Pangram, a New York-based startup specializing in AI detection, has secured $9 million in funding to address the proliferation of AI-generated content across the internet. This investment reflects a strong belief that the demand for tools capable of differentiating between human-created and AI-generated text will continue to escalate.
This recent capital injection, spearheaded by Menlo Ventures with contributions from Haystack, ScOp, Script Capital, and Cadenza, coincides with Pangram's introduction of its advanced AI text detection model, Pangram 4, and a new AI image detection model, Pangram Image.
Pangram asserts that its latest text detection model achieves over 99% accuracy in identifying AI-assisted and hybrid human-AI content, alongside an enhanced capability to detect AI humanizer programs. While the AI image detector is currently in research preview, a broader release is anticipated in the forthcoming weeks.
Pangram was founded approximately two years ago by Stanford AI and machine learning graduates Max Spero and Bradley Emi. Their initiative emerged in response to the widespread adoption of ChatGPT, which led to a surge in internet bots, AI-generated SEO content, and what Spero describes as “LLM-powered Russian disinformation campaigns and UAE-influenced campaigns on Twitter.”
“Understanding whether content is AI-generated or not is incredibly valuable, particularly for text,” Spero explained to TechCrunch. “It fundamentally alters how readers engage with the material. One needs to know if they should approach it skeptically, watching for hallucinations, or if they can trust it as well-researched content from a human journalist.”
At its core, Pangram’s AI detection system is a sophisticated machine learning model, extensively trained on tens of millions of verified human-authored documents. For each of these, the startup developed a “synthetic mirror,” replicating the original's topic, length, and tone, but generated by a leading Large Language Model (LLM).
“Our model identifies the consistent stylistic differences and choices inherent in AI-generated content, enabling it to determine with high confidence what constitutes AI authorship,” Spero stated, emphasizing that their detector does not depend on copy-paste metadata or embedded watermarks.
Pangram's approach to AI detection extends beyond simply identifying entirely AI-written text; it also aims to differentiate varying degrees of AI assistance. This includes scenarios where individuals author content themselves but then leverage AI for editing or refinement. Spero posits that AI assistance is acceptable, provided the writer transparently discloses its use.
The rise of Pangram coincides with the increasing ubiquity of AI. While some instances, such as a Canadian politician inadvertently reading an AI prompt during a legislative speech, lead to public ridicule, others carry more severe repercussions. For example, lawyers who have presented cases relying on fabricated citations generated by ChatGPT have faced potential sanctions and fines.
This growing backlash is not merely impacting individuals through embarrassment or penalties; it is also beginning to shape institutional policies and regulations.
This year, the open-access archive arXiv implemented a new enforcement policy. It stipulates that submissions showing clear evidence of authors failing to adequately review LLM output—such as hallucinated references or meta-comments like “Would you like me to make any changes?”—may result in a one-year submission ban.
Pangram is not alone in anticipating a heightened demand for AI detection solutions. Competitors such as Winston AI, Originality.ai, Copyleaks, and GPTZero are also pursuing this market, each developing their distinct detection technologies.
Although not flawless, Pangram’s technology has the potential to strengthen resistance against the influx of AI-generated content across digital platforms, legal proceedings, and academic publications.
Pangram is available to users through a $20 monthly web subscription or a Chrome extension. The extension offers real-time labeling of posts on platforms like X, LinkedIn, Substack, Reddit, and Medium, alongside a "feed health score" that quantifies the human versus AI content displayed.
Additionally, Pangram provides its technology through an API. Substack, for instance, has recently incorporated Pangram’s solution to inform readers when their preferred authors utilize AI in crafting newsletters. Spero notes that other API clients encompass Quora, educational institutions, publishers, literary agents, and recruiters.
Spero mentioned that approximately one in 10,000 human documents are mistakenly identified as AI-generated by Pangram’s model, prompting a personal evaluation. The text detection model proved highly impressive, though not without minor imperfections. It readily identified entirely AI-generated news articles from ChatGPT and Claude and was seldom deceived by efforts to humanize AI-produced text through editing. However, Pangram occasionally flagged sentences that were entirely rewritten by a human as AI-authored. Notably, the system was completely resistant to attempts to prompt ChatGPT and Claude to evade AI detection.
Further testing involved providing ChatGPT and Claude with one of my articles for refinement. Pangram assigned it a 13% AI-assisted score, which felt largely accurate, yet the model inconsistently detected subtle word-choice alterations, flagging some while overlooking others. It also incorrectly identified some human-written sentences as AI-assisted. This was particularly noteworthy given that when the identical article, in its original human-authored form, was submitted to Pangram, it received a perfect 100% human score.
Considering that news articles can often possess a somewhat dry tone that might inadvertently resemble AI-generated content, I shifted my testing strategy. I evaluated Pangram using my more distinct and personal Substack newsletter content. I submitted the first half of an article as human-written and then tasked ChatGPT and Claude with replicating my style to complete the second half. Pangram largely demonstrated strong accuracy in distinguishing between the human-authored and AI-generated sections.
My preliminary assessments of Pangram’s novel image detection model yielded similarly impressive results.
Pangram’s AI image detection system is designed to identify AI-generated images across various AI models, distinguishing itself from watermark-dependent methods used by OpenAI or Google DeepMind, which primarily detect their own creations. This system analyzes pixel-level distributions, discerning subtle statistical variances between authentic photographs and AI-generated imagery. Spero highlights the model's capacity to even detect an AI image embedded within a real-world photograph.
During my evaluation, the model effortlessly identified AI-generated visuals, regardless of whether they were photorealistic or cartoon-like. I can also corroborate its ability to detect an AI image embedded within a real-world photograph, evidenced by Pangram’s heatmap clearly highlighting the synthetic element. However, in one isolated case, it erroneously categorized a photograph of an AI-generated image as human content.
Spero clarified that his intention is not for this technology to instigate a “witch hunt” against individuals employing AI for writing, but rather to establish a vital mechanism for counteracting the deluge of low-quality AI-generated content.
“I envision a future where AI content continues to proliferate relentlessly,” Spero concluded. “The rate at which new GPUs are emerging surpasses the birth rate of humans. Without deliberate efforts to prioritize human-generated content, we risk being overwhelmed by AI, which could ultimately obscure any authentic human signal.”
The Editorial Staff at AIChief is a team of professional content writers with extensive experience in AI and marketing. Founded in 2025, AIChief has quickly grown into the largest free AI resource hub in the industry.