Max Spero, who runs the AI detection company Pangram, made a point recently that stuck with me: most detection tools can already catch text that’s been run through an online “humanizer.” That’s the easy part. The harder part, he said, is that as large language models get more and more material to learn from, detection keeps getting tougher.
My first reaction was relief. Finally, someone in the detection business not promising a magic button. Because the way most of us talk about AI detection is roughly the way we talk about a metal detector at the airport. Beep or no beep. Real or fake. And that framing is wrong in a way that causes real damage to real people.
Why “real or fake” is the wrong question
Think about what a detector is actually doing. It isn’t reading your essay and recognizing a robot’s handwriting. It’s looking at statistical patterns in word choice and sentence structure, then estimating how likely it is that a machine produced them.
That estimate is a probability, not a verdict. And probabilities get shaky when the thing you’re measuring keeps changing shape. Every new model release shifts the patterns a little. Every writer who edits AI output shifts them more. Every human who naturally writes in clean, tidy prose looks a bit more like a machine than a human who rambles.
So the honest output of a detector isn’t “fake.” It’s something closer to “this looks a lot like machine-generated text, and here’s how confident I am.” Those are very different sentences, and only one of them should ever be used to accuse a student of cheating.
False positives are the part nobody talks about
Pangram’s public positioning leans hard on reducing false positives, and Spero has written comparisons against Turnitin that specifically weigh accuracy, false positive rates, ESL bias, and pricing. I find the emphasis on those first two telling.
A false positive is when the tool flags human writing as AI. If you’re a company selling detection, false positives are your most embarrassing failure mode, because they’re the ones with a name and a face attached. A missed detection is invisible. A wrongly accused student is a story.
The ESL bias question is the one I’d want every school administrator to sit with. Writers working in a second language often use simpler sentence structures and more standard phrasing, which is exactly the profile a detector can misread as machine output. When a tool is wrong in a patterned way rather than a random way, the same group of people absorbs the cost over and over.
What detection can’t do at all
Here’s my favorite detail from Pangram’s own technical report, and I wish it got quoted more often: AI detection is not a way to prove whether text is factually true. Those are completely separate problems.
A detector can tell you a paragraph was probably machine-written. It cannot tell you whether the paragraph is accurate, misleading, or invented. A person can write a lie by hand. A model can write something perfectly correct. Origin and truth are not the same axis, and confusing them is how people end up trusting the wrong things:
- “This passed the detector, so it must be true.” No. It means a human probably wrote it.
- “This got flagged, so it must be misinformation.” Also no. Plenty of AI-assisted writing is fine and honest.
- “The tool said 87 percent, so the case is closed.” That number is a signal to look closer, not a conclusion.
Spero’s framing is that detection still matters for pushing back on misinformation at scale, and I agree with that. Knowing that a flood of identical-sounding posts came from a machine is genuinely useful context. It just isn’t a fact-check.
The volume problem is only getting bigger
Pangram has flagged how much AI text is now flowing through social media, LinkedIn especially. If you’ve scrolled that feed lately, you already knew. And this is where AI agents enter the picture, which is what we care about on this site.
Agents don’t just write one post when you ask. They write continuously, on schedules, across accounts, in whatever tone you configured. The amount of machine-generated text in circulation is growing faster than any detection tool’s ability to keep pace with each new model. That’s the structural reason Spero’s “harder over time” comment matters more than any single accuracy score.
What to actually do with this
The practical takeaway is the least exciting one: human judgment stays in the loop. Detection results are evidence, not proof. Use them to start a conversation, ask for a draft history, check a claim against a source, or notice that fifty accounts posted the same insight within an hour.
If a tool tells you it’s certain, be suspicious of the tool. If it gives you a probability and admits what it can’t measure, that’s the more useful one. That’s the part I appreciate about how Spero talks about his own product. He’s selling a detector while telling you it isn’t an oracle, and in this space, that counts for something.
đź•’ Published: