Start free

How to detect AI writing without tools

Automated detectors are not the only way to spot AI-written text. A careful human reader can catch vocabulary habits, rhythm, structure, and missing specifics that software scores only indirectly.

Why human judgment matters

Detectors score statistical patterns. They miss things a human reader catches: an argument that is not quite connected to the prompt, a specific detail that is wrong, a voice that does not match the rest of a student's portfolio. Humans are better at context. Machines are better at volume. Using both is the honest answer.

The vocabulary tells

LLMs reach for a predictable set of mid-formal words at rates far higher than most humans:
  • leverage, navigate, navigate the complexities
  • multifaceted, tapestry, delve, delve into
  • it's important to note, it's worth noting
  • in today's fast-paced world, in the ever-evolving landscape
  • ensure, foster, robust, seamless
One or two of these per page is normal. Four or five in a single paragraph is a flag. It is not proof of AI authorship, a fluent corporate writer uses the same vocabulary, but combined with other signals, density of tier-1 AI vocabulary is one of the clearest giveaways.

The em-dash fingerprint

ChatGPT loves em dashes. Many models do. If you see an em dash in every other paragraph, and the em dashes are being used to append a polished afterthought rather than to signal a real interruption, that is a tell. Real human writers use em dashes too, but less evenly, and for different purposes. We go deeper on this in our piece on the em dash problem.

The rhythm tell

Read a paragraph out loud. In human writing, sentence lengths swing hard: a short declarative hits, then a long clause-chained analysis unpacks it, then another short line lands the point. AI tends to stay within a tighter range. Paragraphs of medium sentences with similar structure have a mechanical cadence once you hear it.

The structural tell

Ask a model for an essay and you will usually get a predictable shape: an introduction, three evenly weighted body sections, and a conclusion that echoes the intro. Subheadings come in parallel. Lists come in threes. Each section is about the same length.Real essays are lumpy. One section is long because the writer has more to say about it. One paragraph is a single sentence because that sentence is the whole point. Symmetry is AI-coded; asymmetry is usually human.

The nothing specific tell

AI writes around topics. It gestures at examples without naming them. "A recent study showed..." without saying which study. "Many experts argue..." without naming an expert. "Throughout history..." without anchoring a century. When the specifics are missing or generic, the piece may have been generated from a short prompt without a knowledge-base context.

The perfect ending tell

Human essays often end where the writer ran out of thought, or hedge their way out, or stop on a specific image. AI endings tend to reassert the thesis, list the three things covered, and close with an aspirational line about the future. If the last paragraph sounds like a LinkedIn post by default, check the rest for other signals.

Context cues for teachers and editors

If you read a student's earlier work, you have a voice baseline. Sudden shifts, more polished syntax, a different vocabulary range, no more idiosyncratic phrases they used to use, are informative. Guidance from university communication labs on AI in writing walks through how to have the conversation with students when you suspect AI use, which matters as much as the detection itself.

When to use a detector anyway

Eyeballing gets you directionally right. A free in-browser detector like Leap's adds a second signal and can break ties on ambiguous pieces: it scores text from 0 to 100 using writing signals such as sentence-length variation, stock AI phrases, hedging words, em dash density, and repetition, and it highlights the sentences that weigh most. Nothing leaves your browser and no account is needed.Do not treat a detector score as proof. It is a signal, not evidence, and it can be wrong. Treat it as a second opinion that, combined with your own reading, tells you whether the piece warrants a conversation.

A note on false accusations

Never accuse someone of using AI based on vibes alone. The cost of a wrong accusation, to the student, writer, or employee, is real. Use detection cues to open a conversation, not to close one. Our piece on false positives goes deeper on why this matters.

Frequently asked questions