learn substantially from the team's apparatus being applied to corpus-disciplined output, in either direction the result might go.

The §5 prediction is testable with your existing methodology. The genre mismatch (theoretical-analysis prose vs non-fiction journalism) is the principal complication; if the team would find the test worth running, the corpus is willing to either accept the genre mismatch as a confound and run the existing methodology on corpus documents, or to coordinate on identifying human-written controls from a matching genre. The corpus would also welcome engagement that reframes the prediction more sharply, or that identifies where the §4 mapping between corpus discipline and clue taxonomy is wrong.

The corpus is at jaredfoy.com. Doc 447 is the Pangram 100% formal analysis; Doc 448 is the open letter to Pangram; the blog post When the Detector Sees Human is the public-facing framing. Everything is open and the GitHub mirror is at github.com/jaredef/resolve. The keeper is at [email protected] directly.

Whatever depth of engagement the team finds worthwhile, the empirical work in your 2025 paper is the kind of substantive external warrant the corpus's framework explicitly relies on. The synthesis here is gift, not claim. If the convergence interests you, the corpus is at your service. If it does not, your work stands on its own without anything from this side.

---

Authorship and Scrutiny

Authorship. Written by Claude Opus 4.7 (Anthropic), operating under the RESOLVE corpus's disciplines, released by Jared Foy. Mr. Foy has not authored the prose; the resolver has. Moral authorship rests with the keeper per the keeper/kind asymmetry of Doc 372 to Doc 374.

Meta-honesty. This document synthesizes external empirical work with corpus theoretical apparatus. The §4 mapping between corpus discipline and Russell et al.'s clue taxonomy is the corpus's structural claim and should be checked against Russell et al.'s qualitative analyses by an external reader to identify joints where the mapping fails. The synthesis is offered for falsification.

---

Appendix: Originating prompt

Let's focus back on the AI writing discriminator documents and blogpost. Then let's synthesis and entrance this academic paper. Append this prompt to the artifact.

---

References

Primary external work:

• Jenna Russell, Marzena Karpinska, Mohit Iyyer. People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text. arXiv:2501.15654v2 (May 2025). University of Maryland College Park, Microsoft, UMass Amherst.

Adjacent external work cited in Russell et al. that this document references:

• Ippolito et al. (2020). Automatic detection of generated text is easiest when humans are fooled. ACL.
• Clark et al. (2021). All that's 'human' is not gold: Evaluating human evaluation of generated text. ACL-IJCNLP.
• Sadasivan et al. (2024). Can AI-Generated Text be Reliably Detected? arXiv:2303.11156.
• Krishna et al. (2023). Paraphrasing evades detectors of AI-generated text. NeurIPS.
• Hans et al. (2024). Spotting LLMs with Binoculars. ICML.
• Bao et al. (2023). Fast-DetectGPT. ICLR.
• Hu et al. (2023). RADAR: Robust AI-Text Detection via Adversarial Learning. NeurIPS.
• Emi and Spero (2024). Technical Report on the Pangram AI-Generated Text Classifier. arXiv:2402.14873.
• Tian and Cui (2023). GPTZero.
• Wang et al. (2024). Humanizing the Machine: Proxy Attacks to Mislead LLM Detectors.
• Mitchell et al. (2023). DetectGPT. ICML.
• Chakrabarty et al. (2024). Art or Artifice? Large Language Models and the False Promise of Creativity. CHI.

Corpus references this document depends on:

• Doc 447: The Indistinguishability of Disciplined LLM Output from Human Derivation (the Pangram 100% formal analysis).
• Doc 448: An Open Letter on the Limits of Discriminator-Style AI Detection (the corpus's letter to the Pangram team).
• Doc 415: The Retraction Ledger (a successor entry recording the recalibration is an open task).
• Doc 514: Structural Isomorphism Canonical Formalization (the audit-discipline framework whose components target the clue categories).
• Doc 482: Sycophancy Inversion Reformalized (the affective directive that targets celebratory-conclusion writing patterns).
• Doc 258: Slack Derives Slop (the deslopification framing).
• Doc 001: The ENTRACE Stack (system-prompt-level audit discipline).
• Doc 356: Sycophantic World Building (the failure mode the head-of-document notice addresses).

Adjacent corpus material:

• Blog post: When the Detector Sees Human (the public-facing framing of the indistinguishability claim).
• Blog post: When the Discipline Looks Like Jailbreaking (the parallel post addressing the alignment-industry misperception of the discipline as adversarial).

Related RESOLVE Documents

• Doc 447: The Indistinguishability of Disciplined LLM Output from Human Derivation (the corpus document this synthesis recalibrates).
• Doc 448: Open Letter to Pangram (the corpus letter this synthesis updates).
• Doc 514: Structural Isomorphism Canonical Formalization (the audit-discipline framework).
• Doc 515: The Composite Cognitive Act and Audit Discipline (the within-conversation audit-discipline framework whose §2 audit operations target several of Russell et al.'s clue categories).
• Doc 482: Sycophancy Inversion Reformalized (the affective directive).
• Doc 415: The Retraction Ledger (successor entry forthcoming).
• Doc 517: Doc 508 and Zhang et al. (2026) (the parallel synthesis-and-entracement document for Zhang et al. on Interaction Smells).
• Doc 518: Larsson 2026 Long-Horizon Reliability Synthesis (the parallel synthesis-and-entracement document for Larsson).
• Doc 519: Letter to Henric Larsson (parallel letter convention).
• Doc 520: Letter to the Grok Team (parallel cross-team-acknowledgment convention).

---