A student submits an essay that reads a little too smoothly. A freelancer delivers an article that is suspiciously generic. A LinkedIn post from a stranger is perfectly structured, perfectly neutral, and perfectly dull. Did a human write it — or did a machine? And does it even matter?
Sometimes it matters a great deal: teachers grading work, hiring managers reviewing writing samples, publishers checking submissions, readers trying to decide what to trust. This guide covers what AI-generated text actually looks like, what the detection tools can and cannot do, and why catching machine-written content is a far harder problem than most people assume.
First, the Honest Frame: Detection Is Hard
Before the clues, the caveat that shapes everything else: there is no reliable test that proves a text was written by AI. The best-known detection tools have been withdrawn, walked back, or quietly demoted by their own makers — including the companies that build the AI models themselves — because false results are common. AI-written text can read as human, and human-written text can read as AI.
There is also a moving-target problem. Detection patterns are learned from yesterday’s models, while today’s models write differently — and tomorrow’s will write differently again. A checklist of “AI words” published this year ages quickly, because people feed the list back to the models and ask them to avoid those words. Any detection habit you build needs refreshing, and any tool frozen in time is quietly decaying.
That matters because the stakes are real. A false accusation can ruin a student’s reputation; a missed fake can flood a hiring pipeline with invented writing samples. Every signal below is a hint for further investigation, never a verdict. Anyone who tells you they can identify AI writing with certainty is selling you something.
Patterns That Often (But Not Always) Suggest AI
With that warning posted, experienced readers do notice recurring habits in machine-generated text:
- Over-smooth neutrality. AI drafts often sound reasonable about everything and decisive about nothing — every perspective acknowledged, every edge sanded off.
- Repetitive sentence rhythm. Humans vary: a short punch, then a long wandering sentence full of commas and asides. Machine text tends toward a steady, medium-length drumbeat.
- Generic filler phrases. Lines like “in conclusion,” “it is worth noting that,” and long introductions that say little. (Human writers do this too, especially tired ones — another reason clues are not proof.)
- Formulaic structure. An intro, exactly three subheadings, a concluding paragraph that repeats the intro — the outline-first shape of an untuned chatbot response.
- Confident vagueness. Statements that sound authoritative but commit to nothing specific — names, dates, and numbers conspicuously absent. Ironically, confident-but-wrong specifics are also a sign, since AI hallucinations produce exactly those.
- Odd factual slips. A citation to a source that does not exist, an event placed in the wrong year, a statistic no one can find elsewhere — machines invent; humans usually misremember differently.
What AI Detectors Promise and What They Deliver
Commercial AI detectors — the tools that give a text an “AI probability score” — mostly work by measuring statistical patterns like perplexity and burstiness: how “surprising” the word choices are and how much they vary. Machine text is typically smoother and less varied. The problem is the overlap zone. Non-native English speakers, careful formal writers, and anyone following a strict template produce statistically smooth text too — and detectors have repeatedly flagged exactly these people.
Some detector services are more reliable than others on long, unedited passages, and nearly all of them get worse on short texts, heavily edited AI drafts, or text that has been run through paraphrasing tools. A detector score is best treated as a starting point for a conversation, not evidence for a penalty. Several major writing platforms now advise against disciplinary use of detection scores alone — a sensible policy worth imitating.
Real Evidence Beats Vibes
When suspicion arises, there are checks that hold up better than any scanning tool. Version history is the strongest: real writers leave trails of drafts, revisions, and timestamps across documents, while a pasted AI answer usually arrives fully formed. Consistency checks help too — ask the writer about choices inside the piece, and surface-level authors struggle to explain sentences they did not actually compose. Source verification is nearly free: real citations resolve to real pages, while invented references collapse under one click.
None of these are foolproof either, but they share a virtue detectors lack: they produce findings you can show the accused person, discuss, and appeal. A percentage score from a black box produces only an argument.
When Human Judgment Still Wins
For the situations where authorship actually matters — classrooms, hiring, publishing — human context beats software. A teacher who knows a student’s voice can spot a sudden shift in style. A hiring manager can follow up with a short live writing exercise or ask the candidate to explain a choice in their sample. A publisher can simply ask for drafts and notes.
These approaches work because they test the person, not the artifact. The goal is rarely to catch anyone at all: it is usually to confirm the person behind the work can stand behind it. If you are on the receiving end, the practical defense is transparency — writing in your own voice, keeping notes and drafts, and being honest about which tools helped you.
A Note on Images, Audio, and Video
The same principle extends beyond text. Machine-made images sometimes contain warped hands, garbled background text, or impossible physics; our deep-dive on spotting deepfakes catalogs those tells. Cloned audio tends to sound right but feel emotionally flat — see the voice cloning explainer for how those voices are built. In every medium, the story rhymes: eyeball the tells, but verify through context, sources, and corroboration instead of trusting your gut.
Frequently Asked Questions
Can I make my AI-assisted writing undetectable? You can make AI text more human-like with editing — varying rhythm, adding specifics, cutting filler — which is, conveniently, also just good writing practice. Deliberately hiding AI involvement where disclosure is expected (assignments, submissions, client work) is dishonest and can get you fired or failed. Where disclosure is expected, the safe move is simple: say which tools you used and what you changed, reviewed, and rewrote yourself.
Should schools use AI detectors? With extreme caution, if at all. False positives cluster on the students least able to fight back — learners writing in a second language, students following strict rubrics. Process checks (drafts, in-class writing, oral follow-ups) are fairer and harder to game.
What if my business publishes AI-assisted content? Disclose it where readers reasonably care — reviews, tests, and advice pieces deserve honesty about how they were produced, and many platforms now ask for labels on AI-generated media. The content itself is usually not the problem; undisclosed content presented as first-hand human experience is.
Will detection get better over time? Perhaps marginally. Watermarking techniques — hidden statistical signatures in model output — show promise, but they can be weakened by editing or paraphrasing, and they require all model providers to cooperate. The realistic long-term answer is authenticity infrastructure and norms, not a perfect scanner.
