Pangram

What Pangram does

Pangram is an AI-generated content detection startup that offers a highly accurate AI detector with 99.98% accuracy, independently verified by third-party researchers. It uses natural language processing and a massive dataset of human and AI writing to analyze patterns in text generated by popular models like ChatGPT, Gemini, Grok, Llama, and Claude. The platform also includes comprehensive plagiarism detection, AI segment analysis, and AI assistance detection. Pangram's proprietary technology is built on years of research, trained on diverse datasets using hard negative mining and active learning, and can detect content from all major LLMs across more than 20 languages. It is designed to detect even "humanized" AI text and has a near-zero false positive rate. Pangram offers various products including a web-based AI Detector, browser extensions for Chrome and Firefox, and an API for developers and organizations. It caters to individuals, educational institutions, publishing and media, content moderation, law firms, recruiters, and ML engineers, aiming to rebuild trust amidst the increase in generative AI content.

Products

• AI DetectorA web-based tool to detect AI-generated content with 99.98% accuracy, offering AI segment analysis and AI assistance detection.
• Browser Extension (Chrome & Firefox)Detects AI-generated content on any website, including social media platforms like X, LinkedIn, Substack, Medium, Reddit, and Google Docs. It labels content as human or AI-written and allows users to track their feed health.
• Plagiarism CheckerDetects traditional plagiarism and external content by comparing text against billions of webpages, academic papers, news articles, and books. It can be used simultaneously with the AI detector.
• Multilingual AI DetectionDetects AI-generated and human-written content in over 20 languages with high accuracy, supporting languages like English, Spanish, French, Arabic, Chinese, German, and more.
• AI Detection APIAllows developers and organizations to integrate AI detection into their products, review queues, or content pipelines, offering real-time and bulk processing modes with detailed response fields.
• AI Code DetectorDetects AI-generated code from models like ChatGPT, Claude, and GitHub Copilot in languages such as Python, Java, C++, and more, with a focus on low false positives for developers and engineering teams.

Key capabilities

• 99.98%+ accuracy in AI detection
• 3rd party verified accuracy (University of Maryland, University of Chicago)
• AI Segment Analysis to show how much of a document is AI-generated
• AI Assistance Detection to highlight AI-edited segments
• Comprehensive Plagiarism Checker
• Proprietary technology developed from years of research
• Trained on diverse datasets using hard negative mining and active learning
• Detects content from all major LLMs (ChatGPT, Claude, Gemini, Llama, Grok, GPT-4)
• Supports over 20 languages for AI detection and plagiarism checks
• Outperforms trained human readers in AI detection
• Detects "humanized" text designed to evade detection
• Near-zero false positive rate (1 in 10,000)
• FERPA compliant and SOC 2 Type 2 certified
• API for real-time and bulk processing
• Browser extensions for Chrome and Firefox
• Integrations with LMS systems (Canvas, Brightspace, Moodle, Google Classroom, Schoology, Blackboard)
• Syntax-aware AI code analysis for developers
• Short-form AI detection (as short as 75 words)

Use cases

• Upholding academic integrity in universities and schools
• Grading student assignments with certainty
• Verifying content authenticity in publishing and media
• Protecting editorial integrity and SEO rankings
• Content moderation for trust and safety teams
• Detecting fake reviews and ensuring FTC compliance
• Securing software supply chains by detecting AI-generated code
• Verifying technical hiring candidates' originality
• Protecting intellectual property in codebases
• Mitigating security risks from AI-generated code
• Monitoring social media feeds for AI-generated content
• Researching the proliferation of AI content
• Providing writing feedback to students (e.g., Gradpilot)
• Handling AI-written answers on content platforms (e.g., Quora)
• Screening personal statements in university admissions
• Verifying authenticity of peer reviews and grant applications

AI approach

Pangram develops proprietary AI models for detecting AI-generated content and plagiarism. Their models use natural language processing, trained on massive datasets of human and AI writing, and employ techniques like hard negative mining and active learning to achieve high accuracy and low false positive rates. They specifically detect content from major LLMs like ChatGPT, Gemini, Grok, Llama, and Claude, and can identify 'humanized' text. They also offer AI code detection.

Tech named: natural language processing, neural network, classifier head, Hard Negative Mining, Mirror Prompts, active learning, tokenization, embedding, GPT-4, ChatGPT, Gemini, Grok, Llama, Claude, GitHub Copilot, Python, Java, C++, Go, JavaScript/TypeScript

Industries served

• Education
• Publishing
• Media
• Content Moderation
• Law Firms
• Recruitment
• Software Development
• Trust & Safety
• Academia

What it says sets it apart

• 99.98% accuracy, independently verified by third-party researchers
• Proprietary technology built on years of research, not open-source models or commercial LLMs
• Trained on diverse datasets using hard negative mining and active learning for industry-leading false positive rates
• Detects content from all major LLMs, including those designed to evade detection ('humanized' text)
• Supports over 20 languages with high accuracy
• Outperforms trained human readers in AI detection
• Near-zero false positive rate (1 in 10,000), crucial for sensitive applications
• Combines AI detection with comprehensive plagiarism checking in one tool
• Offers granular analysis, distinguishing between fully AI-generated, AI-assisted, and human-written content
• FERPA compliant and SOC 2 Type 2 certified, ensuring data privacy and security
• Specialized AI code detection tuned for low false positives and syntax awareness
• Ability to detect AI content in short-form text (as short as 75 words)
• Founded by AI researchers from Google, Tesla, and Stanford