Hot take: the least useful question you can ask an AI detector is “is this real or fake?” That’s the question everyone asks, it’s the question the marketing copy answers, and it’s the question that has turned this entire product category into a trust liability for schools, newsrooms, and hiring managers. Pangram’s Max Spero has been making a version of this argument publicly, and having spent a lot of time watching detection tools flop in the wild, I think he’s right in a way that should make the rest of the category uncomfortable.
Because “real or fake” is a binary, and binaries invite people to treat a probability score as a verdict. Nobody reads the confidence interval. They read the red banner.
The two failure modes nobody weighs equally
Every detector has two ways to be wrong. It can miss AI-generated text, or it can flag human writing as machine-made. Those errors do not carry the same cost, and pretending otherwise is how this space earned its reputation.
A missed detection is an annoyance. A false positive is a person getting accused of something they didn’t do, often a student, often an ESL writer whose sentence patterns happen to look statistically tidy. Pangram positions its work around driving those false positives down, and comparison pieces from the company put false positive rates and ESL bias right alongside raw accuracy when stacking up against Turnitin. Whatever you think of a vendor grading its own homework, the framing is the correct one. Accuracy alone is a vanity metric. A detector that catches everything and burns innocent writers is worse than useless, because it manufactures conflict and then hands the human a printout to hide behind.
Why the target keeps moving
Spero has pointed out that most detection tools can still catch text that’s been run through “humanizer” services, the paraphrasing layer people bolt on to sneak past checkers. That’s the reassuring part. The less reassuring part is his other observation: as large language models get more and more research and writing to pull from, detection gets harder. The models are training on a wider, richer pool of human text, which means their output drifts closer to the statistical center of how people actually write.
Think about what that implies long-term. Detection works by spotting the fingerprints of a generator, the tells that come from how these systems pick their next word. The better the training data, the fainter the fingerprint. There is no version of this where detectors get permanently ahead. It’s a treadmill, and any vendor selling you a finish line is selling you something else.
The part the vendor says out loud
Here’s what actually moved me on Pangram, and it isn’t a benchmark. Their own technical report states plainly that AI detection is not a substitute for, or a reliable tool for, proving whether textual information like news and media is factual or true. AI does get used for disinformation and scams, sure. But a detector tells you something about how text was produced, not whether the text is honest.
That distinction gets flattened constantly. People want a misinformation button. They want to paste a suspicious post and get a truth score. Detection cannot give you that, and a vendor writing the limitation into its own documentation is doing something that most of this category won’t: reducing its own perceived value in exchange for being accurate about what it does. I’ll take that trade every time. It’s the closest thing to a credibility signal available in a market full of tools that will confidently red-flag the Declaration of Independence.
Where this leaves you
Spero’s framing keeps landing on the same conclusion, and it’s the one nobody wants to hear because it doesn’t scale: human verification still matters. The detector is an input. It narrows where you look. It does not close the case.
Practically, that means a few things if you’re evaluating tools in this space:
- Ask for the false positive rate before you ask for the accuracy number. If a vendor leads with accuracy and buries the rest, that tells you what they optimize for.
- Ask specifically about non-native English writers. Bias against ESL writing is the single most common way these systems cause real harm.
- Never build a policy where the tool’s output is the final decision. Build one where it triggers a conversation.
- Treat any claim of solving detection permanently as a red flag, given the direction the models are moving.
The volume problem isn’t going away either. Pangram’s own coverage notes how much AI content now fills social feeds, LinkedIn especially, and anyone who has scrolled that site for thirty seconds can confirm it without a study. Detection at that scale becomes triage, not judgment.
So the useful reframe is this. Stop asking a detector whether something is real. Ask it where to look, then do the work yourself. Spero is essentially arguing his own product is a flashlight rather than a gavel, and in a category built on selling gavels, that’s the most interesting position anyone’s taken.
🕒 Published: