AI detectors for government and public sector buyers
An AI detector for a public sector buyer is judged on three fields before any accuracy claim: whether it can run inside the agency's own environment, where its servers are, and which security standards the vendor states in writing. Three fields decide which cards appear in this view: a deployment beyond the vendor's cloud (on-premise, self-hosted, private cloud or offline SDK) stated in plans or integrations, a named data region in the retention or legal pages, or a SOC Type 2 or ISO 27001 statement on a vendor trust or security page, compiled into the deployment and compliance fields on 9 November 2026 with the source and read date of each. Of the 72 cards, 20 state at least one of the three (Trinka joined on 10 November 2026 with the tenth wave, McAfee Deepfake Detector with the eleventh): 12 state an on-premise, self-hosted, offline or on-device option, 9 state a data region or regional hosting (8 naming the region), 8 name SOC Type 2 or ISO 27001, and 0 state FedRAMP or StateRAMP on any page read. Every compliance entry is a vendor statement read on the vendor's own page, not a certificate the index has inspected. The retention column is carried for every row because the data retention ranking and the deletion routes question are the pages a procurement officer will read next. This page is one of the views under AI detectors by use case; the compliance view lists every named standard across the index, the offline question the 14 cards that run off the vendor's cloud, and the full index shows all 72 cards.
- Field 1Deployment beyond the vendor's cloud
- Field 2Data region as stated
- Field 3Named standards (where read)
- Field 4Retention as stated
- Field 5Source, read date
Deployment first, because an agency that cannot send documents outside its network has no use for a cloud-only tool however it scores. The deployment cell quotes the plan or integration as stated; a "dedicated server" or a vendor's own self-hosted models are not a customer deployment and sit in the notes. The region cell quotes the country or region the vendor names for its servers or processing. The standards cell lists the named standards and where they were read: a trust or security page at the vendor's domain fetched live on 9 November 2026 (Grammarly, Sapling, Writer, TruthScan, Sightengine), or a page already cited on the card (GPTZero, ZeroGPT, Compilatio, Resemble). Unnamed claims, Reality Defender's "a range of US, UK and EU certifications" and Codeleaks' "Military Grade Security and Compliance", are not standards and do not qualify. Field rules are in the methodology.
Which fields decide fit for a public sector buyer?
| Detector | Deployment beyond the vendor's cloud | Data region as stated | Named standards (where read) | Retention as stated | Source, read date |
|---|---|---|---|---|---|
 Sapling | Enterprise: self-hosted, SSO/SCIM | AWS us-west | SOC Type II, GDPR; HIPAA and PCI support (sapling.ai/security, 9 November 2026) | API results cached 3 days; zero_retention flag | sapling.ai/pricing, sapling.ai/privacy, 8 November 2026 |
 TruthScan | Enterprise: dedicated or on-prem deployment | EU/UK regional processing (Enterprise) | SOC Type 2, GDPR, ISO 27001:2022 (trust.truthscan.com, 9 November 2026) | 30 days by default; Zero Data Retention on Business and above | truthscan.com/pricing, truthscan.com/privacy, 9 November 2026 |
 Resemble Detect | On-prem, air-gapped, private VPC (AWS, GCP, Azure); Carahsoft US government contract vehicles | Not named | SOC Type 2 documentation on Enterprise (pricing page, 9 November 2026) | Media deleted after analysis unless flagged; zero_retention_mode flag | resemble.ai/products/detect, resemble.ai/pricing, 9 November 2026 |
 Hive | On-Prem (Offline) and Enterprise Application tiers | Not named | Not read (trust and security pages served no text) | 14 days by default; changed on request, including none | deploy.com/pricing, docs.thehive.ai, 8 November 2026 |
 Sensity | On-premise server, workstation or USB dongle for air-gapped environments | Not named | Not stated | Biometric data deleted within 3 years; submissions deletable through the API; no period for media | docs.sensity.ai, sensity.ai/privacy-policy, 9 November 2026 |
 AI or Not | Enterprise: on-prem deployment, private instances | Not named | Not stated | Uploaded media deleted immediately after inference | aiornot.com/pricing, docs.aiornot.com, 9 November 2026 |
 WasItAI | Enterprise: option for on-prem installation | Not named | Not stated | Images not stored beyond the analysis request; logs 90 days | wasitai.com/api, wasitai.com/privacy, 9 November 2026 |
 Codeleaks | Enterprise: on-premises or private cloud hosting | "Regional hosting", region not named | Not read (copyleaks.com/security returned 403) | Account deleted within 90 days after an enterprise agreement ends; no period for code | copyleaks.com/pricing, copyleaks.com/legal/privacypolicy, 9 November 2026 |
 Pindrop Pulse | On-premises customers named in the privacy notice | Not named | Not stated | Voice features up to 3 years; Pulse for Meetings DPA: destruction within 90 days after a 30-day return period | pindrop.com privacy notice, 9 November 2026 |
 Deepware | Offline SDK by contact | Not named | Not stated | Samples shared with the service's partners under the terms; no period stated | deepware.ai FAQ, terms, 9 November 2026 |
 ZeroGPT | None stated | Germany | GDPR (privacy policy, 8 November 2026) | Retained only as long as needed to operate the service; no-retention modes for API and enterprise | zerogpt.com/privacy-policy, 8 November 2026 |
 Compilatio | None stated (institutional licence, LMS and API) | France (Chavanod) | GDPR (support centre, 9 November 2026) | 5 years from the end of the contractual relationship | support.compilatio.net, 9 November 2026 |
 Sightengine | None stated | EU datacenters (most processing); region restriction on Enterprise | GDPR with a DPA; ISO 27001, SOC Type 2 and PCI DSS stated for its data-centre partners (sightengine.com/security, 9 November 2026) | Content deleted once processed | sightengine.com/policies/privacy, 9 November 2026 |
 Proofademic | None stated | United States (processing) | Not stated | Retained as long as necessary; no period for submitted text | proofademic.ai/privacy-policy, 8 November 2026 |
 Undetectable AI | None stated | United States (hosting, terms) | Not stated | "Your text is never stored or used to train models" (detector FAQ) | undetectable.ai/privacy, 9 November 2026 |
 Grammarly | None stated | Not named | SOC Type 2, ISO/IEC 27001:2022, GDPR, HIPAA (grammarly.com/compliance, 9 November 2026) | No retention terms on the detector pages read | grammarly.com/ai-detector, 8 November 2026 |
 Grammarly Authorship | None stated (Enterprise through Superhuman Go) | Not named | SOC Type 2, ISO/IEC 27001:2022, GDPR, HIPAA (grammarly.com/compliance, 9 November 2026) | Authorship data up to 10 years; per-document deletion | support.grammarly.com, 9 November 2026 |
 Writer | None stated; detector withdrawn 22 December 2025 | Not named | SOC Type II, ISO/IEC 27001, 27701 and 42001, HIPAA, PCI (writer.com/trust, 9 November 2026) | Retained as long as needed; inputs never used to train models | writer.com/legal/privacy, 9 November 2026 |
 Trinka | On-premise (enterprise); offline grammar checker app on the Confidential Data plan | US (AWS servers based in the US) | SOC Type 2, ISO 27001:2022, HIPAA, FERPA, GDPR (trinka.ai/trust-center, 10 November 2026) | Processed data deleted every 90 days on Basic and Premium; real-time deletion on the Confidential Data plan | trinka.ai/trust-center, trinka.ai/pricing, 10 November 2026 |
 McAfee Deepfake Detector | On-device: inference on the PC's NPU ("perform the entire identification process [...] directly on the PC, maximizing on-device processing to keep private user data off the cloud") | Not named; the tool runs on the device | None read for the detector | "will not store any audio or video data" (mcafee.com Privacy Notice, January 2026) | mcafee.com newsroom, 21 August 2024, mcafee.com, 10 November 2026 |
Values as each vendor states them, read 8 and 9 November 2026, with the trust and security pages read live on 9 November 2026; the live reads covered the /trust, /security, /trust-center and /compliance paths at each card's official domain, and pages that returned 403 (Copyleaks, Scribbr, QuillBot, Phrasly), a script shell with no text (Hive, Deploy) or a redirect to the home page (ZeroGPT) count as not read. Turnitin's trust centre at turnitin.com/trust-center was read and names no security standard, so Turnitin, Turnitin Clarity, Copyleaks, Pangram, GPTZero (FERPA, GDPR and CCPA in its privacy policy, none of them SOC Type 2 or ISO 27001), Originality.ai, Winston AI, Reality Defender and Codequiry (a "dedicated server" on the Defense add-on, with "audited hosting") are absent. Walter Writes states that its models are self-hosted by the vendor, a statement about the vendor's own infrastructure, and is absent.
Why each card is in this view
- AIDI-D-0010Sapling
a self-hosted Enterprise plan, servers on AWS us-west, and a security page stating SOC Type II and GDPR (sapling.ai/pricing, sapling.ai/security, read 8 and 9 November 2026); 0 of 42 human papers flagged in Walters 2023 with 31% uncertain. The one card in this view stating all three deciding fields.
- AIDI-D-0034TruthScan
dedicated or on-prem Enterprise deployment, EU/UK processing and a trust page listing SOC Type 2, GDPR and ISO 27001:2022 (truthscan.com/pricing, trust.truthscan.com, read 9 November 2026); no study record. The second card stating all three fields.
- AIDI-D-0042Resemble Detect
on-prem, air-gapped and private VPC deployment, Carahsoft contract vehicles for US federal, state and local government, and SOC Type 2 documentation on Enterprise (resemble.ai/products/detect, resemble.ai/pricing, read 9 November 2026); no indexed record, and the vendor's 99.5% Podonos figure is its own.
- AIDI-D-0016Hive
On-Prem (Offline) and Enterprise Application tiers with a 14-day default retention adjustable to none (deploy.com/pricing, read 8 November 2026); 2 image records (Ha and Passananti 2024, RAID image 2025), no text record. Its trust and security pages served no readable text on 9 November 2026.
- AIDI-D-0038Sensity
on-premise server, workstation and USB dongle deployments, the only air-gapped field option stated by a video card (docs.sensity.ai, read 9 November 2026); no indexed record, and no accuracy percentage published.
- AIDI-D-0025AI or Not
Enterprise on-prem deployment and private instances, with media deleted after inference (aiornot.com/pricing, read 9 November 2026); 1 image record as Optic in Ha and Passananti 2024.
- AIDI-D-0021WasItAI
an Enterprise option for on-prem installation and images not stored beyond the request (wasitai.com/api, read 9 November 2026); no indexed record.
- AIDI-D-0039Codeleaks
on-premises or private cloud hosting and regional hosting without a named region (copyleaks.com/pricing, read 9 November 2026); the API's AI verdict for source code was withdrawn on 16 December 2025, so the on-premise option covers plagiarism and licence scanning of code.
- AIDI-D-0041Pindrop Pulse
on-premises customers named in the privacy notice of 27 March 2026, with voice features retained up to 3 years (pindrop.com, read 9 November 2026); no indexed record, and Pulse Inspect was launched for governments among its named users.
- AIDI-D-0037Deepware
an offline SDK by contact (deepware.ai FAQ, read 9 November 2026); the terms share uploaded samples with the service's partners, a clause that rules the web scanner out for any confidential video.
- AIDI-D-0011ZeroGPT
infrastructure hosted in Germany under GDPR and no-retention modes for API and enterprise (zerogpt.com/privacy-policy, read 8 November 2026); 0 of 18 human documents in Weber-Wulff 2023, 1.42% at threshold 0.5 in RAID 2024.
- AIDI-D-0032Compilatio
servers in France, a GDPR statement and a 5-year storage period after the contract ends (support.compilatio.net, read 9 November 2026); 0 of 18 in Weber-Wulff 2023.
- AIDI-D-0036Sightengine
most processing in EU datacenters, a GDPR DPA and a security page that states ISO 27001, SOC Type 2 and PCI DSS for its data-centre partners rather than for itself (sightengine.com/security, read 9 November 2026); 1 image record, RAID image 2025.
- AIDI-D-0014Proofademic
processing in the United States stated in the privacy policy, with no training or resale of submissions (proofademic.ai/privacy-policy, read 8 November 2026); no study record.
- AIDI-D-0024Undetectable AI
hosting in the United States stated in the terms and a never-stored statement on the detector FAQ (undetectable.ai/privacy, read 9 November 2026); no study record.
- AIDI-D-0008Grammarly
a compliance page stating SOC Type 2, ISO/IEC 27001:2022, GDPR and HIPAA for the Grammarly service (grammarly.com/compliance, read 9 November 2026); no study record for the detector, a RAID leaderboard entry of 9 February 2026 recorded on the leaderboard question.
- AIDI-D-0026Grammarly Authorship
the same compliance page, with Authorship data kept up to 10 years and an Education admin panel (support.grammarly.com, read 9 November 2026); a writing record, not a probability score.
- AIDI-D-0023Writer
a trust page stating SOC Type II, ISO/IEC 27001, 27701 and 42001, HIPAA and PCI (writer.com/trust, read 9 November 2026); the detector itself was withdrawn on 22 December 2025, so the row records the company's statements for a tool a buyer can no longer licence.
- AIDI-D-0055Trinka
the one tenth-wave card that states all three fields, an on-premise enterprise deployment, US AWS servers and five named standards on a trust centre fetched live (trinka.ai/trust-center, read 10 November 2026); the privacy policy adds that documents may be saved "according to its policy and Indian government regulations", a line a public buyer outside India will want explained; no study record. Binoculars, added the same day, runs locally from open-source code with no vendor deployment statement, so it stays out under the rule that a local run is not a customer deployment.
- AIDI-D-0074McAfee Deepfake Detector
the first card whose deployment is the reader's own device rather than a vendor server or an on-premise install, a value added to the deployment vocabulary on 10 November 2026 as "on-device (Windows AI PC NPU)"; the press releases place the whole inference on the PC's Neural Processing Unit and the Privacy Notice states no stored audio or video; no standard is named for the detector and no study tested it (mcafee.com newsroom, 21 August 2024, mcafee.com Privacy Notice, read 10 November 2026). Hiya's three API regions (EU, US, Asia) are stated for its Audio Intelligence API and not for the free extension, so Hiya is not listed.
What the evidence says for public sector use
No indexed study scored a detector on government documents, and no card states FedRAMP, StateRAMP or an equivalent public sector authorisation on any page read on 8 or 9 November 2026. The closest public body record in the index is a measurement rather than a procurement: the NIST GenAI text-to-text pilot scored 348 detector submissions from 11 anonymised teams on news summaries and found AUC values from about 0.34 to 1.000, the weakest pairs below chance (DOI 10.6028/NIST.AI.700-1, read 9 November 2026; NIST AI 700-1 record), and no indexed card appears under its own name in that report. The leaderboard question lists the 6 cards with a public RAID entry, read 9 November 2026, as the one place an agency can compare several vendors' submissions on a fixed set.
The deployment field splits the index in a way the accuracy record does not. Of the 10 cards stating an on-premise, self-hosted or offline route, 7 are image, voice or video detectors (Hive, Sensity, Resemble Detect, AI or Not, WasItAI, Pindrop Pulse, Deepware) and 3 cover text (Sapling, TruthScan, Codeleaks for code), while the text detectors with the most study records, GPTZero (16), the GPT-2 Output Detector (12) and Copyleaks (9), state no customer deployment; the GPT-2 Output Detector, GLTR and DetectGPT run locally as open-source code with no vendor at all, as the open source view sets out. An agency that needs text detection inside its own network is choosing between tools with no indexed record and open-source models from 2019 to 2023 whose false positive rate on 1,992 human passages reached 0.9503 in Jabarian and Imas 2025 (NBER w34223, read 9 November 2026).
The retention column is where the vendor statements are most specific and most varied: Compilatio's 5 years and Grammarly Authorship's 10 are the longest periods on any card, Hive's 14 days and TruthScan's 30 are defaults with a shorter option, and Sapling, ZeroGPT, TruthScan and Resemble Detect each state a flag or mode that keeps the input out of storage, all as read on 8 and 9 November 2026 and sorted on the data retention ranking. Every standard in the table is the vendor's own statement on its own page; the index has inspected no certificate, report or audit letter, and a buyer should request the document the statement describes. No detector score should be the only evidence in a misconduct, hiring or publishing decision, and the vendors' own caveats on the cards say so in their words.
Sources
- https://sapling.ai/pricing (read 8 November 2026)
- https://sapling.ai/privacy (read 8 November 2026)
- https://sapling.ai/security (read 9 November 2026)
- https://truthscan.com/pricing (read 9 November 2026)
- https://truthscan.com/privacy (read 9 November 2026)
- https://trust.truthscan.com/ (read 9 November 2026)
- https://www.resemble.ai/products/detect (read 9 November 2026)
- https://www.resemble.ai/pricing (read 9 November 2026)
- https://deploy.com/pricing (read 8 November 2026)
- https://docs.thehive.ai/docs/frequently-asked-questions-faq (read 8 November 2026)
- https://docs.sensity.ai/ (read 9 November 2026)
- https://sensity.ai/privacy-policy/ (read 9 November 2026)
- https://www.aiornot.com/pricing (read 9 November 2026)
- https://docs.aiornot.com/ (read 9 November 2026)
- https://wasitai.com/api (read 9 November 2026)
- https://wasitai.com/privacy (read 9 November 2026)
- https://copyleaks.com/pricing (read 9 November 2026)
- https://copyleaks.com/legal/privacypolicy/ (read 9 November 2026)
- https://www.pindrop.com/legal/privacy-notice-for-pindrop-solutions (read 9 November 2026)
- https://deepware.ai/us/faq/index.html (read 9 November 2026)
- https://deepware.ai/us/terms-of-services/index.html (read 9 November 2026)
- https://www.zerogpt.com/privacy-policy (read 8 November 2026)
- https://support.compilatio.net/hc/en-us/articles/4416953631505 (read 9 November 2026)
- https://sightengine.com/policies/privacy (read 9 November 2026)
- https://sightengine.com/security (read 9 November 2026)
- https://proofademic.ai/privacy-policy/?utm_source=aidetectorindex.com&utm_medium=owned&utm_campaign=for_government (read 8 November 2026)
- https://undetectable.ai/privacy (read 9 November 2026)
- https://www.grammarly.com/ai-detector (read 8 November 2026)
- https://www.grammarly.com/compliance (read 9 November 2026)
- https://support.grammarly.com/hc/en-us/articles/29548735595405 (read 9 November 2026)
- https://writer.com/legal/privacy/ (read 9 November 2026)
- https://writer.com/trust/ (read 9 November 2026)
- https://www.turnitin.com/trust-center (read 9 November 2026)
- https://doi.org/10.6028/NIST.AI.700-1 (read 9 November 2026)
- https://www.nber.org/papers/w34223 (read 9 November 2026)
Revision log: 2026-11-09: view created at index version 2026.11.92026-11-10: Trinka added (tenth wave), 19 cards; Binoculars' local run not counted as a deployment2026-11-10: McAfee Deepfake Detector added (eleventh wave) with the new on-device deployment value, 20 cards; Hiya's API regions not entered for the extension
FAQ: AI detectors for government buyers
Which AI detectors can run on-premise or offline?
Ten of 72 cards state a customer deployment beyond the vendor's cloud on 8 and 9 November 2026: Sapling (self-hosted), TruthScan, Resemble Detect, Hive, Sensity, AI or Not, WasItAI, Codeleaks, Pindrop Pulse and Deepware (offline SDK). Three open-source tools run locally with no vendor and are on the offline question.
Does any AI detector state FedRAMP authorisation?
No card states FedRAMP or StateRAMP on any page read on 8 or 9 November 2026, including the five trust pages fetched live. Resemble Detect names Carahsoft contract vehicles for US federal, state and local government, a purchasing route rather than an authorisation.
Which vendors state SOC Type 2 or ISO 27001?
Seven cards: Grammarly and Grammarly Authorship (SOC Type 2, ISO/IEC 27001:2022), Sapling (SOC Type II), TruthScan (SOC Type 2, ISO 27001:2022), Writer (SOC Type II, ISO/IEC 27001, 27701, 42001), Resemble Detect (SOC Type 2 documentation on Enterprise) and Sightengine (ISO 27001 and SOC Type 2 stated for its data-centre partners), as read on 9 November 2026. Each is a vendor statement; the index inspected no certificate.
Which detectors name where the data is held?
Eight: Sapling (AWS us-west), ZeroGPT (Germany), Compilatio (France), Sightengine (EU datacenters), TruthScan (EU/UK regional processing on Enterprise), Proofademic and Undetectable AI (United States) and Codeleaks ("regional hosting", no region named), as read on 8 and 9 November 2026.
Has a public body measured AI detectors?
NIST's GenAI text-to-text pilot (NIST AI 700-1, June 2025) scored 348 submissions from 11 anonymised teams, with AUC from about 0.34 to 1.000 (DOI 10.6028/NIST.AI.700-1, read 9 November 2026). No indexed card appears under its own name in the report, so it measures the field, not a product an agency can buy.
Which on-premise options cover text rather than media?
Three of the ten: Sapling's self-hosted Enterprise plan, TruthScan's dedicated or on-prem Enterprise deployment, and Codeleaks for source code; the other seven are image, voice or video detectors. The text detectors with the most study records state no customer deployment.
