Pangram’s $9M Solution to Combat the Onslaught of AI-Generated Content on the Internet

New York-Based AI Detection Startup Pangram Raises $9 Million to Combat AI Slop Infestation

Pangram, a New York-based AI detection startup, recently secured $9 million in funding to tackle the growing issue of AI-generated content flooding the internet. The company is on a mission to develop tools that can distinguish between human-generated and AI-generated text, anticipating a rising demand for such solutions.

The funding round, led by Menlo Ventures and supported by investors like Haystack, ScOp, Script Capital, and Cadenza, coincides with the launch of Pangram’s latest AI text detection model, Pangram 4, and an AI image detection model called Pangram Image.

Pangram boasts that its new text detection model is more than 99% accurate in identifying AI-assisted writing and mixed human-AI content. It can also detect AI humanizer programs with ease. While the AI image detector is currently available through a research preview, Pangram plans to make it more widely accessible in the near future.

Founded by Stanford AI and machine learning graduates Max Spero and Bradley Emi around two years ago, Pangram emerged in response to the proliferation of AI-generated content on the internet. The launch of ChatGPT led to an influx of bots, AI-generated SEO content, and what Spero describes as “LLM-powered Russian disinformation campaigns and UAE-influenced campaigns on Twitter.”

Spero emphasized the importance of being able to differentiate between AI-generated and human-generated content, particularly in written text. Knowing the origin of the content can significantly impact how it is perceived and trusted by readers.

Pangram’s AI detection system is built on a large machine learning model trained on millions of known human documents. The startup created a “synthetic mirror” for each document, replicating the topic, length, and tone but written by an advanced LLM. This approach allows the model to identify stylistic differences and consistent choices made by AI, enabling it to distinguish AI-generated content with high confidence, without relying on metadata or watermarks.

For Pangram, the focus is not only on identifying content solely created by AI but also on differentiating between varying levels of AI assistance. Spero believes that AI assistance can be acceptable as long as the writer discloses its use.

The rise of AI usage has led to several high-profile incidents where AI-generated content was mistakenly presented as human-created, resulting in ridicule or legal consequences. Institutions like the open-access archive arXiv have implemented stricter policies to prevent the submission of AI-generated content disguised as human work.

Pangram is not alone in recognizing the growing demand for AI detection solutions. Competitors like Winston AI, Originality.ai, Copyleaks, and GPTZero are also developing their own detection tools to address the issue of AI-generated content.

While Pangram’s technology is not flawless, it has the potential to support efforts to combat the influx of AI-generated content across various platforms. The company offers its services through a subscription model and a Chrome extension that provides real-time labeling of posts on platforms like X, LinkedIn, Substack, Reddit, and Medium.

In addition to individual users, Pangram also provides its technology through an API, with clients including Substack, Quora, educational institutions, publishers, agents, and recruiters.

Assessing Pangram’s Effectiveness

Max Spero stated that Pangram’s model incorrectly labels approximately one in 10,000 human documents as AI-generated. To evaluate the model’s accuracy, a series of tests were conducted.

The text detection model demonstrated impressive performance, easily identifying AI-generated news articles and remaining vigilant against attempts to modify the content to sound more human-like. However, there were instances where the model flagged sentences as AI-generated even when they were human-written. Overall, Pangram’s text detection model showed promising results in distinguishing between AI and human-generated content.

Similarly, Pangram’s image detection model proved to be effective in detecting AI-generated images across different models. By analyzing pixel-level distributions, the model can identify subtle statistical differences between real and AI-generated images. It can even detect AI images embedded within real-world photos.

Spero emphasized that while Pangram’s technology aims to address the issue of AI-generated content, it is not intended to target individuals using AI for writing. Rather, it serves as a mechanism to uphold the integrity of human-created content in the face of increasing AI proliferation.

Looking ahead, Spero envisions a future where AI content continues to grow exponentially. He believes that without actively promoting human-generated content, AI-generated content could overshadow authentic human contributions.