How AI detectors actually work.

Most AI-text detectors do one of two things, and understanding which one — and what it measures — explains almost every false positive you've ever seen reported online. No machine-learning background required.

Approach one: perplexity and burstiness

A language model is, at heart, a next-word probability machine. Show it "the cat sat on the" and it assigns a high probability to "mat" and a low one to "helicopter." Perplexity is a measure of how surprised a model is by a text: low perplexity means the text is close to what the model would itself have written; high perplexity means it keeps getting surprised.

The detection logic is simple: AI-generated text was produced by choosing high-probability words, so it tends to have low perplexity. Human writing, full of unexpected choices, tends to score higher. Detectors add burstiness — the variation in sentence length and complexity across a passage — because humans vary their cadence and models tend not to.

It is an elegant idea. It is also the source of the problem, because "predictable, low-perplexity, low-variation text" is not a definition of machine writing. It is a definition of a particular kind of clear prose — and a great deal of human writing qualifies.

Approach two: a trained classifier

The other family of detectors trains a model on labeled examples — "this batch is human, this batch is GPT" — and learns whatever statistical boundary separates them. This can be more accurate than raw perplexity, but it inherits a fatal dependency: it only knows the distribution it was trained on. Most public detectors were tuned on student essays, news copy, and forum text, because that is where the academic-integrity money was. Fiction — especially literary fiction — was barely in the training mix.

Why fiction breaks both

Now stack the two facts. Detection rewards "high perplexity, high burstiness, far from the model's median." Literary craft, over the last century, has been a sustained project of removing exactly those properties:

• Minimalism. Hemingway, Carver, Didion: short declarative sentences, low lexical surprise, deliberately flattened cadence. To a perplexity detector this is indistinguishable from a cautious chatbot.
• Register discipline. A strong authorial voice is consistent. Consistency lowers variance. The better the voice control, the more "machine-like" the burstiness score.
• Genre convention. Thriller and romance prose is engineered to be frictionless and fast. Frictionless reads as low-perplexity.
• Editing. A professionally line-edited manuscript has had its idiosyncratic bumps sanded down by a copyeditor — which is, statistically, the same operation a model performs.

The cruel irony: the more polished and disciplined a piece of human fiction is, the more likely a general-purpose detector is to call it AI. The tool penalizes craft.

The base-rate problem

There is a second, quieter failure. Suppose a detector is 95% accurate — excellent, by industry standards. An agent reads 2,000 submissions a year, of which, say, 40 are actually AI-assembled. Run the math: the 5% error rate on the 1,960 genuine manuscripts produces roughly 98 false accusations — more than twice the number of real cases. A high-accuracy detector, applied to a population that is mostly human, still generates a majority of false positives. Any honest tool has to be designed around this arithmetic, not in spite of it.

What we do differently

Slopsleuth deliberately does not return a single perplexity-style probability. It runs five independent audits that target structural properties of fiction specifically — dialogue texture, voice variance across chapters, rhythmic signature, and so on — each calibrated against both contemporary machine-written fiction and a baseline of published human novels spanning a century. The output is a per-audit breakdown with the false-positive cases documented in the open, not a verdict.

That design will never give you the satisfying certainty of a single number. That's the point. Certainty is precisely what the math does not support, and a tool that pretends otherwise is selling confidence it hasn't earned.

The takeaway for anyone evaluating a manuscript

A detector score is a hypothesis, not a finding. Used as a place to look harder — read the flagged chapters, ask the writer about process, compare against their other work — it is genuinely useful. Used as evidence, it will eventually accuse someone's careful, human, deeply-revised book of being a machine. Knowing how the machinery works is the best protection against trusting it too much.