Pangram's AI Detection Tool: A New Kind of Gatekeeper for Written Content

The Rise of an Unlikely AI Detective

When I first heard about Pangram, it was in the context of a literary scandal—Mia Ballard's novel Shy Girl, allegedly 78 percent AI-written according to a score from this Brooklyn-based startup. The book was ultimately canceled by Hachette, and the incident quickly became a lightning rod for broader concerns about how artificial intelligence is reshaping the publishing landscape.

Pangram, with only 24 employees and a headquarters above a Popeyes in Brooklyn, has raised just $13 million to date—a fraction of what OpenAI has received—but it's already making waves. The company claims its tool can detect AI-generated text with high accuracy, offering a best-guess percentage that helps readers and editors determine whether a piece of writing was created by a human or an artificial intelligence.

What started as an idea born from the founders' discomfort with the growing presence of AI in content has now evolved into a powerful tool used by major publishers, newsrooms, and platforms. As we continue to grapple with the implications of generative AI, tools like Pangram's are becoming crucial players in defining authenticity in our digital age.

How It Works: Synthetic Mirroring and Hard Negative Mining

Pangram's approach to detecting AI-generated content is based on two key methodologies. First, it employs synthetic mirroring, where human writing samples are fed into language models to generate close matches. This teaches the system how AI writes, identifying patterns unique to machine-generated text.

Second, Pangram uses hard negative mining—searching through datasets for false positives and then synthetically mirroring them to augment its training set. It's a clever feedback loop that allows the model to learn from mistakes, improving accuracy over time.

The company also emphasizes proper licensing of its training data—a rare practice in today's AI landscape, especially when compared with giants like OpenAI or Anthropic, whose models were trained on massive public datasets without explicit consent. Pangram's approach, while more limited in scope, reflects a conscious effort toward ethical practices.

Controversy and Criticism: A Tool for Good or Evil?

While Pangram's tool has been embraced by some as a necessary safeguard against deception, others see it as a potential source of harm. Critics argue that the technology can be misused, especially when it comes to false positives or biased outcomes.

"There is such distaste and anger at the AI detection software," says Jane Friedman, an author and publishing expert. "There's this feeling like they are just as evil, if not more evil, than the AI companies themselves."

The controversy deepened when a researcher noted that Pangram's models tend to flag non-native English writers more frequently than native speakers, raising concerns about systemic bias in AI systems.

Despite these issues, Pangram has taken steps to address them. The company states its false positive rate is extremely low—only 0.0041 percent with its latest model—and that it errs on the side of caution when uncertain. This conservative stance means AI-written content may sometimes slip under the radar.

From Scandal to Systemic Concern

The case of Shy Girl was a watershed moment for Pangram. The company's CEO, Max Spero, initially posted his findings on X (formerly Twitter) after receiving a manuscript from a pirating website—a fact that he acknowledged during our interview. Despite the questionable origin of the document, Spero didn't hesitate to run it through Pangram's system.

As the story unfolded across media outlets and social platforms, it became clear that the tool wasn't just a side effect of a single controversy but part of a larger conversation about authorship and transparency in an age where AI can mimic human expression.

Another example is the Commonwealth Short Story Prize winner who was later found to have significant AI content by Pangram. The company's analysis of past winners revealed other potential cases of AI use, further cementing its reputation as a powerful force in the literary world.

Substack Integration and Public Perception

In late July, Substack announced its integration with Pangram, allowing readers to quickly determine whether content was AI-generated. While Substack maintains that its philosophy is not anti-AI, it believes users should know what they're consuming—a sentiment echoed by many who support the idea of transparency in digital communication.

Some authors remain wary, fearing that a single scan could destroy their careers. There's a sense that an entire professional life can be derailed with just one click of a button—an outcome that many find unsettling.

However, the technology itself is evolving. As Spero puts it, Pangram aims to become more granular in its analysis, providing even greater detail about how AI was used—even for light editing or assistance.

The Human Element: Bias, Discretion, and Responsibility

Pangram isn't just about algorithms—it's also about people. When authors are confronted with high scores from the tool, some become defensive, others honest. But the real challenge lies in how these tools are used. As Todd Shuster, co-CEO of Aevitas, noted, “It was instantly a very helpful tool.”

He and his team have started having difficult conversations with authors who've used AI without full disclosure. In some cases, they've asked for rewrites that better reflect the author's own voice and ideas.

The question remains: Who gets to decide what constitutes acceptable AI use? And how do we ensure fairness in such assessments? These are complex questions that extend beyond technology into ethics and equity in creative industries.

Looking Ahead: The Future of Detection Tools

With ongoing developments in AI, the stakes for tools like Pangram continue to rise. As generative models grow more sophisticated, so must detection methods. Pangram's team is already working on improvements, including higher granularity and better contextual analysis.

Spero's vision extends beyond publishing—into education, law, recruitment, and other sectors where authenticity matters. His goal is clear: to become the gold standard for determining originality across industries.

But as these systems become more pervasive, they must also be held accountable. Transparency, ethical training, and community engagement are crucial elements in ensuring that tools like Pangram don't just police AI use but foster healthier, more equitable practices.

Final Thoughts: Trusting the Machines

After our conversation, I couldn't shake off a sense of ambiguity about Max Spero. His demeanor was curious—sometimes distracted, sometimes animated—and yet he remained steadfast in his convictions. Whether it's due to fatigue, media pressure, or something else entirely, there's no doubt that Pangram has entered a new era.

In the end, we're not just dealing with an algorithm—we're navigating a cultural shift in how we define and value human creativity. Tools like Pangram may be the gatekeepers of this transition, but ultimately, they depend on us to determine whether trust is warranted.