Sonny Valindra
New York-based AI detection startup Pangram has successfully closed a $9 million funding round, signaling strong investor confidence in its mission to combat the escalating issue of AI-generated content polluting the internet. The funding, led by Menlo Ventures with significant contributions from Haystack, ScOp, Script Capital, and Cadenza, underscores a growing market demand for tools capable of distinguishing between human and artificial intelligence authorship. This strategic investment coincides with Pangram’s launch of its next-generation AI text detection model, Pangram 4, and a new AI image detection model, Pangram Image, positioning the company at the forefront of this critical technological arms race.
The proliferation of AI-generated content, often referred to as "AI slop," presents a multifaceted challenge across digital platforms. From SEO spam and deceptive marketing to misinformation campaigns and academic integrity concerns, the pervasive nature of AI-generated text and images necessitates robust detection mechanisms. Pangram’s latest advancements aim to address these growing threats head-on.
Pangram 4, the company’s updated text detection model, boasts an impressive accuracy rate exceeding 99% in identifying AI-assisted writing and content that blends human and AI authorship. Crucially, this iteration demonstrates enhanced capabilities in detecting AI "humanizer" programs, which are designed to obscure the artificial origins of text. Simultaneously, Pangram has introduced Pangram Image, a dedicated AI image detection model. While currently in a research preview phase, its wider release is anticipated in the coming weeks, promising to bring similar discernment to the visual realm.
The genesis of Pangram can be traced back approximately two years, to the advent of large language models like ChatGPT. Max Spero and Bradley Emi, both Stanford graduates with backgrounds in AI and machine learning, recognized the immediate implications of this technology’s widespread adoption. The launch of these powerful AI tools opened the floodgates to an internet increasingly populated by automated content, including what Spero describes as "LLM-powered Russian disinformation campaigns and UAE-influenced campaigns on Twitter." This surge in AI-generated output created a clear and present need for reliable methods to verify content authenticity.
"I think it’s just incredibly valuable to know whether what you’re looking at is something that’s AI-generated or not," Spero stated in a recent interview. "Especially text that you’re reading, because it changes how people approach the text. Is this something that I’m going to have to look out for hallucinations and jump in skeptically, or is this something that I trust was well-researched from an actual journalist?" This sentiment highlights the profound impact AI-generated content can have on user trust and critical evaluation of information.
Pangram’s AI detection system operates on a sophisticated machine learning framework. The core of its technology involves training a large model on a vast corpus of tens of millions of known human-authored documents. To create a contrasting dataset, the startup developed a "synthetic mirror" for each human document. This synthetic version replicates the original’s topic, length, and tone of voice but is generated by a frontier large language model. This comparative approach allows Pangram’s model to meticulously learn the subtle, yet consistent, stylistic differences and decision-making patterns inherent in AI-generated text. "Our model is learning the stylistic differences and the choices that AI makes consistently and is able to use that to learn what makes something AI-generated with high confidence," Spero explained, emphasizing that their detection method does not rely on superficial metadata or hidden watermarks.
Beyond simply identifying purely AI-generated text, Pangram’s approach acknowledges the spectrum of AI assistance. The company’s detection system is designed to differentiate between various levels of AI involvement, including instances where humans use AI for editing or refinement. Spero suggests that AI assistance can be acceptable, provided that its use is transparently disclosed by the author. This nuanced perspective positions Pangram not as an outright ban on AI tools, but as an advocate for ethical and transparent AI integration.
The timing of Pangram’s funding and product launches is particularly significant, given a series of high-profile incidents that have underscored the potential pitfalls of unchecked AI content generation. Recently, a Canadian politician inadvertently read an AI prompt aloud during a speech to lawmakers, resulting in widespread ridicule. More critically, legal professionals have faced sanctions and fines for relying on AI-generated citations that proved to be fabricated by tools like ChatGPT, demonstrating the real-world consequences of AI-driven inaccuracies in professional contexts.
These incidents are not isolated; they reflect a broader trend of AI-generated content infiltrating various sectors. Academic institutions and research platforms are also grappling with the implications. For example, the open-access archive arXiv has implemented a new policy this year. Submissions found to contain evidence of authors failing to adequately review LLM output—such as hallucinated references or conversational AI prompts embedded within the text—can now result in a one-year submission ban. This policy signals a growing institutional awareness and a proactive stance against the uncritical acceptance of AI-generated scholarly material.
Pangram is not alone in recognizing the burgeoning market for AI detection solutions. Competitors such as Winston AI, Originality.ai, Copyleaks, and GPTZero are also actively developing and refining their own detection technologies, indicating a robust and competitive landscape. However, Pangram’s advanced accuracy rates and its dual focus on both text and image detection position it as a significant player. The company’s technology, while acknowledging it is not infallible, offers a crucial mechanism to counter the overwhelming influx of potentially unreliable AI-generated content across the internet, legal documents, and academic publications.
Pangram offers its services through a tiered subscription model. Users can access its web-based platform for $20 per month. Additionally, a Chrome extension is available for download, which provides real-time AI content labeling across popular platforms like X (formerly Twitter), LinkedIn, Substack, Reddit, and Medium. This extension also furnishes a "feed health score," offering a percentage breakdown of human versus AI-generated content visible to the user. For enterprise-level integration, Pangram provides API access. Notably, Substack has already integrated Pangram’s technology into its platform, enabling readers to identify newsletters authored with AI assistance. Other significant API clients include Quora, educational institutions, publishers, agents, and recruitment firms, reflecting the broad applicability of Pangram’s detection capabilities.
Does Pangram Work? An Empirical Assessment
To gauge the efficacy of Pangram’s AI detection capabilities, a series of tests were conducted. Spero indicated that Pangram’s model incorrectly labels approximately one in 10,000 human documents as AI-generated, a remarkably low error rate for such a complex task.
Initial tests focused on entirely AI-generated news articles produced by both ChatGPT and Claude. Pangram’s text detection model proved highly effective, accurately flagging these articles without issue. Attempts to humanize the AI-generated text through editing were also largely unsuccessful in deceiving the model. Pangram consistently identified the AI origin, even when subtle modifications were made. In some instances, however, the model flagged sentences that had been entirely rewritten by a human as AI-generated, suggesting that certain phrasing styles can inadvertently mimic AI patterns. Importantly, Pangram demonstrated resilience against attempts to prompt ChatGPT and Claude into evading AI detectors, indicating a robust underlying detection methodology.
A further test involved providing ChatGPT and Claude with an existing article and instructing them to "polish" it. Pangram assigned a 13% AI-assisted score to this content, a figure that appeared to be a reasonable approximation of the actual AI involvement. The model successfully detected subtle word-choice alterations in some sentences, while overlooking similar changes in others. Intriguingly, Pangram flagged some human-written sentences as AI-assisted in this scenario. However, when the same article, in its original, entirely human-written form, was submitted, Pangram confidently assigned a 100% human score. This observation suggests that the presence of minor AI edits, even when subtle, can influence the model’s perception, and that the inherent style of certain types of writing, such as news articles, might occasionally lead to misclassifications.
To explore this further, a more stylistic approach was adopted. A personal Substack newsletter, characterized by a distinct voice and personal anecdotes, was split in half. The first half, written by a human, was fed into Pangram, while ChatGPT and Claude were tasked with replicating the author’s style for the second half. In this test, Pangram largely succeeded in distinguishing between the human-written and AI-generated portions, accurately identifying the AI-generated text.
The new Pangram Image detection model, while still in preview, also showed promising results. The system is designed to identify AI-generated images across various AI models, a notable advantage over watermark-dependent solutions like those from OpenAI or Google DeepMind, which primarily detect their own output. Pangram Image analyzes pixel-level distributions and learns subtle statistical disparities between authentic photographs and AI-generated visuals. According to Spero, the model can even detect AI-generated imagery embedded within real-world photographs.
During testing, Pangram Image effectively identified both photorealistic and cartoonish AI-generated imagery. The model’s ability to detect an AI image embedded within a real-world photo was confirmed, with the provided heat map clearly highlighting the artificial elements. However, in one specific instance, a photograph of an AI-generated image was misclassified as human content, indicating that the model, like its text counterpart, is still undergoing refinement.
Spero expressed a clear vision for Pangram’s role: not to instigate a "witch hunt" against AI users, but to provide a necessary countermeasure against the escalating volume of low-quality, AI-generated content. "The future that I see is that AI content just continues to proliferate," Spero predicted. "We’re getting new GPUs faster than new people are being born. If we do not actively discriminate in favor of human content, then we’re just gonna see more and more AI, and it’s just gonna drown out any human signal that we have." This outlook underscores the critical need for tools that can help preserve the integrity and authenticity of the digital information landscape. The substantial funding secured by Pangram and its ongoing technological advancements suggest that the market is keenly aware of this imperative.
