A smoke alarm doesn’t know whether your kitchen is on fire or whether you just burnt toast. It knows one thing: particles in the air crossed a threshold. It screams. You investigate. That division of labor is the whole reason smoke alarms are useful and also the reason nobody lets one testify in court.
I’ve been building bots long enough to have watched a lot of engineers treat classifiers like judges. AI text detection is where that habit gets expensive. Pangram’s Max Spero has been talking about this space publicly, and the thread running through his comments is not “we solved it.” It’s that detection matters for fighting misinformation, and that even good tools sit inside a process that still requires humans.
Real or Fake Is the Wrong Question
If you’ve ever shipped a classifier, you know the framing problem. “Real or fake” implies a clean binary with a stable boundary. Text doesn’t work that way. A human writes a draft, a model polishes it. A model writes a draft, a human rewrites half of it. An ESL writer produces prose that is entirely their own but statistically resembles the flattened register that models default to. Where does the line go?
Spero has noted that most detection tools can catch copy that’s been run through “humanizer” services. That’s encouraging as an engineering result. But he also points at the harder trend: as large language models train on an ever-wider pool of research and text, detection gets more difficult. The models are absorbing more of the range of human expression, which narrows the gap a detector is trying to measure.
That’s a moving target, not a fixed one. Anyone who has maintained a spam filter recognizes the shape of it. You’re not solving a problem once; you’re subscribing to it.
False Positives Are the Metric That Matters
Pangram’s stated focus is reducing false positives to improve accuracy, and the company has published comparisons against Turnitin covering accuracy, false positive rates, ESL bias, and pricing. I want to sit with the false positive point because builders routinely get it wrong.
When you evaluate a detector, overall accuracy is the least interesting number in the report. What matters is the cost asymmetry:
- A false negative means AI-written text slips through. Annoying, sometimes consequential.
- A false positive means a real person gets accused of something they didn’t do.
Those are not equivalent errors, and no single accuracy figure captures the difference. If you’re wiring a detector into a moderation pipeline, a hiring flow, or a grading system, the false positive rate is your actual risk surface. Ask for it specifically. Ask what population it was measured on. A tool that looks great on clean synthetic samples can behave very differently on writing from non-native speakers.
What Pangram Says About Its Own Limits
One detail from Pangram’s technical report deserves more attention than it gets: the company states plainly that AI detection is not a substitute for, nor a reliable tool for, proving whether text is factually true. Detection tells you something about how text was produced. It tells you nothing about whether the claims inside it hold up.
That’s a vendor volunteering a boundary on its own product, which is rarer than it should be. It also lines up with what Spero has said about human verification remaining necessary. AI is frequently used for disinformation and scams, so detection has real defensive value. But “a model probably wrote this” and “this is false” are different findings, and collapsing them produces bad decisions in both directions. Plenty of true things are drafted by models. Plenty of lies are typed by hand.
How I’d Wire It Into a Bot
Practical takeaways for anyone putting a detector into a system:
- Treat the score as a signal, not a verdict. Route it to a queue, not to an automated punishment.
- Log the score alongside the input so you can audit drift later. Detection performance changes as models change.
- Never surface a raw confidence number to an end user as if it were a fact about them.
- Keep a human in the loop for anything with consequences attached. That’s not a temporary crutch until the tech improves; it’s the design.
- Separate your provenance checks from your fact checks. Different tools, different questions.
Spero’s broader message, as I read it, is a mature one for a founder in this space: the tools are getting better, the problem is getting harder, and neither of those cancels the other out. Detection is worth doing. It’s also worth being honest about what a detector can and cannot know.
Build the alarm. Just don’t let it hand down sentences.
🕒 Published: