Pangram: AI Content Detection Platform and Pangram 4 Classifier

A comprehensive synthesis of Pangram (Pangram Labs), the enterprise AI content and text detection platform, its foundational classifier architecture, and the Pangram 4 Technical Report (arXiv:2607.271

Pangram Labs (Brooklyn, NY, USA; Founded 2023 by Max Spero and Bradley Emi) Platform: pangram.com | Python SDK: pypi.org/project/pangram-sdk | GitHub: pangramlabs/pangram-sdk Technical Reports:

• Pangram 4 Technical Report (arXiv: 2607.27183, 29 July 2026) by Ben Glickenhaus, Katherine Thai, Jenna Russell, Elyas Masrour, Yue Han, Max Spero, Bradley Emi

• Technical Report on the Pangram AI-Generated Text Classifier (arXiv: 2402.14873, February 2024)

---

Executive Summary

The rapid diffusion of frontier Large Language Models (LLMs)—including OpenAI's GPT-4o and o-series reasoning models, Anthropic's Claude 3.5/3.7, Google's Gemini 1.5/2.0, Meta's Llama 3 series, and DeepSeek V3/R1—has precipitated a profound crisis of authenticity across academic research, peer review, education, and clinical communication. Modern LLMs generate syntactically flawless, contextually plausible text that completely circumvents traditional n-gram matching and verbatim plagiarism engines (such as Turnitin, iThenticate, or standard string search).

Early heuristics-based AI detectors (e.g., measuring scalar perplexity, burstiness, or raw entropy) suffered from severe fatal flaws:

1. Unacceptable False Positive Rates (FPR): Incorrectly penalizing human writers, particularly non-native English speakers (ESL) whose stylistic patterns mimic lower-perplexity synthetic corpora.

2. Vulnerability to Paraphrasing and "Humanization": Easily bypassed by commercial adversarial tools (QuillBot, Undetectable AI, WriteHuman, StealthGPT) that swap synonyms and perturb sentence lengths.

3. Binary Coarseness: Forcing documents into an all-or-nothing binary classification, failing to capture modern collaborative workflows where a human author drafts an idea and uses an LLM for formatting, copy-editing, or grammar polishing.

Pangram (developed by Pangram Labs) is a deep-learning-based content authenticity and AI text classification ecosystem designed to resolve these challenges. Its flagship architecture, Pangram 4 (introduced in July 2026; arXiv:2607.27183), establishes an empirical benchmark for production detection:

• Ultra-Low False Positive Rate: Achieves an FPR of 0.0041% (approximately 1 false positive per 24,390 human documents), establishing an empirical standard for high-stakes decision-making.

• High Sensitivity: Maintains a False Negative Rate (FNR) of 0.3396% (compared to 1.99% in Pangram 3) with an overall AUROC of 0.9916 on English long-form prose.

• Granular 3-Tier Classification: Classifies text into Human-Written, AI-Assisted, and AI-Generated, with character- and token-level span attribution.

• Adversarial & Humanizer Robustness: Retains 98.83% detection accuracy across 13 major commercial text humanizers.

• Multilingual Scope: Calibrated across 20+ languages including English, Arabic, Chinese, French, German, Italian, Japanese, Portuguese, Russian, Spanish, and Turkish.

• Multi-Modal Extension: Integrates a research preview for AI Image Detection, identifying synthetic visuals from diffusion and autoregressive generators (Flux.1, Midjourney, Stable Diffusion 3, DALL-E 3).

---

Architectural Blueprint & Detection Mechanics

┌──────────────────────────────────────────────────────────────────────────────────────────────────┐
│                                PANGRAM 4 DETECTION PIPELINE                                      │
└──────────────────────────────────────────────────────────────────────────────────────────────────┘

 [ Input Document (Text / PDF / Word / HTML) ]
                      │
                      ▼
 ┌─────────────────────────────────────────────────────────────────────────────────────────────┐
 │ 1. TEXT NORMALIZATION & PREPROCESSING                                                       │
 │    • Zero-width character stripping, homoglyph normalization, encoding cleanup             │
 │    • Language detection & length validation (optimal: ≥50 words)                            │
 │    • Boundary-preserving semantic chunking & sentence tokenization                          │
 └──────────────────────────────────────────────┬──────────────────────────────────────────────┘
                                                │ Clean token stream
                                                ▼
 ┌─────────────────────────────────────────────────────────────────────────────────────────────┐
 │ 2. DEEP TRANSFORMER FEATURE EXTRACTOR (6× Parameter Scale-Up vs. Pangram 3)                 │
 │    • Bidirectional multi-layer transformer backbone (Mixture-of-Experts routing)           │
 │    • Captures long-range statistical syntax, stylistic cadence, and token-transition priors│
 │    • Active-learning hard negative mining (adversarial prompts, reasoning traces, ESL text) │
 └──────────────────────────────────────────────┬──────────────────────────────────────────────┘
                                                │ High-dimensional contextual embeddings
                                                ▼
 ┌─────────────────────────────────────────────────────────────────────────────────────────────┐
 │ 3. DUAL-BRANCH INFERENCE & ATTRIBUTION ENGINE                                               │
 │    ├── A. DOCUMENT-LEVEL CLASSIFIER:                                                        │
 │    │      • Calibrated probability score: P(AI), P(Assisted), P(Human)                      │
 │    │      • Conservative threshold gating (tuned for FPR ≤ 0.0041%)                         │
 │    │                                                                                        │
 │    └── B. "REPEAT2" TOKENWISE & SPAN BOUNDARY DETECTOR:                                     │
 │           • Sliding-window per-token scoring with dense boundary refinement                │
 │           • Identifies interleaved AI edits within predominantly human writing             │
 └──────────────────────────────────────────────┬──────────────────────────────────────────────┘
                                                │
                                                ▼
 ┌─────────────────────────────────────────────────────────────────────────────────────────────┐
 │ 4. STRUCTURED PREDICTION & VERIFIABLE AUDIT REPORT                                          │
 │    • Overall Label: "Human-Written" | "AI-Assisted" | "AI-Generated"                       │
 │    • Document AI Percentage (0.0% – 100.0%)                                                 │
 │    • Color-Coded Heatmap: green (human), yellow (AI-edited/assisted), orange/red (AI)       │
 └─────────────────────────────────────────────────────────────────────────────────────────────┘

Architectural Principles

1. Classifying Statistical Signatures, Not Surface Keywords:

LLMs generate text by sampling from conditional probability distributions over token vocabularies. Regardless of prompt temperature or system persona, autoregressive generators leave characteristic artifacts: specific distributional uniformity, predictable syntactic transitions, and an absence of genuine idiosyncrasy. Rather than searching for "telltale AI phrases" (e.g., "delve", "testament", "tapestry"), Pangram's deep encoder models the latent probability manifolds of human vs. synthetic language across diverse registers.

1. The "Repeat2" Boundary Detection Algorithm:

In collaborative writing, authors frequently compose an original draft and instruct an LLM to rewrite a single paragraph, add a transition, or rephrase a conclusion. Early classifiers either missed localized edits or flagged the entire manuscript. The Repeat2 mechanism operates tokenwise across overlapping contextual windows, calculating differential likelihood ratios to localize precise edit boundaries without bleeding false positives into surrounding human prose.

1. Multi-Model Benchmark Validation:

Pangram is continuously benchmarked against 26+ active frontier model families (OpenAI GPT-4o/o1/o3/GPT-5; Anthropic Claude 3.5 Sonnet / 3.7 / Opus; Google Gemini 1.5 Pro / Flash / 2.0; Meta Llama 3/3.1/3.3; Mistral Large; DeepSeek V3/R1; xAI Grok). In internal evaluations on millions of synthetic and natural documents, Pangram 4 detects synthetic samples across these families with 99.7% accuracy.

---

Technical Performance & Empirical Benchmarks

1. Accuracy vs. False Positive Trade-Off (English Prose)

The table below contrasts Pangram's performance progression from Pangram 3 to Pangram 4 based on the July 2026 technical report (arXiv:2607.27183):

Metric	Pangram 3.3	Pangram 4 (Ours)	Absolute Delta / Improvement
AUROC	0.9842	0.9916	$+0.0074$
False Positive Rate (FPR)	0.012% (~1 in 8,300)	0.0041% (~1 in 24,390)	$2.9\times$ reduction in false accusations
False Negative Rate (FNR)	1.99%	0.3396%	$5.8\times$ reduction in missed detections
Humanizer Defense Accuracy	87.4%	98.83%	$+11.43\%$ across 13 tools
Model Size / Parameter Capacity	Baseline ($1\times$)	$6\times$ Parameter Scale	Enhanced expressive representation
Output Granularity	Binary + Windows	Human / Assisted / AI	Discrete 3-tier semantic labeling

2. Resistance to Commercial "Humanizers" and Paraphrase Engines

Commercial humanization engines attempt to evade detection through intentional perturbations: inserting orthographic noise, swapping rare synonyms, forcing irregular sentence structures, and round-trip translation. On benchmark testing against 13 leading evasion services:

[ Pure AI Text (e.g. GPT-4o) ] ──────────► Passed through Commercial Humanizers
                                                        │
                 ┌──────────────────────────────────────┴──────────────────────────────────────┐
                 ▼                                                                             ▼
       Legacy AI Detectors                                                            Pangram 4 Classifier
    (Accuracy drops to 20% - 45%)                                                  (Maintains 98.83% Accuracy)
  (Vulnerable to perturbed syntax)                                              (Invariant to superficial noise)

Pangram's training corpus includes active adversarial data mining, training the MoE encoder to remain invariant to superficial noise while tracking core syntactic entropy and semantic coherence.

---

Multi-Modal Detection: AI-Generated Images

In addition to text, Pangram Labs has deployed a research preview for AI Image Detection on pangram.com:

• Target Modalities: Identifies images generated by latent diffusion models (Stable Diffusion XL, SD3, Flux.1), cascaded diffusion architectures (Midjourney v5/v6, Imagen 3), and generative adversarial networks (GANs).

• Artifact Analysis: Evaluates spatial frequency distributions, Fourier transform power spectra, unnatural lighting consistency, anatomical micro-structures, and latent noise residual patterns.

• Biomedical Relevance: High utility for scientific publishing integrity, screening Western blot manipulations, synthetic micrograph fabrications, and simulated histopathology patches (e.g., stress-testing synthetic pipelines like [[HistoGen: Histopathology Cell Nuclei Image Generation Tool]]).

---

Developer Integration: Python SDK & REST API

Pangram provides a native Python SDK (pangram-sdk on PyPI) and REST API for high-throughput batch auditing.

Installation

pip install --upgrade pangram-sdk

Self-Contained Python Inference Example

import os
import time
from pangram import Pangram

# Initialize client using environment variable PANGRAM_API_KEY
client = Pangram(api_key=os.environ.get("PANGRAM_API_KEY"))

text_to_audit = """
Histopathological evaluation was performed on formalin-fixed paraffin-embedded 
sections stained with haematoxylin and eosin. The neoplastic proliferation 
demonstrated marked nuclear pleomorphism, hyperchromasia, and an infiltrative 
cribriform glandular architecture invading into the deep muscularis propria.
"""

# Submit asynchronous text classification task
task = client.text.create_task(
    text=text_to_audit,
    model="pangram-4"  # Defaulting to Pangram 4
)

print(f"Task submitted with ID: {task.id}")

# Poll for completion
while True:
    status = client.text.get_task(task.id)
    if status.status in ("COMPLETED", "FAILED"):
        break
    time.sleep(0.5)

if status.status == "COMPLETED":
    result = status.result
    print(f"Prediction: {result.label}")              # e.g., 'human', 'ai_assisted', 'ai_generated'
    print(f"AI Probability: {result.ai_score:.4f}")   # 0.0 to 1.0
    print(f"Assistance Score: {result.assisted_score:.4f}")
    
    # Inspect fine-grained segment predictions
    print("\nSegment Breakdown:")
    for segment in result.segments:
        print(f"[{segment.label.upper()}] ({segment.start}:{segment.end}): {segment.text.strip()}")
else:
    print(f"Audit failed: {status.error}")

Ecosystem Integrations

• Learning Management Systems (LMS): Direct LTI integrations for Canvas LMS, Blackboard, and Google Classroom.

• Browser Extensions: Dedicated Chrome and Firefox add-ons allowing instantaneous one-click audits across web forms, Google Docs, and social feeds.

• Enterprise Workplace: Native Gmail integration and webhooks for publishing workflow automation.

---

Implications for Medicine, Pathology & Scientific Integrity

1. Combatting Synthetic Paper Mills & Fabricated Research

The rise of generative AI has fueled automated "paper mills" producing entirely fabricated scientific literature, complete with convincing case reports, synthetic patient cohorts, and plausible histopathological narratives.

• Pangram serves as an automated first-line triage filter for medical journal editorial offices and peer-review portals.

• Screens submitted manuscripts for uncredited AI drafting, hallucinated citations, and synthetic discussion sections before peer review.

2. Clinical Documentation & Boundary Objects

In [[The pathology report as a boundary object: From clinical communication to computational representation]], Dr. Shuoshuo Wang emphasizes that pathology reports are delicate boundary objects requiring precise semantic grounding.

• When pathologists or residents utilize LLMs to format synoptic surgical reports, distinguish between:

  • Legitimate Formatting Assistance: Correcting grammar, standardizing synoptic formatting, or expanding acronyms.

  • Illegitimate Narrative Fabrication: Permitting an LLM to generate diagnostic impressions or interpolate unobserved microscopic findings.

• Pangram’s per-segment attribution enables institutional auditing: verifying that diagnostic syntheses remain human-authored while administrative scaffolding may be computationally assisted.

3. Protecting Non-Native English Researchers (ESL Equity)

A major ethical risk of legacy AI detection is the disproportionate flagging of non-native English writers whose formal, repetitive phrasing mimics low-entropy LLM text.

• Pangram 4’s low false positive rate (0.0041%) and training on diverse international scientific corpora significantly minimize wrongful accusations.

• Distinguishes benign language editing (e.g., using an LLM to refine spelling, syntax, or phrasing) from substantive intellectual generation of scientific hypotheses.

4. Algorithmic Due Process & The "Gold Standard Paradox"

As detailed in [[The Gold Standard Paradox in Digital Image Analysis: Manual Versus Automated Scoring as Ground Truth]] and [[Cognitive biases in AI-assisted medical decision making: A structured review as a primer for veterinary and human pathology]]:

• No Detector is an Infallible Judge: Despite an AUROC of 0.9916 and FPR of 0.0041%, statistical classifiers cannot establish legal or moral culpability in isolation.

• Mandatory Human-in-the-Loop Review: AI detector output must be treated as probabilistic evidentiary indicators, never as autonomous punitive triggers. Institutions must preserve editorial recourse, student defense, and expert human adjudication.

---

Comparative Matrix: Pangram vs. Alternative Detection Approaches

Feature / Dimension	Pangram 4	Turnitin AI Writing Detector	GPTZero	Open-Source Perplexity (e.g. Binoculars)
Model Architecture	Deep Transformer MoE Classifier	Proprietary Transformer	Ensemble Transformer	Perplexity / Cross-Perplexity Ratio
Document FPR	0.0041% (1 in ~24,000)	~1.0% (1 in 100)	~1.0% – 2.0%	Highly variable ($>3.0\%$)
Classification Tiers	Human / AI-Assisted / AI	Binary (AI % score)	Binary + Sentence highlight	Continuous metric threshold
Paraphrase / Humanizer Robustness	98.83%	Moderate (vulnerable to QuillBot)	Moderate	Very Low (collapses upon synonym swap)
Reasoning Model Detection (o1/o3/R1)	Yes (calibrated on traces)	Variable	Partial	Low
Multilingual Support	20+ languages	English only	Primarily English	Multilingual if base model is
Image Detection	Yes (Research Preview)	No	No	No
Developer API & Python SDK	Yes (pangram-sdk)	Enterprise LTI only	Yes	Self-hosted code

---

Related Notes & References

• Scientific Integrity & Writing: [[Plagiarism]], [[Text Editing]], [[Authorship]], [[Reproducibility]], [[Selecting a Journal]]

• AI Ethics & Cognitive Oversight: [[What AI Can and Cannot Do in Pathology]], [[Cognitive biases in AI-assisted medical decision making: A structured review as a primer for veterinary and human pathology]], [[When Two Wrongs Don't Make a Right: Examining Confirmation Bias and the Role of Time Pressure During Human-AI Collaboration in Computational Pathology]], [[Ethical guidelines for deploying artificial intelligence applications in the pathology field: Lessons learned from a prospective framework in a large tertiary care academic medical center]]

• Representational Theory & Quality: [[The pathology report as a boundary object: From clinical communication to computational representation]], [[The Gold Standard Paradox in Digital Image Analysis: Manual Versus Automated Scoring as Ground Truth]], [[NPIC Quality Coordination Centre: Digital Pathology Quality Assurance and Metrology]]

• Computational Tools & Generative AI: [[askLLM]], [[HistoGen: Histopathology Cell Nuclei Image Generation Tool]], [[Micro-Manager]]