AI detectors for law firms and courts: an index fit check
An AI detector for a law firm is judged on where a document goes before it is judged on what the score says, because a draft contract, a witness statement or a brief is privileged and a detector that keeps it, trains on it or shares it has disclosed it. Three fields decide which cards appear in this view: a text card whose retention field states a period or a route that keeps the text out of storage (a deletion-on-analysis statement, a zero-retention flag or mode, or an API that stores nothing), and in the same card either an on-premise or enterprise deployment in plans or integrations, or a report, export or audit record a reviewer can file. Of the 72 cards, 9 text cards state both halves on the pages read on 8 and 9 November 2026. No court filing, hearing transcript or disciplinary decision in the index records a detector score entered as evidence; the one institutional document on the question is Vanderbilt University's statement of 16 August 2023 disabling Turnitin's indicator, read beside the results as evidence question and its 7 primary documents. This page is one of the views under AI detectors by use case; the business view reads retention beside team seats, the data retention ranking sorts every stated period, and the full index shows all 72 cards on one board.
- Field 1Retention period or no-storage route as stated
- Field 2On-premise or enterprise deployment
- Field 3Report, export or audit record
- Field 4Source, read date
Confidentiality first, then the record a reviewer keeps. The retention value is quoted as the vendor states it, because "deleted after analysis" (Phrasly, Smodin), "cached for 3 days" (Sapling) and "stored until you delete the report" (Winston AI) are different undertakings, and only the first two are a stated period or a no-storage route under this page's rule. The deployment column quotes the plan or integration that puts the detector inside the firm's own environment or under an enterprise contract. The record column quotes the report, export or history the vendor states, because a reviewer who flags a document needs something to file beside the flag. Field rules are in the methodology; the deployment list is the deployment field compiled on 9 November 2026 and shared with the government view.
Which fields decide fit for a law firm?
| Detector | Retention period or no-storage route as stated | On-premise or enterprise deployment | Report, export or audit record | Source, read date |
|---|---|---|---|---|
 Sapling | API results cached on the text for 3 days; zero_retention: true scores a text without a copy kept; teams with retention disabled get it on every request | Enterprise: self-hosted, SSO/SCIM, team controls; servers on AWS us-west | API responses and web reports; file batches of 10 on Pro and Enterprise | sapling.ai/privacy, sapling.ai/pricing, 8 November 2026 |
 Pangram | Dashboard text stored while the account is active and deletable from the history tab; zero data retention options for the enterprise API | Enterprise API; Team plan at $20 per seat | Dashboard history of queries; PDF, DOCX and RTF upload of up to 100 files | pangram.com/data-privacy, pangram.com/pricing, 8 November 2026 |
 Copyleaks | API: content submitted to the AI Detector API is not stored, shared, or used to train public AI models; platform users may opt out of model training by contacting support | Education and Enterprise plans, custom | Reports through the platform; documents up to 2,000 pages | copyleaks.com/privacy-policy, copyleaks.com/api/ai-detector, 8 November 2026 |
 ZeroGPT | Private content retained only for the period necessary to operate the service; no-retention modes for API and enterprise; infrastructure in Germany | Enterprise plan by quote; API plans | PDF reports and detection history on paid plans | zerogpt.com/privacy-policy, zerogpt.com/pricing, 8 November 2026 |
 Hive | Default retention 14 days; changed on request, including no retention | On-Prem (Offline) and Enterprise Application tiers, contact sales | API results per 2,048-character chunk; moderation dashboard | docs.thehive.ai FAQ, deploy.com/pricing, 8 November 2026 |
 Grammarly Authorship | Authorship data kept on Grammarly's servers for up to 10 years; a shared report is stored until the account is deleted and can be deleted per document | Enterprise through Superhuman Go; Grammarly for Education admin panel | A shareable Authorship report recording typed, pasted and AI-assisted text | support.grammarly.com, grammarly.com/pro, 9 November 2026 |
 Compilatio | Documents stored 5 years from the end of the contractual relationship on servers in France | Magister+ licensed to institutions; REST API; SSO | Passage-level report with the estimated proportion of AI text highlighted | support.compilatio.net, compilatio.net/magister, 9 November 2026 |
 TruthScan | Content retained 30 days by default; Zero Data Retention on Business and above for images and PDFs, video or audio on request | Enterprise: dedicated or on-prem deployment, regional (EU/UK) processing, MSA and DPA | Audit-ready detection reports, CSV export, detection history (Starter and above) | truthscan.com/privacy, truthscan.com/pricing, 9 November 2026 |
 AI or Not | Uploaded media processed and deleted immediately after inference; free text tool: no data stored | Enterprise: on-prem deployment, private instances | API reports by modality; reverse image search report | docs.aiornot.com, aiornot.com/pricing, 9 November 2026 |
Values as each vendor states them, read 8 and 9 November 2026. Nine other text cards state one half and not the other and are absent: Winston AI stores submitted text until the report or account is deleted and sells PDF reports, a stored-until-deleted policy rather than a period; GPTZero's API stores nothing while its dashboard inputs are stored and used in aggregate, with a team plan and no report export on the card; Originality.ai keeps scan history for 30 days or a year by plan and may train on it unless the user opts out; Turnitin is licensed to institutions and deletes whole submissions on institutional request, with no period stated; Scribbr, Phrasly, Smodin and Undetectable AI state deletion on analysis or no storage with no enterprise deployment or report export; Walter Writes and Proofademic state no period for submitted text. Hive's text row is the one card in this table whose text detection has no price line on the vendor's pricing page.
Why each card is in this view
- AIDI-D-0010Sapling
a 3-day cache on API results, a zero_retention flag, and an Enterprise plan that is self-hosted with SSO/SCIM (sapling.ai/privacy, sapling.ai/pricing, read 8 November 2026); 0 of 42 human papers flagged in Walters 2023 with 31% returned uncertain, and 1 of 7 detectors in Liang 2023. The firm's own counsel should read the clause that content "is stored by us and used only to improve the Service" for the web product.
- AIDI-D-0006Pangram
zero data retention for the enterprise API and a history tab that deletes content from the servers (pangram.com/data-privacy, read 8 November 2026); 0.0000 of 1,992 human passages flagged in Jabarian and Imas 2025 and 0 of 40 human papers in Van Vlasselaer 2026, which also saw it flag 45.5% of 1,163 real theses when run alone. Detection is trained on a dataset the vendor says excludes user submissions.
- AIDI-D-0003Copyleaks
the API page states that submitted content is not stored, while the platform policy allows model training unless a direct customer opts out and a Shared Data Hub holds user-submitted documents (copyleaks.com/privacy-policy, read 8 November 2026); 0 of 42 in Walters 2023, 5 of 10 flagged in Perkins 2024. The two routes carry two different undertakings and a firm should know which one it is on.
- AIDI-D-0011ZeroGPT
no-retention modes for API and enterprise, hosting in Germany, and PDF reports with history on paid plans (zerogpt.com/privacy-policy, read 8 November 2026); 0 of 18 human documents in Weber-Wulff 2023, 2% of 42 in Walters 2023, 1.42% at threshold 0.5 in RAID 2024, and a 45.7% true positive rate in Pegoraro 2023, the best of that study's online tools.
- AIDI-D-0016Hive
14-day default retention adjustable to none on request, and On-Prem (Offline) and Enterprise Application tiers (docs.thehive.ai, read 8 November 2026); 2 image records and no text record in the index, so the text detector a firm would use has no indexed test.
- AIDI-D-0026Grammarly Authorship
a stated period of up to 10 years and per-document deletion, with Enterprise and Education administration (support.grammarly.com, read 9 November 2026); the product is a record of how a document was written, not a probability score, and returns nothing for text it did not watch being typed. The 10-year period is the longest stated on any card in the data retention ranking.
- AIDI-D-0032Compilatio
5 years of storage after the contract ends, on servers in France, under an institutional licence with a passage-level report (support.compilatio.net, read 9 November 2026); 0 of 18 human documents in Weber-Wulff 2023, its only record. Magister+ is sold to institutions, so a firm would buy as an organisation, not a user.
- AIDI-D-0034TruthScan
30-day default retention, a Zero Data Retention toggle on Business and above, and an Enterprise tier with on-prem deployment and EU/UK processing (truthscan.com/privacy, truthscan.com/pricing, read 9 November 2026); no study record. The terms state that User Content "is not your confidential information", a clause a firm should read before any upload.
- AIDI-D-0025AI or Not
media deleted immediately after inference, no data stored by the free text tool, and Enterprise on-prem deployment with private instances (docs.aiornot.com, read 9 November 2026); 1 image record (as Optic, in Ha and Passananti 2024) and no text record.
What the evidence says for legal documents
No indexed study scored contracts, pleadings or legal memoranda, and no indexed document records a detector score admitted as evidence in a court or hearing; the results as evidence question, added 9 November 2026, holds 7 primary documents on how a score should be used and cites no case. The one institutional decision is Vanderbilt University's: it enabled Turnitin's indicator on 4 April 2023 with the warning that the tool should be "a guideline, not a grading metric", then disabled it on 16 August 2023 on the arithmetic of the vendor's own claim, a 1% false positive rate applied to the 75,000 papers the university had submitted in 2022, "around 750 student papers could have been incorrectly labeled" (vanderbilt.edu, read 9 November 2026; Vanderbilt 2023 record). A firm that reads a score into a dispute inherits the same arithmetic at its own volume.
The register of formal prose is the text class with the worst false positive record in the index. Liang 2023 reported an average false positive rate of 61.22% across seven detectors on 91 essays by non-native English writers against near-perfect accuracy on 88 US 8th-grade essays (arXiv 2304.02819, read 8 November 2026; Liang 2023 record), and the authors attribute the gap to low lexical variety and simpler syntax, the features of a careful drafter writing in a constrained register. Weber-Wulff 2023 found that detectors' accuracy averaged 94% on human-written text and 69% on human text machine-translated into English, and 42% on AI text edited by a human (arXiv 2306.15666, read 8 November 2026; Weber-Wulff 2023 record). Templated and formulaic language is excluded or caveated by several vendors themselves: Turnitin's FAQ says its model eliminates "lists, headers, references, data tables, and text that appears to be formulas or code" before scoring, and the bullet points and code question lists the 10 cards that state a non-prose limit.
Retention is the field a firm can verify before any score exists. Of the 9 cards above, Sapling, ZeroGPT, Pangram and TruthScan offer a flag, mode or plan that keeps the text out of storage, AI or Not states deletion on inference, Hive and Copyleaks state a default and a request route, and Grammarly Authorship and Compilatio state fixed periods of 10 years and 5 years, all as read on 8 and 9 November 2026. The deletion routes question lists the 17 cards with a stated removal mechanism, and the offline question the 14 that can run off the vendor's cloud. No detector score should be the only evidence in a misconduct, hiring or publishing decision, and a privileged document should not leave the firm to obtain one.
Sources
- https://sapling.ai/privacy (read 8 November 2026)
- https://sapling.ai/pricing (read 8 November 2026)
- https://www.pangram.com/data-privacy (read 8 November 2026)
- https://www.pangram.com/pricing (read 8 November 2026)
- https://copyleaks.com/privacy-policy (read 8 November 2026)
- https://copyleaks.com/api/ai-detector (read 8 November 2026)
- https://www.zerogpt.com/privacy-policy (read 8 November 2026)
- https://www.zerogpt.com/pricing (read 8 November 2026)
- https://docs.thehive.ai/docs/frequently-asked-questions-faq (read 8 November 2026)
- https://deploy.com/pricing (read 8 November 2026)
- https://support.grammarly.com/hc/en-us/articles/29548735595405 (read 9 November 2026)
- https://www.grammarly.com/pro (read 9 November 2026)
- https://support.compilatio.net/hc/en-us/articles/4416953631505 (read 9 November 2026)
- https://www.compilatio.net/en/magister (read 9 November 2026)
- https://truthscan.com/privacy (read 9 November 2026)
- https://truthscan.com/pricing (read 9 November 2026)
- https://docs.aiornot.com/ (read 9 November 2026)
- https://www.aiornot.com/pricing (read 9 November 2026)
- https://www.vanderbilt.edu/brightspace/2023/08/16/guidance-on-ai-detection-and-why-were-disabling-turnitins-ai-detector/ (read 9 November 2026)
- https://arxiv.org/pdf/2304.02819v3 (read 8 November 2026)
- https://arxiv.org/pdf/2306.15666 (read 8 November 2026)
Revision log: 2026-11-09: view created at index version 2026.11.9
FAQ: AI detectors for law firms
Which AI detectors state both a retention undertaking and an enterprise route?
Nine of 72 cards on 8 and 9 November 2026: Sapling, Pangram, Copyleaks, ZeroGPT, Hive, Grammarly Authorship, Compilatio, TruthScan and AI or Not, each quoted in the table with its source. Four of the nine (Sapling, ZeroGPT, Pangram, TruthScan) offer a zero-retention flag, mode or plan.
Has an AI detector score been used in court?
The index records no court filing, hearing transcript or disciplinary decision with a detector score as evidence, and this page cites none. The 7 primary documents it holds on the question, from Vanderbilt's 16 August 2023 statement to Turnitin's FAQ, are listed on the results as evidence question.
Which detectors can run inside a firm's own environment?
Four of the nine state an on-premise, self-hosted or private deployment: Sapling (self-hosted Enterprise), Hive (On-Prem Offline), TruthScan (dedicated or on-prem Enterprise) and AI or Not (on-prem, private instances), as read on 8 and 9 November 2026. The offline question lists 14 cards in all, including open-source tools.
What is the false positive record on formal writing?
Liang 2023 measured an average 61.22% false positive rate across seven detectors on 91 essays by non-native English writers (arXiv 2304.02819, read 8 November 2026); Weber-Wulff 2023 found 0 false accusations for 8 of 14 tools and 9 of 18 for GPT Zero on 18 human documents (arXiv 2306.15666, read 8 November 2026). No indexed study scored legal documents.
Does any vendor keep a document for a fixed number of years?
Two of the nine: Grammarly Authorship states up to 10 years for Authorship data and Compilatio 5 years from the end of the contract, as read on 9 November 2026. The data retention ranking sorts every stated period across the index.
Should a firm act on a single detector score?
The vendors say no in their own words: Turnitin's FAQ says the indicator "should not be used as the sole basis for action", and Writer, before withdrawing its tool, wrote that low false positives matter "for stopping false accusations", both as read on 8 and 9 November 2026 and quoted on the cards. Vanderbilt's statement puts the arithmetic on a 1% claim at 750 wrong flags in a year. The page's rule is the same: no score as the only evidence.
