Pangram Labs explained whether AI text can be reliably distinguished from human text

Pangram Labs is attempting to solve a task that becomes increasingly challenging with each new generation of models: determining whether text was written by a human or generated by AI. The company's head, Max Spero, explained in the Odd Lots podcast how such detectors work and why widespread AI-generated writing is already changing the very structure of the internet.

How They Search for Traces

According to Spero, the task of detection is not about finding a single 'secret marker,' but rather evaluating a set of characteristics that are more frequently found in synthetic text. These include predictability of phrasing, repetitive structures, excessively uniform tone, and overall statistical similarity to responses from large language models.

These systems do not read text like a literary critic. They attempt to measure the probability that the material before them was assembled by a machine from the most typical and safe language templates.

This is particularly important now, when models have learned to write noticeably better than they did a year ago. While early AI-generated texts easily gave themselves away through dryness and clichés, newer versions can imitate natural speech, add rhythm, and even maintain an author's intonation.

Therefore, modern detectors function more like a probabilistic filter. They are more useful on large volumes of content—for example, when checking thousands of articles, reviews, or posts—than as an absolute verdict on a single short paragraph.

Where This Will Be Useful

Interest in such tools is not limited to academic verification. AI-generated text has quickly occupied a space where editors, copywriters, SEO teams, and moderation services were previously required. The internet is increasingly filled with automatically generated pages, product cards, pseudo-expert advice, and clones of news articles. For platforms and publishers, the question is no longer whether such content exists, but how to distinguish useful material from cheap content noise that clogs search results and erodes trust.

• Verification of academic and competitive work
• Filtering of SEO spam and content farms
• Moderation of reviews, comments, and submissions
• Verification of materials in editorial offices and marketplaces
• Risk assessment for branded content publication

However, the mere fact of using AI does not make the text bad. For many teams, it is already a common working tool: the model helps assemble a draft, reduce research time, or rewrite a passage in a specific style. The problem begins where synthetic content is masked as an independent opinion, human experience, or original expertise. This is why the discussion about detectors quickly goes beyond technical verification and comes down to a question of transparency.

Limits of Detection

The main difficulty is that the boundary between 'human' and 'machine' text is blurring. An author can take an AI draft, substantially rewrite it, add facts, remove templates, and make the material truly their own. The opposite situation is also possible: a person writes dryly, monotonously, and predictably, causing an automated system to mistakenly increase the probability of AI authorship. Therefore, any such tool inevitably exists in a world of false positives, disputed cases, and gray zones.

Against this backdrop, the future of the internet is increasingly discussed not as a battle between humans and machines, but as a struggle for trust in content. If a significant portion of texts, reviews, answers, and instructions will be released automatically, platforms will need to build additional layers of verification: origin labels, editing history, reputation signals, rules for disclosing AI use, and stricter moderation of content spam networks. Detectors like Pangram Labs' solution in such a framework look not like a final answer, but as one element of a broader trust infrastructure.

What Does This Mean

Tools for recognizing AI-generated text are becoming a separate market because the internet is rapidly being filled with synthetic content. The winner will not be the one who finds the perfect detector, but rather the one who best integrates text origin verification into editorial, educational, and platform processes.