What Is the Decision Threshold on an AI Detector and Who Gets to Set It?

Every AI detector you’ve ever used told you a small lie by omission. It showed you a tidy verdict — “98% AI” or a red banner or a thumbs-down — and hid the one number that actually produced that verdict. That number is the decision threshold, and once you understand it, detector results stop looking like facts and start looking like what they are: a judgment call someone made on your behalf.

Key takeaways

• A detector doesn’t output “AI” or “human.” It outputs a probability, then compares that number to a hidden cutoff called the decision threshold.
• Who sets the threshold varies. The vendor picks a default, but in some tools an institution or instructor effectively decides where suspicion begins.
• Moving the threshold trades one error for another. Lower it and you catch more AI but flag more innocent writers; raise it and the reverse happens.
• The same essay can read as “94% AI” under one tool and unflagged under another, even when the raw score barely moved.
• Because thresholds are judgment calls, a score near the line is weak evidence, not proof.

A detector’s real output isn’t a verdict

Here’s the part most people never see. When you paste text into a detector, the model doesn’t decide “human” or “AI.” It produces a probability — a number somewhere between 0 and 1 that represents how confident it is that a machine wrote the words. Maybe 0.31. Maybe 0.88.

That probability is continuous. Your writing isn’t sorted into two bins; it lands somewhere on a sliding scale. To turn that smooth scale into the blunt label you actually see, the tool needs a rule: *above this point, call it AI; below it, call it human.* That point is the decision threshold.

So the headline result is really two things glued together: a measurement (the probability) and a policy (the threshold). The measurement comes from the model. The policy comes from a person. People tend to argue about the model and ignore the policy, which is backwards, because the policy is often where your result actually gets decided.

Want the deeper version of how that probability gets calculated in the first place? We break down the perplexity and burstiness math under the hood separately. For now, just hold onto the split: a number, then a line.

What moving the threshold actually does

Picture every document the detector will ever see plotted on a line from 0 to 1. Human-written texts cluster toward the low end. AI-written texts cluster toward the high end. But the two clusters overlap in the middle — that’s the messy reality detectors live in. Some human writing scores high. Some AI writing scores low.

The threshold is a vertical line you drop somewhere into that overlap. Wherever you put it, you make two promises and break two others.

• Slide the line down (say, to 0.40). You catch more of the genuine AI text that scored low. But you also sweep up more humans whose writing happened to score above 0.40. More true catches, more false accusations.
• Slide the line up (say, to 0.80). You stop flagging most innocent writers. But now AI text scoring 0.65 sails right through undetected. Fewer false accusations, more misses.

There’s no setting that gives you everything. Lowering false positives raises false negatives, and the other way around. That trade-off is baked into the math, not a flaw someone can engineer away. It’s also why no detector will ever be perfectly accurate — the overlap doesn’t disappear, you just choose which mistake you’d rather make.

The threshold is how a vendor encodes that choice. A tool aimed at catching academic dishonesty might set an aggressive line and accept more false alarms. A tool built for content marketers might set a forgiving one to avoid annoying paying customers. Same underlying model, very different experience, entirely because of where the line sits.

Who actually sets the line?

This is the question almost nobody asks, and the answer is more scattered than you’d hope.

The vendor sets the default. An engineer or data scientist picks a threshold during development, usually by testing on a labeled dataset and choosing the cutoff that hits some target — often a low false-positive rate, since accusing real writers is the costly mistake. That default ships with the product.

The institution sometimes overrides it. In education especially, the picture gets layered. Turnitin, for instance, surfaces an AI-writing percentage to instructors, and individual schools write their own policies about what percentage is worth a conversation versus a formal academic-integrity case. The software reports a number; the institution decides what the number *means*. That policy is a threshold too — just one applied by people reading a dashboard rather than by code.

The instructor improvises the rest. A teacher who decides “I only worry above 50%” has, in that moment, set a personal threshold. Two teachers at the same school, looking at the same tool, can effectively run different cutoffs. Your outcome depends partly on whose class you’re in.

So when you ask “who decided I look like AI,” the honest answer is a chain: an engineer chose a default, an admin maybe adjusted it, and a person interpreting the result drew their own line on top. The score is objective-ish. The thresholds stacked on top of it are not.

A mini-scenario: the same paragraph, three fates

This example is invented to show the mechanics, but the pattern behind it is real. Maria, a non-native English speaker, writes a careful, somewhat formal lab report. Her prose is clean and a little repetitive — exactly the profile that tends to score high on detectors even when it’s entirely her own work. That isn’t a guess: a 2023 study by Liang and colleagues at Stanford found GPT detectors disproportionately flag non-native English writers.

Say her report gets a raw AI-probability of 0.68.

• Detector A uses a default threshold of 0.50. Verdict: “AI-generated.” Red banner. Maria gets an email asking her to explain.
• Detector B uses a default threshold of 0.80, tuned to avoid false alarms. Verdict: “Likely human.” No flag. Nobody emails anyone.
• Maria’s actual instructor has a personal rule: only escalate above 75% *and* only after reading the writing herself. Verdict: a quiet second look, then nothing.

The writing never changed. The probability never changed. Three different thresholds produced an accusation, a clean pass, and a shrug. If you only saw Detector A’s output, you’d think the tool “caught” something. It didn’t catch anything — it applied a line, and the line happened to fall below 0.68.

Why this matters for how you read any score

A few practical habits fall out of all this.

Prefer the number to the label. A tool that says “82%” is more honest than one that says only “Likely AI,” because the percentage lets you see how close you are to wherever the line sits. A 51% and a 99% can wear the same red badge. They are not remotely the same situation.

Treat the threshold zone as a coin flip, not a confession. Scores hovering near the cutoff are exactly the cases where the model is least sure. That’s the worst possible place to draw a hard conclusion. If your result is “barely over,” the right reading is “the detector couldn’t tell,” not “you got caught.”

Don’t argue with detectors as if they’re objective. When two tools disagree, neither is broken. They’re running different models *and* different thresholds. Disagreement is the expected outcome, which is also why a single detector score should never stand as proof on its own — it’s a prompt to look closer, ideally at your drafts, version history, and process.

Know that “passing” is threshold-relative too. If you revise a draft and it drops from 90% to 60%, whether that counts as “passing” depends entirely on the threshold it’s measured against. The honest goal isn’t hitting a magic number; it’s writing that genuinely reads as yours, so the result holds up no matter where the line is drawn.

The bigger point

The decision threshold is the quiet place where a measurement becomes a judgment, and it’s almost always set by someone other than you — an engineer, an admin, a teacher — using a value they rarely disclose. That’s not a conspiracy. It’s just how converting a probability into a label works. But it does mean a detector result is never the neutral fact it pretends to be.

So the next time a tool hands you a confident-looking verdict, pull it apart in your head: what was the actual probability, and whose line decided it? You’ll read every result more clearly. If you want to compare tools, it helps to know what you actually get on each plan, and you can always browse the rest of our writing on detection for more on the statistics behind those probabilities.

When you’d rather see a real score on your own writing instead of a borrowed one, run your draft through a checker and watch the actual probability move, not just the verdict.