Which AI detectors run on-premise or offline? The four that run on your own machine, the one open-source text component, the one on-device PC tool and the ten vendors that state an on-premise option

Three, all research code rather than products, and all for text. GLTR runs a local server against GPT-2 117M and colours each word by its rank under that model; its repository was created on 11 April 2019 and last pushed on 18 January 2024 (github.com, read 9 November 2026). The GPT-2 Output Detector is OpenAI's 2019 RoBERTa classifier, whose weights and code are MIT-licensed and whose local web UI answers at localhost:8080; the hosted demo on Hugging Face was paused as of 9 November 2026, so the local route is the one that works (github.com, read 9 November 2026). DetectGPT is the official implementation of the 2023 Stanford paper, last pushed on 4 March 2023, and runs the paper's experiments from the command line with a source model and a T5-3B or T5-11B mask-filling model, perturbing each passage about 100 times (github.com, read 9 November 2026). The test record of the three is on their cards: the GPT-2 Output Detector has 12 study records, GLTR 4 and DetectGPT 4 in the ranking by independent testing, and Jabarian and Imas 2025 recorded a 0.9503 false positive rate for the RoBERTa detector on 1,992 human passages (NBER w34223, read 9 November 2026). Running locally says nothing about accuracy.

SynthID is the partial case. Google DeepMind publishes the SynthID Text logits processor and Bayesian detector in Hugging Face Transformers v4.46.0 and later and in the google-deepmind/synthid-text repository, last pushed 26 October 2026, but that code detects only text from models that applied a shared watermark configuration; the image, video and audio checks run on the hosted SynthID Detector portal and in the Gemini app (synthid.com, read 9 November 2026; open-source view).

Ten, nine of them on a plan with no published price. Sensity is the most specific: an On-premise plan with a server on dedicated NVIDIA GPUs, a single-investigator workstation, or a USB dongle for air-gapped field work, and a documentation note that on-premise deployments "do not enforce fixed media limits" (sensity.ai; docs.sensity.ai, read 9 November 2026). Resemble Detect names on-prem, air-gapped and private VPC deployment on AWS, GCP and Azure with Docker and Kubernetes, and lists Carahsoft contract vehicles for US government buyers (resemble.ai, read 9 November 2026). Hive's pricing page has an "On-Prem (Offline)" tier by contact (deploy.com, read 8 November 2026); AI or Not, WasItAI and Codeleaks put on-premise deployment or installation in an Enterprise plan (aiornot.com; wasitai.com; copyleaks.com, read 9 November 2026); TruthScan's Enterprise plan, "$0.005 or less per result", offers "dedicated or on-prem deployment" (truthscan.com, read 9 November 2026); Deepware's FAQ offers "an offline environment via SDK" by contact (deepware.ai, read 9 November 2026). Sapling is the one text detector with a stated self-hosted tier, on an Enterprise plan that starts at 10 seats at $15 a seat a month (sapling.ai, read 8 November 2026). Pindrop's privacy notice refers to "on-premises customers" without a corresponding plan on the pages read (pindrop.com, read 9 November 2026).

Seven of the ten, Sensity, Resemble, Hive, AI or Not, WasItAI, Deepware and Pindrop, are image, video or voice detectors first, which fits the buyers those pages name: forensic teams, contact centres and government. Among the text-first products, Sapling and Codeleaks are the only two with the option, and Codeleaks no longer returns an AI verdict for code (docs.copyleaks.com, read 9 November 2026).

The government view, added 9 November 2026, reads these 10 on-premise options beside each card's data region, named standards and retention, and notes that 7 of the 10 are image, voice or video detectors; the law firms view, added the same day, requires an on-premise or enterprise route beside a retention undertaking.

The other 28 cards state a web tool, an API or an LMS integration and no local or on-premise route on the pages read. Two cards state the opposite of offline in their own words. Turnitin Clarity's instructor FAQ says all writing must happen inside the browser writing space, "there is no offline editing and the editor freezes when the connection drops" (guides.turnitin.com, read 9 November 2026). Walter Writes says it processes text "using our own self-hosted AI models", a statement about where the vendor runs its models, not an option for the customer (walterwrites.ai, read 8 November 2026).

Cards with no local or on-premise option stated

Delivery as the integrations and plans fields of each card describe it, read on the date shown.

An on-premise option moves the data; it does not move the model's error rate, and none of the ten vendors publishes a separate accuracy figure for the on-premise build.

McAfee's on-device detector runs on a Windows PC, not a phone; the mobile apps view lists the five cards whose integrations name an iOS or Android app, read 10 November 2026.

All read 9 November 2026 unless an earlier date is shown; the source of every row is linked in the tables.

• GLTR repository: https://github.com/HendrikStrobelt/detecting-fake-text
• GPT-2 Output Detector README: https://github.com/openai/gpt-2-output-dataset/tree/master/detector
• DetectGPT repository: https://github.com/eric-mitchell/detect-gpt
• SynthID Detector: https://synthid.com/
• SynthID Text repository: https://github.com/google-deepmind/synthid-text
• Sensity deepfake detection: https://sensity.ai/deepfake-detection-for-video-image-audio/
• Sensity formats and limits: https://docs.sensity.ai/resources/formats-limits
• Resemble AI Detect: https://www.resemble.ai/products/detect
• Resemble AI pricing: https://www.resemble.ai/pricing
• Deploy (Hive) pricing: https://deploy.com/pricing (read 8 November 2026)
• AI or Not pricing: https://www.aiornot.com/pricing
• TruthScan pricing: https://truthscan.com/pricing
• WasItAI API: https://wasitai.com/api
• Deepware FAQ: https://deepware.ai/us/faq/index.html
• Copyleaks pricing: https://copyleaks.com/pricing
• Copyleaks release notes: https://docs.copyleaks.com/resources/updates/release-notes
• Sapling pricing: https://sapling.ai/pricing (read 8 November 2026)
• Pindrop privacy notice: https://www.pindrop.com/legal/privacy-notice-for-pindrop-solutions
• Turnitin Clarity instructor FAQs: https://guides.turnitin.com/hc/en-us/articles/37198804513165-FAQs-for-Instructors-using-Turnitin-Clarity
• Turnitin AI writing detection FAQs: https://guides.turnitin.com/hc/en-us/articles/28477544839821-Turnitin-s-AI-writing-detection-capabilities-FAQs (read 8 November 2026)
• GPTZero: https://gptzero.me/ (read 8 November 2026)
• Pangram data privacy: https://www.pangram.com/data-privacy (read 8 November 2026)
• Walter Writes privacy policy: https://walterwrites.ai/privacy-policy/ (read 8 November 2026)
• Jabarian and Imas 2025, NBER w34223: https://www.nber.org/papers/w34223

Revision log: 2026-11-09: page created from the plans, integrations and free tier fields of 42 cards at index version 2026.11.7.; 2026-11-09: eighth-wave pages linked; 2026-11-10: RADAR (local run) and McAfee Deepfake Detector (on-device) added to the table, Ghostbuster (local code that calls OpenAI), Hiya and five web tools to the second table (eleventh wave); 2026-11-10: twelfth-wave pages linked