How to Tell if Writing Was Generated by AI. A Practical Guide to Detection

AI-generated content is everywhere now, and the ability to spot it has become a genuinely useful skill - for editors, marketers, researchers, and anyone who cares about authenticity online. Max Spero, CEO of Pangram Labs, recently joined Bloomberg's Odd Lots podcast with Tracy Alloway and Joe Weisenthal to dig into exactly how detection software works and what it means for the future of the web.

What Makes AI Writing Different From Human Writing?

There are subtle but consistent patterns in AI-generated text that detection tools are trained to recognize. AI language models tend to produce writing that is statistically "safe" - meaning the word choices are highly predictable given the surrounding context. Human writers take risks. They use unexpected words, make odd stylistic calls, repeat themselves in ways that feel personal rather than mechanical. Pangram Labs built software specifically to measure these statistical patterns at scale, looking at things like token probability distributions - essentially how surprising or unsurprising each word choice is within a sentence. The higher the predictability across a long passage, the more likely it came from a model. One key insight from Spero's discussion is that no single sentence gives AI away - it's the cumulative pattern across hundreds of words that becomes telling. A human might write one very predictable paragraph, but they almost never sustain that level of uniformity across an entire article. That consistency is the signature.

The Technical Reality of AI Content Detection

Detection is not a solved problem - and Spero is clear about that. Current detection tools have meaningful error rates, and false positives are a real concern, particularly for non-native English speakers whose writing can sometimes pattern similarly to AI output. Pangram Labs developed its approach by training classifiers on large datasets of known human and AI writing, looking for signals beyond just vocabulary - things like sentence rhythm, structural uniformity, and semantic flatness. Semantic flatness is particularly interesting: AI models rarely go on tangents, rarely contradict themselves within a piece, and rarely insert the kind of irrelevant-but-human detail that makes real writing feel alive. The challenge intensifies as models improve. GPT-4-class writing is detectably harder to identify than earlier outputs, which means detection tools have to evolve continuously. There's a real arms race dynamic here, and it's not slowing down.

AI Detection Technology Identifies Statistical Patterns in Text. It Does Not Prove Authorship

This is the factual boundary that matters most in this conversation. AI detection tools identify statistical likelihood - they do not and cannot confirm with certainty that a specific human or AI wrote a given piece of text. What Pangram Labs' software does is measure how closely a piece of writing matches the output characteristics of known language models. It does not access metadata, writing history, or authorship records. A high AI-probability score means the text resembles AI-generated content statistically; it does not mean the content was definitively machine-written. The technology also does not assess the quality or accuracy of content - a piece can be entirely human-written and still score poorly on originality metrics. For organizations thinking about content authenticity at scale, tools like JackSEO take a different angle entirely - rather than detecting AI content after the fact, it focuses on producing brand-aligned, SEO and GEO-optimized content that is transparent and purposeful from the start. Detection and creation are separate problems requiring separate thinking.

What Does the Rise of AI Writing Actually Mean for the Internet?

The Odd Lots conversation touched on something that doesn't get enough direct attention: the internet's information ecosystem is changing structurally, not just in volume. When the marginal cost of producing text drops to near zero, the web fills with content - much of it optimized for search ranking rather than genuine usefulness. This is already measurable in certain content categories. The downstream effect is that trust signals are shifting - readers and algorithms alike are developing new ways to evaluate whether content carries real authority or is just statistically competent filler. For content teams navigating this shift, the approach JackSEO takes - analyzing niche trends from trusted sources and aligning output with brand tone and industry context - reflects how serious operations are responding. The question isn't really "was this written by AI?" anymore. The more practical question is "does this content carry genuine signal, or is it noise?" That reframe is where the industry is slowly heading, and detection technology is only one piece of a much larger puzzle.