Writer AI content detector

Vendor Writer, Inc. Official site writer.com/ai-content-detector How this card is sourced

AIDI-D-0023

Updated ·

Writer's AI content detector classified 27 of 54 documents correctly in Weber-Wulff 2023, a 50% binary accuracy that ranked 10th of 14 tools (arXiv 2306.15666, Table 5, read 9 November 2026). The vendor is Writer, Inc. of San Francisco, and the tool holds a card in the detector index although its product page no longer exists: the page announced a sunset on December 22, 2025, and the address has returned a redirect to writer.com since 24 December 2025 (archived detector page, captured 9 December 2025, read 9 November 2026). How this card is sourced applies the same fourteen fields to a withdrawn tool as to a live one, and the studies that tested it remain in the record.

What Writer detected

The Writer register entry: field, value, source and read date
FieldValueSource
Content types detectedTextweb.archive.org/web/20251209115654/https:...
Generators coveredChatGPT, GPT-4 (page title: ChatGPT detector and AI checker for GPT-4)web.archive.org/web/20251209115654/https:...
Input limitsWeb tool: pasted text or a URL, up to 5,000 words per check (archived detector page, 9 December 2025). API: POST /v1/tools/ai-detect took a single text input and returned a label (fake or real) with a score from 0 to 1; the legacy enterprise endpoint /content/organization/{organizationId}/detect is marked deprecated. The tool was sunset on December 22, 2025 and the page now redirects to writer.com.web.archive.org/web/20251209115654/https:...
Free tierFree web tool, up to 5,000 words per check, no account stated (archived page, 9 December 2025); withdrawn on December 22, 2025, so no free tier is available at the read dateweb.archive.org/web/20251209115654/https:...
Data retentionPrivacy Policy (effective April 17, 2026): We will only retain your information for as long as we need it to fulfil the purposes for which we collected it. This includes the purposes of our legitimate business interests and satisfying our legal, contractual, or reporting requirements. The same policy states that in no event will we use any data you provide to us when you access or use our Services to train our models, including any inputs or outputs. The archived detector page carried no retention statement of its own.writer.com/legal/privacy
OwnershipWriter, Inc., 111 Maiden Lane, Fourth Floor, San Francisco, CA 94108, USA (privacy policy effective April 17, 2026); terms of use governed by California law. The detector page announced that the tool would be sunset on December 22, 2025 with the AI detection API endpoint (archived capture of 9 December 2025); since 24 December 2025 the page returns a 301 to writer.com and the vendor's platform pages no longer mention AI detection.writer.com/legal/privacy
Independent studies5 recordsStudies index

Writer's detector scored pasted text or a URL and returned a percentage labelled "detection score", split into "AI-generated content" and "human-generated content". The archived page describes the tool as a "free AI detector to check up to 5,000 words", with a page title naming ChatGPT and GPT-4 as the targets (archived detector page, read 9 November 2026). The modality is text only; no image, audio or code detection was offered, and no language list or count was published. The vendor's disclaimer states that "our Writer product is better at detecting English text than other languages" (archived disclaimer, revision January 2024, read 9 November 2026).

The same disclaimer places the tool outside production use: "The prediction we give is just for showing how our product works. It's not meant to be used as a real product." It adds that Writer does not recommend "using Writer (or any detector) without a human watching over it" (archived disclaimer, read 9 November 2026). Those sentences are the vendor's, and they sit on the card as the vendor caveat.

The API was a single endpoint. Writer's developer documentation describes POST /v1/tools/ai-detect, which took one text input and returned a label of "fake" or "real" with a score from 0 to 1; the page carried a deprecation notice saying the endpoint "is deprecated and will be removed on December 22, 2025" (archived API page, captured 13 December 2025, read 9 November 2026). The older enterprise endpoint is still listed in the vendor's legacy reference as deprecated (legacy API reference, read 9 November 2026). The card records the API as not available. Two detectors tested beside Writer in the same 2023 studies are Copyleaks and GPTZero, and the GPT-2 Output Detector is the other card for a tool its vendor no longer offers.

Limits and plans

The web tool accepted up to 5,000 words per check and no monthly cap was stated, so the free tier field records the per-check limit and the rank field for free words per month stays empty (archived detector page, read 9 November 2026). The detector was never sold on its own: Writer's plans page lists a Starter tier with a 14-day free trial and an Enterprise tier priced through sales, and neither plan mentions AI detection (writer.com/plans, read 9 November 2026). The card therefore carries no plan table.

On retention, the live privacy policy of April 17, 2026 states: "We will only retain your information for as long as we need it to fulfil the purposes for which we collected it", and separately that "in no event will we use any data you provide to us when you access or use our Services to train our models, including any inputs or outputs" (privacy policy, read 9 November 2026). The archived detector page carried no retention statement of its own, so the retention rank field is empty. Ownership comes from the same policy, which gives Writer, Inc. at 111 Maiden Lane, Fourth Floor, San Francisco, and from the terms of use, which apply California law (terms of use, read 9 November 2026).

What the vendor claims

We focus on having a very low number of false positives. This is important for stopping false accusations in situations like academic honesty cases. (AI content detector disclaimer; no accuracy figure published.)
web.archive.org/web/20251209115654/https:..., read

Writer published no accuracy percentage for the detector. The disclaimer says: "Our AI-generated text detectors try to be accurate, but it's hard to be 100% accurate", and on errors: "We focus on having a very low number of false positives. This is important for stopping false accusations in situations like academic honesty cases" (archived disclaimer, read 9 November 2026). No false positive rate accompanies that sentence, so the claim gap field is empty and the false positive figure on this card comes from the studies below.

The archived tool page also told users what to do with a high score: "You should edit your text until there's less detectable AI content" (archived detector page, read 9 November 2026). Weber-Wulff 2023 quotes that prompt in its Figure 13 and observes that "some of the detection tools such as Writer are clearly aimed to be used to hide AI-written text" (arXiv 2306.15666, read 9 November 2026). The card carries both sentences.

What independent studies found

  1. AIDI-S-0001Weber-Wulff 20232023

    Writer: 27 of 54 documents correct (50% binary accuracy, rank 10 of 14); 1 false accusation in 18 human-written documents (5.6%, on a machine-translated text); 19 of 36 AI documents undetected (52.8%); 2 of 9 manually edited and 1 of 9 QuillBot-paraphrased ChatGPT texts classified correctly. Tested March 7 to 28, 2023.

  2. AIDI-S-0003Walters 20232023

    Writer: called 14% of the 42 human-written papers AI, tied with Crossplag for the highest false positive rate of 16 detectors (Table 4); Crossplag, GPT Radar and Writer had false negative rates of 23% to 36%, while the others reached at most 17%. Tested June 25 to July 12, 2023.

  3. AIDI-S-0004Elkhatat 20232023

    Writer: 12 of 15 GPT-3.5 paragraphs flagged, 1 missed, 2 uncertain; 1 of 15 GPT-4 paragraphs flagged, 11 missed, 3 uncertain; of 5 human paragraphs, 2 cleared, 2 rated Likely AI-Generated and 1 Uncertain (Table 4).

  4. AIDI-S-0009Chaka 20232023

    Writer.com AI Content Detector: 7 of 21 English chatbot responses classified as AI-generated, 14 misclassified as human-written, 0 indecisive (rank 4 of 5); 5 of the 6 ChatGPT responses were classified as human-written. Tested 30 March to 2 April 2023.

  5. AIDI-S-0018Pegoraro 20232023

    True positive rate 6.9% on ChatGPT responses and true negative rate 94.5% on human responses, the lowest TPR of the 20 scored methods on a ChatGPT and human Q&A benchmark, March 2023 (Table I).

Four records in the studies index tested Writer, all in 2023, and the results run in the same direction. Weber-Wulff 2023 scored Writer at 27 of 54 documents correct, with 1 false accusation in 18 human-written documents (5.6%, on a machine-translated text) and 19 of 36 AI documents undetected (52.8%); on the obfuscated classes, Writer classified 2 of 9 manually edited and 1 of 9 QuillBot-paraphrased ChatGPT texts correctly (arXiv 2306.15666, Tables 5, 9 and 10, read 9 November 2026).

Walters 2023 found that Writer called 14% of the 42 human-written papers AI, tied with Crossplag for the highest false positive rate among 16 detectors, and that Crossplag, GPT Radar and Writer had false negative rates of 23% to 36% while the others reached at most 17% (DOI 10.1515/opis-2022-0158, Table 4). Elkhatat 2023 tallied Writer at 12 of 15 GPT-3.5 paragraphs flagged but only 1 of 15 GPT-4 paragraphs, and on the 5 human controls Writer rated 2 "Likely AI-Generated" and 1 "Uncertain" (DOI 10.1007/s40979-023-00140-5, Table 4). Chaka 2023 placed Writer fourth of five, with 7 of 21 English chatbot responses classified as AI-generated and 5 of the 6 ChatGPT responses classified as human-written (DOI 10.37074/jalt.2023.6.2.12).

The lowest false positive rate among those records is the Weber-Wulff figure of 1 in 18 (arXiv 2306.15666, Table 9), which is the number the ranking uses, with the Walters 14% and the Elkhatat 2 of 5 recorded beside it in the data file (detectors.json, read 9 November 2026). Every one of these tests ran on the 2023 tool against GPT-3.5 and GPT-4 text, and no record tested the tool after 2023; the vendor withdrew it before any newer test could.

Pegoraro 2023, added 9 November 2026 as the fifth record, ran 20 detection methods on a ChatGPT and human question-and-answer benchmark in March 2023 and records the Writer AI Content Detector at a 6.9% true positive rate on ChatGPT responses, the lowest of the 20 scored methods, and a 94.5% true negative rate on human responses (arXiv 2304.01487, Table I, read 9 November 2026); no ranking field takes a value from the printed rates.

Where Writer appears in the index

Rankings

Writer ranks in four lists. AI detectors ranked by independent studies counts its 4 records; false positive rates sorts on the Weber-Wulff 5.6% of 18; false negative rates sorts on the Elkhatat 12 of 30 AI paragraphs missed (40%); and robustness to edited and paraphrased text sorts on the Weber-Wulff 2 of 9 manually edited texts (22.2%), all read 9 November 2026 in the data file.

Writer does not appear in the free tier, language or retention rankings, because the tool published no monthly word figure, no language count and no retention clause. Every detector card on the register shows Writer with the same fields as the live tools, so a reader who meets a 2023 Writer score in an old integrity file can see what the studies of that year measured. The false positive evidence review covers the 2023 cohort of studies this card draws on.

Questions about Writer AI content detector

Is the Writer AI content detector still available?

No. The vendor's page announced that "this tool will be sunset on December 22" 2025 "along with the AI detection API endpoint", and writer.com/ai-content-detector has redirected to the homepage since 24 December 2025, per the archived captures read 9 November 2026.

How accurate was Writer's AI detector in independent tests?

Weber-Wulff 2023 scored Writer at 27 of 54 documents (50%), rank 10 of 14 (Table 5); Elkhatat 2023 found it flagged 12 of 15 GPT-3.5 paragraphs but 1 of 15 GPT-4 paragraphs (Table 4); Chaka 2023 ranked it fourth of five with 7 of 21 responses correct, each read in the primary paper on 9 November 2026.

What was Writer's false positive rate?

1 of 18 human-written documents (5.6%) in Weber-Wulff 2023 (Table 9), 14% of 42 human papers in Walters 2023 (Table 4), the highest in that study, and 2 of 5 human paragraphs in Elkhatat 2023 (Table 4), each read in the primary paper on 9 November 2026.

How many words could Writer check at once?

Up to 5,000 words per check in the free web tool, per the archived detector page of 9 December 2025, read 9 November 2026. No monthly cap was stated. The withdrawal is set beside the two other withdrawals on the record on which AI detectors were shut down.

Did Writer publish an accuracy figure?

No. The disclaimer says "it's hard to be 100% accurate" and that the tool focuses on "a very low number of false positives", without a rate, per the archived disclaimer read 9 November 2026.

Sources and revisions

Revision log

  1. Card created2026-11-09: pegoraro 2023 record added to the study list in the fifth wave2026-11-10: generator count recounted from 2 to 1 in the data file (chatgpt and gpt-4 are one gpt family under the sitewide rule applied since the ninth wave)

Corrections log

Related reading