Pangram 4

Model Description

Pangram 4 is an AI detection model released by Pangram Labs on July 29, 2026. It is the successor to Pangram 3.3.2, with significant improvements to overall accuracy, robustness to adversarial attacks, and the more fine-grained detection of AI-assisted and mixed-authorship text.

As with the Pangram 3 series, Pangram 4 uses technology from the EditLens technical report. However, Pangram 4 expands on the EditLens work substantially, introducing an MoE backbone; humanizer, tokenwise, and mixed-authorship prediction heads; and a two stage training process.

Check out the full announcement on the blog here, the technical write-up here, and the complete technical report here.

Inputs

Pangram 4 accepts text inputs of at least 50 words. It is designed for natural-language prose written in complete sentences.

Outputs

Pangram 4 returns a complete, non-overlapping sequence of variable-length segments covering the analyzed text. The model predicts at the token level and decodes one of three provenance labels for each segment: Human, AI-Assisted, or AI-Generated.

Segment-Level Outputs

Document-Level Outputs

Model Resolution

Pangram 4 produces substantially finer-grained boundaries than Pangram 3, with a minimum predicted segment length of around two sentences.

Supported Languages

Pangram can work effectively on languages outside the official support set due to generalization. Please contact us to request an evaluation on another language.

Model Training

Training Datasets

Pangram's human-written corpus contains long-form prose from a wide variety of domains, including essays, creative writing, reviews, books, news articles, scientific papers, professional writing, and general web text. All training data is licensed for commercial use. Pangram 4 is not trained on user-submitted text or customer API data.

Pangram's AI-generated corpus is produced in-house using the synthetic mirroring technique introduced in the original Pangram Technical Report by Emi and Spero (2024). AI-assisted training examples are created by applying AI edits to human-written documents. Pangram uses lexical and semantic alignment between the original and edited text to assign tokenwise Human, AI-Assisted, and AI-Generated labels to the training data.

Architecture

Pangram 4 uses a causal, open-weight sparse mixture-of-experts language model as its shared backbone. Four linear classification heads are attached to the shared representation:

• +A 15-class segment-level head estimates the overall level of AI involvement in a text window.
• +A three-class tokenwise provenance head predicts Human, AI-Assisted, or AI-Generated for every token.
• +A binary mixed-authorship head estimates whether the the segment contains text of mixed provenance.
• +A four-class humanizer head distinguishes Human, AI-Generated, AI-Edited, and Humanized-AI text.

The humanizer head is trained as a stop-gradient probe so that its objective does not update the shared backbone.

Pangram 4 is trained in two stages using LoRA. The first stage trains the segment-level and humanizer objectives. The first adapter is then merged into the backbone, a new adapter is initialized, and training continues with the tokenwise provenance and mixed-authorship objectives alongside the segment-level and humanizer objectives from stage 1.

Tokenwise training uses a method called Repeat2. Each 512-token training window is repeated twice. The tokenwise loss is applied only to the second copy, allowing every supervised token to use context from the complete source window despite the causal backbone.

For more details on model training, please see Sections 4.1 and 4.2 in the Pangram 4 Technical Report.

Model Inference

Preprocessing

Pangram removes adversarial Unicode characters and normalizes mixed-script homoglyph attacks before tokenization. File inputs are converted to text before inference.

Inference

Pangram 4 has a 512-token inference window, so longer documents are divided into overlapping windows. Repeat2 is applied to each window, and predictions for tokens that appear in multiple windows are aligned and averaged.

The model produces tokenwise, segment, and mixed-authorship, and humanizer predictions for every window. A calibrated three-state linear-chain conditional random field combines the first three observations across the full document (the humanizer evidence does not contribute). Viterbi decoding computes the most likely label sequence, while a forward-backward pass computes token-level class probabilities.

Postprocessing

Product output is constrained to sentence-level resolution, so tokens are assigned to their respective sentences by character midpoint, then sentences inherit the majority label of their tokens. An additional merging step is conducted to ensure that contiguous label sequences are of a minimum length (which we have currently tuned to around 2 sentences). The resulting non-overlapping segments cover the full input document. Finally, document fractions and headline document results are calculated from the final labeled segments.

For additional details on model inference and postprocessing, please refer to Section 4.3 in the Pangram 4 Technical Report.

Evaluations

Methodology

For the purposes of evaluation, we use the prediction_short field returned by the Pangram API. For binary evaluation datasets, i.e. datasets containing only fully human-written and/or fully AI-generated texts, a Mixed prediction counts as an error.

Human Datasets — False Positive Rate by Language

At the production operating point, Pangram 4 achieves a 0.0041% false positive rate (roughly 1 in 24,000) on 1,000,000 human-written English FineWeb evaluation examples. Multilingual false positive rates are measured on pre-2022, per-language subsets of FineWeb2.

Overall False Positive Rate

FineWeb2 False Positive Rate by Supported Language

AI Datasets — False Negative Rate by Language + Generator Model

The English evaluation contains successfully scored generations from 26 frontier open-source and closed-weight models. The multilingual evaluation contains synthetic mirrors of English how-to articles, news articles, and reviews across 18 supported languages.

Overall False Negative Rate

False Negative Rate by Generator Model

False Negative Rate by Language

3rd Party Benchmarks

For additional evaluation results on both internal and 3rd party benchmarks, please see Section 5 of the Pangram 4 Technical Report.

Changes from Pangram 3.3

• +Replaces previous algorithm for boundary detection with a tokenwise model and document-level structured decoder
• +Substantially improves heterogeneous-mixed text recall
• +Adds humanizer detection feature that flags AI-generated text that has been perturbed to evade detection

Intended Usage and Limitations

Pangram 4 is intended for long-form natural-language writing samples of at least 50 words, written primarily in complete sentences.

Short conversational replies, answers to questions with a single factual answer, source code, tables of contents, reference sections, templated or automated writing, instructions and technical manuals, and text dominated by mathematical notation are outside the model's primary scope or may be more susceptible to errors.

For best results, human-written headers, footers, instructions, and other extraneous formatting should be removed before checking a document.

Raw text and .docx files are recommended over PDFs when available because PDF parsing can introduce unintended artifacts.

Ethics & Safety

False accusations of AI usage can lead to serious consequences, including reputational damage, emotional trauma, and other undue harm.

We acknowledge that our model has a non-zero error rate and its errors may result in such harms. We commit to continuing to engage with our users and the academic community to educate others on appropriately contextualizing and communicating the results of AI detection software.

We take reports of false positives extremely seriously and work to mitigate their occurrences to the best of our team's ability.

We commit to training and releasing models with the lowest possible false positive rate, and improve our evaluations to continue to thoroughly test and monitor future model releases for regressions.