As artificial intelligence produces a growing share of online text and images, distinguishing human-created material from machine-generated content is becoming increasingly difficult. Pangram, a New York-based AI detection startup, is positioning itself as a solution to that problem.
The company has raised $9 million in a funding round led by Menlo Ventures. Haystack, ScOp, Script Capital, and Cadenza also participated in the investment. Pangram plans to use the funding to improve its detection technology and expand access to its tools.
The announcement comes alongside the launch of Pangram 4, the company’s latest text detection model, and Pangram Image, a new system designed to identify AI-generated images.
Generative AI tools have made it possible to create articles, marketing copy, social media posts, academic material, and images within seconds. While these systems can improve productivity, they have also contributed to a sharp increase in low-quality, misleading, and automated content.
This expansion has created challenges for publishers, schools, recruiters, online platforms, and readers. Organizations increasingly need ways to determine whether content was written by a person, produced entirely by AI, or created through a combination of human and machine input.
Pangram was founded by Stanford AI and machine learning graduates Max Spero and Bradley Emi. The founders launched the company after recognizing that generative AI would make online authorship more difficult to verify.
Rather than treating every use of AI as inappropriate, Pangram aims to provide greater transparency. Its technology is designed to identify different levels of AI involvement, including cases where a person writes an original document and later uses an AI assistant to edit or polish it.
Pangram says its new text detection model can identify fully AI-generated writing, AI-assisted material, and content that has been modified using tools intended to make machine-written text sound more human.
The company claims Pangram 4 achieves accuracy of more than 99 percent when detecting AI-assisted and mixed human-AI content. As with any detection system, however, results may not be perfect in every situation.
Pangram trains its technology using tens of millions of documents known to have been written by humans. It then creates AI-generated versions that match the topic, length, and tone of those documents.
By comparing the human material with its synthetic counterpart, the model learns to recognize patterns associated with machine-generated writing. These patterns can include stylistic decisions, word choices, sentence structures, and other subtle signals.
Pangram says its system does not depend on hidden watermarks or copy-and-paste metadata. This could allow the detector to evaluate content even when it was created using AI systems that do not provide built-in identification markers.
Pangram is also moving beyond written content with the introduction of Pangram Image. The image detection tool is currently available as a research preview, with a wider release expected later.
The system analyzes pixel-level patterns to distinguish authentic photographs from AI-generated visuals. Pangram says this approach can work across images produced by different generative AI models.
That differs from watermark-based detection systems, which may be most effective when identifying content created by a specific company’s own tools.
Pangram Image is also designed to detect generated visuals that appear inside real photographs. For example, the system may be able to identify an AI-created picture displayed on a screen, poster, or printed page within an otherwise authentic photograph.
This capability could become increasingly important as synthetic images become more realistic and easier to distribute through news sites, social platforms, and private messaging services.
Pangram offers a web subscription priced at $20 per month, as well as a browser extension that labels content while users browse major online platforms.
The extension can analyze posts on services such as X, LinkedIn, Substack, Reddit, and Medium. It also provides a feed health score showing the estimated proportion of human and AI-generated content appearing on a user’s screen.
For larger organizations, Pangram provides access through an application programming interface. This allows businesses and platforms to integrate the detection system directly into their existing products and workflows.
Substack has already integrated Pangram’s technology to provide information about possible AI use in newsletters. The company also works with customers in publishing, education, recruitment, and online content platforms.
Schools and universities could use the technology to review student submissions, while recruiters may use it to assess applications and written assignments. Publishers and literary agents may apply similar tools when evaluating articles, manuscripts, and proposals.
Pangram is entering a growing market that already includes companies such as GPTZero, Copyleaks, Originality.ai, and Winston AI.
Demand for these products is rising, but AI detection remains technically challenging. Human writing can sometimes appear formulaic, particularly in professional, academic, or news-related contexts. At the same time, AI-generated text can be edited until it closely resembles natural human writing.
These conditions create the risk of false positives, where genuine human work is incorrectly classified as AI-generated. They can also produce false negatives, where machine-generated content avoids detection.
For this reason, detection scores should generally be treated as indicators rather than unquestionable proof. Institutions using these tools may need additional review processes before making serious decisions about a writer, student, employee, or applicant.
The debate over AI-generated content is gradually shifting from whether people should use AI to how that use should be disclosed.
AI tools are already part of many writing and editing workflows. A journalist may use AI to organize research, a student may use it to improve grammar, and a business may use it to prepare an early draft.
The central issue is often whether readers, employers, teachers, or customers understand how the final material was produced.
Pangram’s approach reflects this changing environment. Instead of focusing only on identifying content written entirely by AI, the company is attempting to measure different degrees of machine involvement.
As generated content continues to spread, tools that support transparency may become an important part of the internet’s trust infrastructure. Pangram’s new funding gives it an opportunity to compete for that role, but its long-term success will depend on whether its technology can remain accurate as AI models and human editing techniques continue to evolve.
Share your thoughts about this article.
Be the first to post a comment!
