You can spot likely AI writing by counting a handful of tells: heavy em-dash use, rule-of-three phrasing, unusually even sentence lengths, hedging, and stock vocabulary like “delve”, “seamless”, and “landscape”. What you cannot do is prove it. Every one of those signals also shows up in careful human writing, and any text can be edited past a detector in seconds. Treat this as informed skepticism, a prompt to read more closely, not a verdict machine.
The AI writing heuristic checker counts these patterns in text you paste and shows you the numbers, without pretending to deliver a yes/no answer.
The countable tells
These are the patterns you can actually measure, which is why a heuristic tool focuses on them:
- Em-dash density. Models lean on em-dashes far more than most people do. Several per paragraph is a flag.
- Rule-of-three. Lists and phrases bundled in threes (“fast, cheap, and reliable”) appear constantly in AI prose.
- Low burstiness. Human writing mixes short punchy sentences with long winding ones. AI output tends toward a steady, medium length. Low variance in sentence length is one of the stronger statistical signals.
- AI vocabulary. A recurring stock set: delve, tapestry, seamless, robust, moreover, furthermore, elevate, realm, landscape, testament, “when it comes to”. A cluster of these is telling.
- Hedging. Frequent “it’s important to note”, “generally”, “in many cases”, and similar softeners that pad without committing.
Any single tell means little. A formal human writer uses em-dashes; a good essay uses a rule-of-three on purpose. It is the combination and density across a whole passage that shifts the odds.
What the numbers look like
Here is roughly how the signals separate typical human writing from unedited model output. These are illustrative ranges, not thresholds to trust blindly.
| Signal | Typical human prose | Unedited AI prose |
|---|---|---|
| Em-dashes per 100 words | 0–1 | 2–5 |
| Sentence-length variance (burstiness) | High | Low |
| Rule-of-three instances | Occasional | Frequent |
| AI-vocabulary hits per 500 words | 0–2 | 5–15 |
A passage scoring high across every row is worth a second read. But notice the overlap: a careful, formal human writer can land in the right-hand column, and a lightly edited AI draft can land in the left. That overlap is the whole problem.
Why no detector is reliable
This is the part most tools bury. Detection is fundamentally unreliable, for reasons that are not going away:
- False positives hit real people. Detectors reward predictable, low-variance text. Fluent non-native English speakers, technical writers, and anyone using grammar assistants naturally produce that pattern. Their genuine work gets flagged as AI. There have been repeated, documented cases of human writing, including published essays and student work, scored as machine-generated.
- Evasion is trivial. A few minutes of editing, or one pass through a “humanising” rewriter, breaks most detectors. Because defeating detection is so easy, a clean score proves nothing either.
- The models keep changing. Detectors are trained on yesterday’s model outputs. Each new model generation shifts the patterns, so accuracy drifts down over time.
- There is no ground truth in a probability. Commercial “99% AI” badges are marketing. Under the hood it is a classifier returning a likelihood, and likelihoods are wrong a meaningful fraction of the time.
The practical upshot: never make a high-stakes decision, especially an accusation, on a detector score alone.
Worked example: reading a suspicious paragraph
Suppose you paste a 300-word paragraph and the checker reports: 4 em-dashes, 6 AI-vocabulary hits, three rule-of-three constructions, and low sentence-length variance. That is a strong cluster. Reasonable conclusion: this reads like unedited model output, so scrutinise the claims and ask the author about their process.
Now suppose the same paragraph came from a non-native speaker who writes in a formal, even register and likes dashes. Identical signals, completely human. The tool cannot tell these two apart, and neither can any detector. That is exactly why the output is a set of counts to interpret, not a verdict.
How to use this well
- Use the counts to decide where to look harder, not to conclude anything.
- Combine with context you actually have: does the author know the material, can they explain their draft, does the version history make sense.
- Never accuse someone based on a score. Ask about process instead.
- If you write and want to avoid tripping these signals yourself, vary sentence length, cut stock vocabulary, and drop unnecessary hedging. Our AI vs human writing cost guide looks at where machine drafting genuinely helps and where a human pass is still worth the money.
Run your own sample through the AI writing heuristic checker to see the patterns laid out. Just remember what it is: a lens, not a judge.
