What you can actually detect about AI-written text

You cannot reliably tell whether AI wrote something just by reading it. There is no detector that works consistently across all AI systems, and the ones that exist produce false positives and false negatives regularly. What you can do instead is look for patterns that are common in AI output — repetitive phrasing, generic reasoning, a lack of specific examples or contradictions — and treat those as signals to dig deeper, not proof of anything.

The reason detection is so hard is that large language models like GPT-4, Claude, and Gemini are trained on human text. They learn to mimic human writing well enough that the difference is often invisible. As these models improve, the gap closes further. A detector that catches 80% of AI text today might catch 40% of it in six months.

If you need to know whether something was AI-written for a real reason — checking student work, evaluating content for a publication, or assessing a job application — the most honest answer is that you should ask the person directly or use context clues about how the text was produced, not a detector tool.

Key Takeaways

  • No AI detector is reliable enough to use as proof that something was or was not written by AI.
  • AI-written text often has patterns like repetitive sentence structure, generic examples, and reasoning that sounds smooth but lacks specifics.
  • If you need to verify authorship for work or publication, asking directly or checking the writing process is more useful than running a detector.
  • Detectors improve and break constantly as AI systems change, making any tool you find today potentially useless in weeks.
  • Some AI writing is intentionally designed to sound human, while some human writing accidentally sounds like AI, so pattern-matching alone will mislead you.

Why AI detectors fail so often

AI detectors work by looking for statistical patterns — things like the distribution of word choices, sentence length, or how often certain phrases appear. The problem is that human writing also has patterns, and those patterns overlap with AI patterns more than you might think. A formal essay by a human and a formal essay by Claude can look statistically similar because both are trying to sound authoritative and clear.

The other problem is that AI systems are not a single thing. GPT-4 writes differently than Gemini, which writes differently than open-source models like Llama. A detector trained to catch one system might miss another entirely. And as companies release new versions, the patterns change.

Several studies have tested popular detectors — including Turnitin's AI detection, GPTZero, and Originality.AI — and found they flag human-written text as AI at rates between 10% and 50% depending on the detector and the text. That means if you use one to check ten human-written essays, you might get false alarms on one to five of them. That is not useful for making decisions.

Patterns that show up often in AI writing

While you cannot use these as proof, they are worth noticing. AI systems tend to repeat sentence structures. You might see multiple sentences in a row that follow the pattern "X is important because Y" or "On one hand, Z. On the other hand, W." Humans do this too, but usually with more variation.

AI writing often lacks specific details. Instead of naming a particular study, date, or example, it might say "research shows" or "studies have found." When it does include specifics, they are sometimes wrong — a real risk with AI systems that can confidently state false information. Human writers usually either know the detail or admit they do not.

Another pattern is that AI writing can sound very smooth and logical while saying very little. A paragraph might flow beautifully but contain no actual argument, just restated versions of the same point. Human writing, especially first drafts, is usually choppier and more willing to contradict itself or change direction.

AI systems also tend to hedge heavily. You see phrases like "it could be argued," "some might say," and "it is worth considering" more often than in human writing, because the model is trained to avoid making strong claims it might get wrong.

What to look for in the writing process itself

If you are trying to figure out whether a student or employee actually wrote something, the process matters more than the text itself. Ask them to explain a specific sentence or defend a choice they made. Ask them to show you their notes or outline. Ask them to write something new on the spot on a related topic. Someone who wrote the piece will usually be able to do these things; someone who used AI to write it often cannot, or will produce something noticeably different in quality or style.

For published content or social media, look at the account history. Does this person usually write like this? If someone who normally posts short, casual updates suddenly publishes a formal 800-word essay, that is worth noticing — though it could mean they hired a human writer, not that they used AI.

If you are evaluating content for accuracy, fact-check the specific claims rather than trying to detect the source. An AI-written article that cites real sources and gets the facts right is more useful than a human-written article full of errors.

Tools that claim to detect AI, and what they actually do

Several companies sell AI detection services. Turnitin, which many schools use for plagiarism detection, added an AI detection feature. GPTZero is a standalone detector. Originality.AI, Copyleaks, and others offer similar services. Most of them work by analyzing statistical patterns in the text and assigning a probability that it was AI-written.

The honest limitation: these tools are not validated against real-world use. They are tested in labs where researchers know which texts are AI and which are human, but that is not how you would use them. You would use them on text where you do not know the answer, which is exactly where they are least reliable. A tool that scores 85% accuracy in a lab test might score 60% accuracy when you use it on actual student essays, because real student essays are different from the test set.

Some detectors also flag text as "likely AI" when the probability is only 50% or 60%, which is barely better than a coin flip. Others require you to pay per check. None of them will tell you which AI system wrote something, only whether they think it was AI at all.

When you actually need to know the source

If you are checking student work, the best approach is to build in process checks from the start. Require outlines, drafts, or in-class writing. Ask students to cite their sources and explain their reasoning. These practices catch plagiarism and AI use without needing a detector.

If you are hiring and you want to know whether a cover letter or writing sample was AI-generated, ask the candidate to discuss their writing process or to write something new during the interview. You will learn more about their actual skills that way than you would from a detector.

If you are a publisher or content manager concerned about AI-generated submissions, the practical approach is to check facts, verify sources, and read carefully for the patterns described above. You are looking for signs that the content is low-effort or inaccurate, not trying to prove its origin.

The real risk: AI that is designed to sound human

As AI systems improve, companies are specifically training them to sound more natural and less detectable. OpenAI, Anthropic, and others are working on making their outputs harder to distinguish from human writing. This means detectors will become less useful over time, not more useful.

At the same time, some human writing — especially formal writing, technical writing, or writing in a second language — naturally sounds like what we think of as "AI-like." A non-native English speaker writing formally might use hedging language and generic examples not because they used AI, but because they are being careful with a language they do not use every day.

This is why context and process matter so much more than any detector. If you have reason to trust someone and their work checks out factually, a detector flagging it as AI should make you skeptical of the detector, not of the person.

Frequently Asked Questions

Can I use a free AI detector online?

You can, but the results will not be reliable. Free detectors have the same accuracy problems as paid ones, and they may also store your text on their servers. If you are checking something confidential, avoid uploading it to any online detector.

What if a detector says something is AI but I wrote it myself?

This happens regularly, especially with formal writing, technical content, or writing in a non-native language. Detectors produce false positives. If you wrote it, you can explain your process and reasoning, which a detector cannot do.

Do schools use AI detectors to catch cheating?

Some do, including schools that use Turnitin. However, many educators are moving away from detectors because of the false positive rate and because they do not address the real problem — whether students are learning. Process-based checks like drafts and in-class writing are more effective.

Can AI write something that a detector will definitely miss?

Yes. If you prompt an AI system carefully and edit the output, you can produce text that detectors struggle with. This is one reason detectors are unreliable — they can be fooled by relatively simple techniques.

Is it illegal to use AI to write something without saying so?

It depends on the context. In academic settings, most institutions have policies against submitting AI-written work as your own. In professional writing, disclosure requirements vary by industry and publication. For personal use, there are no legal restrictions, though there may be ethical ones depending on what you are writing.