Can AI detectors reliably identify ChatGPT-generated text?

Most consumer AI detectors produce a probability score rather than a definitive answer, and they generate false positives on polished human writing as often as they catch genuinely generated text. They work by measuring statistical patterns like perplexity and burstiness, which overlap heavily with well-edited human prose. Treat any detector output as a starting point for closer reading, not a verdict.

What single signal is most reliable for spotting AI-written text?

The absence of specific, first-person detail is the hardest tell to fake. AI text summarizes experiences, generalizes examples, and avoids naming a particular person, place, or moment. A human author will include one small concrete fact — a street name, a phrase someone said, a number they looked up at 2 a.m. — that no model would generate because it is tied to a specific life.

Is using ChatGPT to help write an email or report unethical?

It depends on the context and the expectations set by the other party. In most workplace settings, using an AI tool to draft or edit is increasingly treated as a productivity aid, similar to using Grammarly or a spell-checker. It becomes a problem when someone presents fully generated text as their own analysis in a context where original reasoning was explicitly required, such as a thesis defense, a legal filing, or a hiring assessment.

How can I make my own writing harder to mistake for AI output?

Write with specificity and opinion. Name the person you were talking to, include the number you actually looked up, state the one thing you are least sure about and why it matters. Let a sentence be too long or too short because that is how your brain processed the thought. The goal is not to sound messy but to sound like a particular person with a particular perspective, which is exactly what AI text struggles to produce.

AI Detection

How to Tell if Someone Is Using ChatGPT or Another Writing Tool

By VisibleAISearch · October 9, 2026 · 6 min read

AI detectionwriting authenticityprofessional communicationsignal spotting

A weathered granite lighthouse standing on a windswept rocky headland at golden hour, its stone base scarred by salt spray and lichen. In the foreground on a flat rock ledge rests a brass mariner's compass with its needle frozen, beside a short length of frayed hemp rope coiled once. The sea churns in broad strokes of pewter and amber behind the structure. Wide environmental composition, natural light, no text, no people.

The Rhythm and Structure That Give It Away

AI-generated prose has a metronomic quality that most readers feel before they can name it. Paragraphs tend to land at similar lengths, rarely two sentences, rarely twelve, clustering around six to nine. Sentence length alternates in a predictable pulse: one longer clause, then a shorter one, then a mid-range connector. Human writing is messier. A person thinking through an idea will produce a three-word sentence followed by a sprawling run-on that backtracks and corrects itself mid-thought. The AI version of that same passage would give you two evenly weighted sentences with a semicolon or em-dash in the middle, as if the rhythm were set on a timer.

Another structural tell is the insistence on triads. Lists of three, three-part summaries, three reasons followed by a synthesis paragraph. Humans do use lists of three, of course, but they also use lists of two, four, or seven items without apologizing for the number. AI text gravitates toward the cleanest possible count because its training rewards symmetry. If every section of a long document resolves into exactly three points, each with one example and one transition sentence, the probability that a human sat down and thought through each choice drops noticeably.

There is also a uniformity of purpose in AI paragraphs that is hard to replicate manually. Each paragraph opens with a topic sentence, develops it over two or three supporting sentences, and closes with a soft bridge to the next idea. A human writer will sometimes open mid-thought, will let a paragraph drift into an anecdote and then snap back, will forget to connect two ideas and simply move on. That looseness is not a flaw; it is evidence of a person deciding in real time what matters.

This structural regularity matters beyond casual suspicion. In professional settings, hiring reviews, client deliverables, academic submissions, the inability to distinguish a carefully edited human draft from an AI-generated one has created real friction. The good news is that the signals are learnable. You do not need software or a probability score; you need to train your ear for the metronome and notice when the writing sounds too resolved, too complete, too free of the small stumbles that mark a person actually thinking.

Vocabulary Patterns That Cluster Unnaturally

Certain words and phrases appear in AI-generated text with a frequency that no individual writer would naturally sustain. Words like delve, moreover, leverage, tapestry, landscape (used metaphorically), and the phrase it is important to note carry a particular weight in machine output because they are statistically safe, they fit nearly any context without committing to a specific meaning. A human might use leverage once in a technical document; AI text will sprinkle it across three paragraphs as though it were a neutral connector rather than a jargon-heavy verb.

Hedging language is another cluster worth watching. Phrases such as while it is worth considering, one might argue, it could be said that, and in many cases create a soft, non-committal tone that reads as carefully worded but oddly risk-averse. Humans hedge too, especially in legal or diplomatic contexts, but they usually hedge once and then make their point. AI text hedges repeatedly, stacking qualifiers like while one can acknowledge that it is worth noting that, because its objective function rewards avoiding a definitive claim over making one.

Contractions behave differently as well. In informal human writing, contractions are common and varied: I'm, don't, won't, ain't, y'all. AI text in a semi-formal register will either avoid contractions entirely for a polished effect or use a narrow set (I've, it's, you'll) with mechanical consistency. The absence of one specific contraction, say, the word couldnt instead of couldn't, or the phrase its worth rather than it's worth, in an otherwise casual paragraph is a small but telling seam.

None of these words are proof by themselves. A thoughtful human writer can use delve or moreover, and a casual AI output might skip contractions entirely depending on how it was prompted. The signal is density and pattern: when five or six of these markers appear in the same three-hundred-word stretch, the probability that a single person chose each one deliberately becomes hard to sustain.

A magnifying glass with a dark wooden handle lying across a small stack of five blank cream-colored index cards bound together with a single loop of red waxed twine. Beside the cards sits a short dried lavender sprig and a tarnished brass skeleton key, all resting on rough-hewn weathered oak planks with visible grain and a faint water ring. Tight close-up framing, warm directional light from the left, shallow depth of field, no text, no screens, no people.

What Is Missing From AI-Generated Text

The most reliable indicators of AI authorship are often absences rather than presences. AI text rarely contains a specific, sensory, first-person memory. It will say the client raised concerns about timeline risk; a human who was in that meeting might write the client kept tapping her pen on the table and said we need this by Friday or the whole thing falls apart. The second version carries the weight of a person who was there, who noticed the pen-tapping, who chose to include it because it mattered to them. AI text summarizes the event; human text inhabits it.

Contradiction and self-correction are another absence. A person writing under pressure will assert something, then immediately qualify or reverse it: I think the pricing model is wrong, well, not wrong, but it assumes a usage pattern we have not validated yet. That mid-sentence pivot is the sound of thought happening in real time. AI text, by contrast, presents a single coherent position per paragraph because its generation process favors local consistency. It does not talk itself out of a claim and then walk back into a better one, because that would introduce an inconsistency in the sequence.

There is also an absence of genuine stakes. AI-generated recommendations read as if the author has nothing to lose by being wrong. A human who stands behind a proposal, a freelancer pitching a project, a manager recommending a hire, a student defending a thesis, will let their uncertainty or their conviction bleed through. They will use I believe where a safer phrasing would be it is evident. They will admit the part they are least sure about. That vulnerability is, paradoxically, the strongest authenticity signal in any piece of writing.

This absence theme has a broader implication that extends well beyond detecting a single email or essay. In professional and public life, recognizability is becoming as important as accuracy. If your writing, your business voice, your public-facing content all sound like a smoothed-out average of everything the internet has ever said, you become interchangeable. The same principle that makes AI text hard to attribute to a specific person also makes it hard for readers to trust what they are reading. Specificity, contradiction, and stake are not stylistic preferences; they are the markers that say a particular human chose these words for these reasons.

Context Changes What Counts as a Signal

A sentence that would trigger suspicion in one context is completely unremarkable in another. A graduate student writing a literature review will naturally produce dense, hedged, triadic prose because that is the register of the field. A non-native English speaker drafting a professional email may write in full sentences without contractions not because an AI generated it but because their native grammar does not favor them. A marketing team's brand voice guide might explicitly require three-bullet summaries and transition sentences, making every output structurally uniform by design. Before you reach for the suspicion lever, ask whether the context itself explains the pattern.

The stakes of the situation should calibrate your detection effort. If a freelancer sends you a project update that sounds slightly too polished, it may be because they spent twenty minutes editing it and are tired. If a vendor submits a forty-page strategic proposal that reads like a well-organized Wikipedia article with no named case studies, no specific client names, no acknowledgment of trade-offs they personally weigh, the absence of lived detail becomes harder to explain away. The same linguistic pattern carries different weight depending on what is at risk and how much time the author plausibly spent.

It is also worth acknowledging that AI-assisted writing is not binary. Many professionals use an AI tool to brainstorm outlines, generate first drafts, or check grammar, then revise heavily in their own voice. The result is a hybrid: structurally sound but carrying personal asides, domain-specific jargon, and the occasional idiosyncratic phrasing that no model would produce unprompted. Your detection should be calibrated to identify fully generated text, not every instance where someone used a tool to lower their initial friction.

What to Do When You Suspect AI Authorship

The most effective follow-up is a specific question that requires lived experience to answer well. If a colleague's email recommends a particular vendor, ask them to describe the onboarding process they went through with a similar tool last year, or what specifically frustrated them about the previous contract. An AI-generated text will have summarized the recommendation; a human who actually evaluated it will be able to pull out a concrete, slightly awkward detail that no model would invent because it carries personal context.

You can also ask for reasoning on a single point rather than the whole document. Pick one claim, one example, or one transition and ask why they chose that particular framing over an alternative. A human who wrote the text will be able to explain their thought process: I put that first because the client was most anxious about cost, so I wanted to address it before they could. An AI-generated text does not have a thought process to recount; the answer will either be vague or will restate what is already on the page.

Frame the conversation as curiosity rather than accusation. I am trying to understand your reasoning here because we need to present this to the board, and that phrasing invites explanation without implying fraud. In academic or hiring contexts, institutional policies may already address AI use; in those cases, follow the stated process. But in most workplace situations, a direct but low-stakes question about a specific detail will resolve your uncertainty faster and more fairly than any detection tool, because it tests for something no model can fake convincingly: a person who remembers, who noticed, who cared enough to include that particular fact.