Pangram AI Detector Explained

The Pangram artificial intelligence (AI) detector inspects submitted documents by breaking copy into segments and labeling each section as Human, AI-Assisted, or AI-Generated. At Layer3Labs, we build and run AI workflows inside small and mid-sized business teams, so we know an automated scanner belongs in front of a staff reviewer rather than in place of one. Pangram then calculates the document-level percentage of machine writing, helping reviewers identify sections altered by language models.

Users can run 20 free checks every day on the Pangram home page, while individual paid subscriptions begin at $20 per month.

How Pangram Detects AI Text

The Pangram AI detector evaluates text by dividing documents into variable-length sequences, calculating token-level predictions, and decoding those calculations into three specific labels: Human, AI-Assisted, or AI-Generated. The Pangram 4.1 model card was released on October 19, 2026. It reports start and end positions, segment word counts, an assistance score, and a humanizer score. The model also returns class probabilities for each classified portion.

Pangram explains on its home page that its classifier operates on a traditional language model architecture. Pangram tokenizes incoming copy, maps tokens to vector embeddings, and routes them through a neural network with a classification head to output predictions. Pangram trained the foundational model on roughly 1 million publicly licensed human documents alongside AI-generated text from GPT-5 and other frontier models.

On its home page FAQ, Pangram states that its architecture does not use perplexity or other metrics that correlate with AI writing. Pangram reads style, word choice, syntax and grammar. The text check runs purely on the submitted draft without inspecting cut-and-paste telemetry or querying external search indexes, while a separate Similarity Checker handles plagiarism cross-referencing.

Pangram processes documents across more than 20 languages, with the official model card documenting support for 24 distinct languages. Users access 20 daily free checks directly through the browser interface, and larger scanning volumes connect via paid subscriptions outlined in our Pangram pricing by plan review.

In the independent Chicago Booth Review report covering commercial tools, Pangram maintained the lowest false-positive rate among tested options. A flag is a reason for a person to review. It is not proof. This is detailed in our guide on whether Pangram is accurate.

Pangram's model card lists operational limits for its 4.1 text model.

A direct contradiction exists between Pangram product pages regarding programming languages. The Pangram recruiter use case claims Pangram accurately detects AI-generated code written in Python, Java, and C++. In contrast, the technical model card explicitly designates source code as out of scope. Readers evaluating software screening should verify current test parameters on the Pangram 4.1 model card before scanning candidate codebases.

• Writing samples must contain at least 50 words arranged primarily in complete grammatical sentences.
• The dashboard shows less certainty under 150 words.
• Out-of-scope materials include source code, Markdown text, reference sections, tables of contents, templated or automated writing, mathematical notation, and machine translations where results become less reliable.
• Pangram recommends raw text and .docx files over PDFs.

Pangram Image 1.0 and Image Flash 1.0

On November 4, 2026, Pangram announced Pangram Image 1.0 for static picture inspection and Pangram Image Flash 1.0 for frame-by-frame video screening. As documented in the Pangram Image 1.0 launch post, the image model overlays visual heatmaps directly onto synthetic graphics to pinpoint regions showing artificial synthesis. It also demonstrates robustness to edits, compression, and low resolution.

Pangram reports that Image 1.0 achieved 99.96 percent accuracy on an internal benchmark comprising 1,200 synthetic images generated across 12 distinct synthesis engines matched against 100 authentic photos. For video processing, Pangram states that Image Flash 1.0 attained 98.7 percent accuracy on the academic DF26 benchmark, supported by programmatic Application Programming Interface (API) endpoints.

A documentation discrepancy exists regarding minimum resolution requirements for image uploads. The home page web widget requires minimum dimensions of 512 by 512 pixels for image files and 720 by 720 pixels for video clips. In contrast, the official launch post states that Image 1.0 accepts files down to 256 by 256 pixels. It includes an explicit notice of reduced classification confidence below 512 pixels. Readers should review technical specifications directly in the Pangram Image 1.0 launch post when configuring ingestion pipelines.

Teams evaluating multimedia verification can explore broader industry capabilities across our detailed roundups of AI image detectors compared and AI video detectors compared.

Business Uses for the Pangram AI Detector

Pangram's use-case pages describe screening incoming documents so that a person reads the flagged segments first, and none of them recommends an automatic rejection.

Every operational deployment requires placing Pangram in front of qualified team members rather than automating adverse actions. On its own marketing pages, Pangram explicitly clarifies that its scores constitute a probabilistic signal rather than dispositive proof. It advises recruitment teams to treat output as an interview prompt. Its law-firm page says detection supports decisions and does not replace legal judgment.

Pangram is not for anyone who needs proof of who wrote a text. A score alone should not drive automated discipline, termination, or rejection.

What Pangram Costs

Pangram structures its commercial tiers across self-serve monthly plans and pay-as-you-go developer endpoints as published on November 8, 2026.

Prospective buyers can evaluate volume-based cost modeling and break-even consumption thresholds inside our dedicated breakdown of Pangram pricing by plan.

Pangram leaves several purchase parameters unspecified on its public tiers. As of November 8, 2026, the Pangram pricing page advertises annual savings figures of $60 for Individual, $240 for Professional, and $120 for Team plans. It does not publish the total upfront annual cost. Furthermore, Pangram omits a precise monthly word allowance for the Team tier and does not state the volume threshold required to activate the 20 percent bulk text API discount.

Where the Pangram AI Detector Does Not Fit

The Pangram AI detector does not provide forensic proof of who sat at a keyboard and drafted a passage. Because the classifier analyzes writing style and syntax rather than verifying individual identity, teams should never treat scores as proof of authorship.

Organizations with sporadic scanning requirements do not need commercial subscriptions. The 20 free daily scans available through the browser interface comfortably accommodate occasional document checks without recurring software overhead.

Prospective enterprise customers must evaluate conflicting statements regarding Pangram data retention. The Pangram data privacy FAQ states that submitted user content is encrypted at rest and in transit. Content is retained for the active lifetime of the user account. It is modifiable via history tab deletions. Conversely, Pangram sales pages for law firms and publishers state that processed documents remain strictly ephemeral and are not retained on servers. Enterprise buyers handling confidential records should confirm which retention policy applies with the Pangram data privacy FAQ before transmitting sensitive documents. Pangram does confirm across its documentation that it never sells user information to external parties and does not train foundational AI models like ChatGPT on customer submissions.

Readers asking whether a detector score affects search performance can review our guide on does Google penalize AI content. For a wider look at alternative tools across the industry, compare feature sets within our review of AI content detectors compared.

The answer would change under specific operational conditions. A published third-party academic study newer than the September 2025 Chicago Booth working paper could demonstrate altered accuracy bounds across newer generative engines. A price change to a Pangram plan would also change the answer. So would a monthly volume under 40,000 words, where the API costs less than the Individual plan.

To test these capabilities against real operational workflows, run the free tier of the Pangram AI detector on a sample document you possess explicit permission to inspect. Then evaluate the highlighted segments against human editorial judgment.