/anti-detection-writing

Detector research: use and limits

Verified September 6, 2026. Use these findings to assess an editing hypothesis, not to promise a detector outcome. Refresh dated claims when the service changes.

Later dated sections retain the subsequent research and measurements. The September 8 technical-format diagnostic compared one freshly scanned baseline with a reference card and mini-FAQ on the same exposed ETag case. All three passed separate quality review and stayed at 100% AI / 0% mixed / 0% human in Basic 4.9b. The three scans used 394 word credits; there was no local model run or repair loop. This supports reporting a bounded no-gain diagnostic, not a new detector-reduction instruction, a measured cost-efficiency gain, or an impossibility claim.

• GPTZero says it stopped using perplexity and burstiness for detection in autumn 2023. Sentence-length variation and unusual vocabulary are not documented controls for its current score. Historical thresholds farther down that support page are not current instructions. Official support
• Its August 10, 2026 product report says release 4.8b began masking detected headers from the classification features on August 1. That report describes other surrounding material, such as bibliographies, as future work rather than an implemented universal exclusion. It also distinguishes importance-based Advanced highlighting from ordinary sentence classes. This is vendor-reported behavior, not a reproduction of our Basic 4.9b scans. Header edits are therefore a weak new hypothesis; do not infer that changing Markdown or an arbitrary highlighted span will improve a current result. Product report
• Its 2026 report describes supervised document/sentence classification and training against paraphrasing, plus whitespace cleaning and probability remapping. Table 1 identifies version 4.1b; our September 6 UI scans displayed 4.9b. Architecture and hyperparameters remain proprietary. The report does not establish what caused an individual current-service result. A sentence's predicted class and its contribution to a document prediction are different quantities, so a highlight alone is not a causal editing rule. Report, §§3.2–3.5, §4 and Table 1
• DetectGPT, Fast-DetectGPT, and Binoculars evaluate model probabilities through specific algorithms. Their results do not validate a natural-language editing prompt or transfer automatically to another detector. Inspect the actual intervention and tested setting before adapting a paper's lesson. DetectGPT, Fast-DetectGPT, Binoculars
• Adversarial Paraphrasing uses detector-guided token generation and evaluates eight open detectors, without GPTZero. Its aggregate detection rates are not individual-document confidences or a current GPTZero pass rate. Copying its prompt omits the implemented mechanism. Paper, authors' implementation
• Check a paper's quality rubric before borrowing its success criterion. That implementation permits minor factual changes in rating 4, so pooling ratings 4 and 5 does not establish exact claim preservation. Keep material factual failures separate from prose preference; a favorable average rating cannot override changed advice. Quality rubric

Additional research checked September 6, 2026

These papers do not identify a GPTZero version matching our Basic 4.9b observations. Read each outcome in its own units: mean class probability, classification rate, and a single document's confidence are different measurements.

When considering a released paraphraser, pin its base model, adapter and inference settings. Check input fit instead of silently truncating. Retain the raw generation separately from the finished candidate: analysis, repeated drafts and incomplete endings must not contaminate the detector input. Independently check exact claims and protected spans before spending scans. A rejection is an unmeasured quality failure, not a zero detector score. A small local feasibility test is not a paper reproduction or evidence that the ordinary skill acquired the model's behavior.

Our September 6 local pilot ran one pass of the released HIP 0.6B adapter on two reused technical answers. Both failed separate factual review: the Kubernetes reply lost a necessary health-condition distinction, and the PostgreSQL reply reversed snapshot timing and added invalid syntax. Original command payloads and URLs survived, illustrating why literal-presence checks alone are insufficient. Neither output was scanned. These failures apply to the two saved outputs, not all HIP sizes or domains; the model was not adopted by the skill.

Distinguish an implemented detector algorithm, a tested editing intervention, and a proposed explanation. Contextual rewriting and an authentic supplied writing sample can be useful editorial choices; a detector benefit remains a separate hypothesis. Apply the experiment and plateau guidance in measurement.md. Research plausibility alone is not transfer evidence, and these research notes are not a validated score-reduction update.

Reader trust and the editing boundary

Checked September 6, 2026 after reading Bryan Cantrill's “The revolt of the reader”. The essay argues for author responsibility and using an LLM as an editor after the author has developed the piece. Its praise of Pangram reflects personal experience, not a controlled GPTZero comparison.

• The underlying survey collected 668 voluntary, anonymous responses through social media, with no guaranteed demographics. Its 78% stopping-reading and 98% original-writing preference figures describe stated reactions to suspected AI involvement. They are not measured detection accuracy or a blinded comparison of edited texts.
• Oxide's policy distinguishes public and personal writing from operational documents, where verified accuracy and completeness can justify generation. Its Pangram publication requirement is an organizational policy, not a requirement inherited by this skill or other communities.
• The Pangram 4 report, §§3.1, 5.5 and 5.9 reports 44 false negatives among 10,223 documents rewritten with BLADER instructions. Its commercial-humanizer result of 98.83% counts detection as AI or mixed; it is not a confidence score. The polished-human-text benchmark starts from human writing, so avoiding a fully-AI classification there does not establish successful rewriting of AI originals. The report excludes short conversational replies under 50 words and acknowledges errors and context sensitivity. These are developer-reported Pangram results, not our reproduction or GPTZero transfer evidence.
• Vincent Bernat's editing account links a human draft and subsequent copyediting revisions. It offers inspectable editing history, without detector measurements or a controlled reader-preference test.

These sources support testing restrained editing without adding another general style rule. A useful hypothesis is to diagnose specific clarity problems before revising only the affected passages. Compare that workflow with the current skill on the same drafts; retain factual checks and separate readability from detector outcomes. Keep human-origin and AI-origin cases separate: preserving an author's supplied voice does not demonstrate removing the signature of a generated draft. No new scan or validated score reduction accompanies this research note.

Drafting methods checked September 7, 2026

Separating facts from prose is a testable workflow, not an established detector fix. Step-by-Step separates symbolic planning from trained text realization and audits omitted, wrong and added relations. Graph-to-text evaluation supplies counterevidence: zero-shot changes to graph linearization and triplet expression did not improve its results; stronger PlanGTG gains required training. DiscoFuse shows why sentence fusion needs checks for altered connectives and ambiguous references. None establishes a current GPTZero or Pangram reduction from an ordinary editing instruction.

The local studies/writing-methods-1 trial compared three procedures with the current skill on two new synthetic drafts. Relation-map reconstruction failed one factual review because the preparer invented a copying actor. Core expansion passed quality but was not selected or scanned. Topic continuity won the frozen simplicity tie-break, but preserved both originals unchanged; it did not exercise sentence-topic editing. All readability comparisons tied. Pangram 4.0 returned AI Generated with 100% displayed AI text for both complete pairs, with short-text warnings. The separately fixed technical GPTZero Basic 4.9b pair returned 100% AI, 0% mixed, 0% human on both sides. Confirmation cases remained unused. These observations compare preservation with rewriting, not the effect of edits that never occurred.

Keep populated fact-table fields source-supported. A passive source statement with no actor must not acquire one merely to complete the table or prefer active voice. Audit both omissions and additions before drafting; checking that every original claim appears is insufficient. Retain unselected methods as unmeasured, not detector failures. No new score-reduction method was adopted from this trial.

A separate adaptive follow-up then scanned the pre-existing, quality-passing core-expansion technical draft against a fresh baseline. Both Pangram 4.0 results remained AI Generated with 100% displayed AI text and short-text warnings. This is one reused development pair, with no GPTZero or confirmation result for that method; the business core-expansion draft remains unmeasured. The original round and follow-up are recorded separately.

Reader dialogue checked September 7, 2026

STORM, NAACL 2024, §§3.2, 5.2 and 6 uses responsive questions during research before outlining and writing. Its conversation ablation and editor assessments motivate testing the preparation process, but concern retrieved evidence and long articles. They establish neither a benefit from a two-question adaptation on fixed evidence nor a commercial detector reduction. The paper also identifies unsupported connections between facts as a failure mode. The authors' conversation loop was inspected on September 7, not executed or reproduced.

The local studies/reader-dialogue-1 pilot compared the current skill with a reader-dialogue workflow on one new synthetic decision memo. Both writers received the same source contract and article request. A separate reader asked an initial question, read the writer's analytical answer, and asked one follow-up before composition. The first candidate preserved the facts but was judged slightly less clear than the baseline. After the single pre-score clarity repair allowed by the protocol, a fresh masked review found both final articles accurate and equally clear across all 20 required claims. The first candidate remains saved and unscanned; the measured treatment includes the dialogue, extra calls/context and review-informed repair.

Both exact reviewed texts received Pangram 4.0 AI Generated with 100% displayed AI text, high segment confidence and no short-text warning. The completed pair showed no reduction. GPTZero and confirmation cases were not measured. This observation does not isolate article length, the dialogue or the repair, and it does not demonstrate a readability advantage over the baseline. No new default writing rule was adopted.

Retain the first-pass judgment when a repair is needed: a passing final revision does not erase an initial quality disadvantage or identify which preparation step helped. Record responsive exchanges and their effect on the final explanation before claiming the proposed method was exercised. A longer-case result without a short-text warning is useful additional evidence, but cannot diagnose why earlier short cases received the same score.

Released models and local tooling checked September 7, 2026

The revised StealthRL paper reports low detection rates alongside poor judged quality and similarity. Its released local implementation can run a PEFT adapter, but an unscored single-output path assigns a placeholder zero. Neither that value nor the paper's open-detector results establish a current GPTZero/Pangram reduction. Retain unmeasured values as null and require independent factual review of raw outputs.

The released AuthorMist model provides another learned rewriting mechanism. A single generation using its card prompt is not the paper's candidate-selection experiment. Its model and tokenizer advertise different context limits; use the conservative model limit, account for the complete prompt plus output allowance, and decode continuation tokens directly instead of splitting a text marker. Record the base model's research license separately from the derived card's MIT label. No weights are bundled with this skill.

The new local detector CLI exercises a pinned open Desklib model and rejects overlength inputs instead of silently truncating. Its first eight-text diagnostic returned AI probabilities below 0.5 for every known AI-origin development text; seven had prior Pangram AI Generated/100% AI-text observations. An independent implementation matched the CLI's baseline logit and tokenization exactly. The model also scored its publisher's AI-labelled example highly. This supports numerical implementation correspondence while revealing a weak proxy for our cases; it is not a general accuracy estimate or a calibrated ranking for the commercial endpoint. Do not treat lower local scores as GPTZero/Pangram gains. The complete diagnostic is retained in studies/local-tools-1.

The same local study exercised AuthorMist on two reused development originals using the published plain-completion prompt, one completion each, fixed sampling settings and MPS bfloat16. Both ended on EOS but failed separate exact-source review. The HTTP answer lost a required header and caching qualifications; the business update changed staffing and rollout decisions. Both roughly doubled in length and added editing commentary. No repair, candidate selection or detector scan followed. This is a two-output feasibility failure, not a general verdict on the model or a reproduction of its paper. The raw outputs, finish reasons and rejection record are retained; the model is not adopted.

Bounded model-assisted search checked September 8, 2026

The later twenty-method study tested a different complete workflow: pinned AuthorMist generation, a bounded factual-repair pass and a separate purpose/fluency pass. Its business development example changed from 100% to 0% GPTZero AI confidence, and a fresh process plus fresh final editing repeated that change. The HTTP example remained at 100%. All measured results used Basic Scan 4.9b; intensity zero disabled contrastive decoding. This is a model-and-editing result, not evidence that a Markdown instruction reproduces the trained model or that any one component caused the decrease.

Both previously sealed cases preserved every required fact, but the independent reviewer preferred the incumbent's readability. The confirmation package was therefore left unscanned and the model workflow was not adopted. Sentence movement had retained stiff noun phrases and dense instructions: factual repair needs a final reading for those defects, without claiming that such a check will lower a detector score. The full record retains all twenty attempt dispositions, unsuccessful results and the corrected derived decision.

The technical plateau does not prove that technical prose cannot improve. Nor does a 100% confidence result identify its cause or mean that the answer is incorrect. The official perplexity/burstiness support update was rechecked on September 8; its current-architecture qualification remains in place. Avoid treating old perplexity recipes, local proxy scores or sentence highlights as a diagnosed cause. Use the cost-conscious comparison guidance for a specific next hypothesis.

Register and demonstration methods checked September 9, 2026

Register Analysis for Arbitrary Style Transfer compares source and reference, derives descriptors, then rewrites. Its style-classifier and automatic meaning measures motivate testing reference-guided preparation; they do not establish exact claim preservation or a GPTZero decrease. Our two-stage adaptation used a licensed professional explanation and one justified operation per case. The article preserved its claims but lost the readability comparison. After a separately declared paragraph repair, both full measured pairs remained 100% AI in Basic 4.9b. The study retains the formal rejection and additional diagnostic separately.

AdParaphrase tests preferred and non-improving revision demonstrations on Japanese advertisements, with human paraphrase and attractiveness judgments. It reports no detector experiment. A small set of English editing demonstrations would be a new adaptation requiring examples and selection criteria fixed before target scores; no such set was tested here. A separate choice between two incumbent drafts was tested and rejected for a temporal qualification error. Do not attribute evidence from trained search or Japanese advertising to that procedure.

Preserve the time scope of claims during evidence restructuring. A date on a measurement must not narrow a limitation that still applies. Review paragraph density after moving accurate sentences: grouping intervention, results, screenshot scope and confounds can make the explanation harder to follow. These observations justify explicit editorial checks, without implying lower detector confidence.

Full-article generation checked September 9, 2026

The later article transfer study distinguishes a custom forward loop from native generation using the pinned AuthorMist configuration. Their prompt format, sampling parameters and sampling device differed; any behavior difference is bundled. Both runs failed, and a separately declared recovery of the native output lost its final readability comparison. No detector scan followed. Loading a model card's parameters is insufficient unless the executed decoder applies the corresponding processors; record actual behavior and raw failure evidence.

A public PyTorch MPS attention issue motivated a small numerical check. Six seeded synthetic attention shapes showed no nonfinite values or large discrepancies against an explicit reference. This did not exercise complete model generation, and the native sampler's later invalid-probability exception did not identify its cause. Neither observation justifies calling an attention change a verified fix or claiming a speedup. The study also retains the caveat that its saved native protocol did not explicitly state the stricter EOS eligibility rule later attributed to it.

Sequential context checked September 9, 2026

DIPPER, Appendix C, compares contextual paraphrasing and multi-sentence windows using automatic quality measures. Its inference implementation appends each output to the next window's prefix. This motivates a continuity experiment, but its trained T5 model and literary training data differ from an instruction-following writer. GPTZero's February 2026 report also explicitly includes DIPPER outputs in detector training; historical bypass results cannot establish current performance.

Our five-section adaptation preserved the claims but lost a slight readability preference and remained unscanned. A missing study-date detail in one writer's packet creates an evidence-availability confound, alongside different call counts and contexts. Preserve relevant qualifiers from supporting sources when constructing partial packets. Neither this rejection nor the paper establishes a current GPTZero benefit from windowing.

Paired demonstrations checked September 9, 2026

The later paired-example trial tested three fixed synthetic English original/revision/judgment examples, including one accurate but less useful revision. This adapts AdParaphrase's conditioning idea without reproducing its human annotations, language, domain or example counts. The candidate also adds a brief diagnosis instruction, so those components are not isolated.

The reviewer slightly preferred both complete candidates, but the HTTP and article prose pairs each remained 100% AI / 0% mixed / 0% human in Basic 4.9b. Prose scope, code exclusions and full-message review were fixed before scanning. The round supplies no detector-reduction rule; confirmation stayed locked. A follow-up check found that the released HIP adapter had already failed factual review in the earlier two-output pilot, so it was not treated as an untried model or rerun merely to establish feasibility.

Vocabulary review checked September 9, 2026

The phrase review reference records a contextual checklist from GPTZero's October 2024 vocabulary article and the limits stated in its vocabulary FAQ. Frequency associations do not establish a causal score-reduction method. A separate offline check found none of the ten listed entries or the article's named variants in any of the four prose inputs above. This hypothesis therefore supplied no edit for those cases; no new scan or reduction is claimed. Keep the checklist as editorial guidance, not a validated detector intervention.

A broader source review added corpus observations from COLING 2025 and biomedical abstract research, separately labeled from vendor and GitHub editing heuristics. A BEA 2025 paper evaluates custom vocabulary classifiers on student and generated essays, with substantial generator dependence; it does not test GPTZero scores after editing. The expanded reference therefore preserves technical meanings, attribution and uncertainty instead of turning association lists into universal bans.

Exact crossover checked September 10, 2026

Language Model Crossover distinguishes generated offspring from ordinary text splicing. Its classical crossover implementation targets image prompts; its sentence experiment uses different generation and evaluation machinery. Neither establishes a GPTZero benefit for professional prose.

Our exact-crossover diagnostic tested a new constrained adaptation: two complete incumbent drafts, independently reviewed matching claim boundaries, exact reciprocal splicing, and separate selection within the parent and hybrid pools. The comparator was the selected parent, not a single incumbent call. The final masked reviewer slightly preferred the selected hybrid, but both fresh prose-only Basic 4.9b scans returned 100% AI / 0% mixed / 0% human, with no short-text warning. Two confirmation cases remained sealed. This supplies no detector-reduction rule; exact inheritance prevents unrecorded rewriting but does not itself establish semantic compatibility or lower confidence.