Not reliably. AI detectors can sometimes spot unedited text straight out of a chatbot, but they regularly flag genuine human writing as AI-generated and are easily fooled by light editing, so their scores shouldn't be used on their own to make decisions about people. When it comes to AI detectors accuracy, the honest summary is "a weak hint, not evidence."

That matters if you're screening job applications, reviewing freelance work, or running a training institute. This post explains how detectors work, why they produce false positives, who gets hurt by them, and what better policies look like for hiring, education and client work.

How do AI detectors work?

Most AI detectors don't have a secret record of everything ChatGPT has ever written. They look at the text itself and estimate how "AI-like" it is, based on statistical patterns.

Two ideas sit behind many of them:

  • Predictability. Language models tend to choose likely, common words. Text where each next word is very predictable scores as more AI-like. The technical term is low perplexity.
  • Uniformity. Human writing often varies more in sentence length and structure, while AI writing can be more even. This is sometimes called burstiness: humans are burstier.

Some detectors also use a classifier, a model trained on examples of human and AI text to tell them apart.

An everyday analogy: it's like guessing whether a dish came from a chain restaurant by how standard it tastes. Often you'll be right. But plenty of home cooks make a perfectly standard pasta, and a chain dish with an extra handful of chilies will fool you.

Why do AI detectors get it wrong?

Because "predictable and even" isn't a property only AI has.

False positives, where human work is flagged as AI, are common in exactly the situations where you'd use a detector:

  • Formal and formulaic writing. Cover letters, legal summaries, technical documentation and SOPs follow conventions, which makes them predictable.
  • Non-native English writers. People writing carefully in a second language often use simpler, more common phrasing, which detectors can read as machine-like. This is one of the most widely raised fairness concerns.
  • Short texts. A paragraph or two gives the detector very little to go on.
  • Writing improved by grammar tools. Spelling and style assistants smooth out the irregularities detectors look for.
  • Well-known passages. Famous texts and common phrasings can score as AI because models learned them so thoroughly.

False negatives, where AI text passes as human, are just as easy to produce. Asking the AI to vary sentence length, adding a few personal details, paraphrasing with another tool, or simply editing by hand can move a score dramatically.

Detection also degrades as models improve and as people mix human and AI writing in the same document, which is now the normal way people write.

What do detector scores actually mean?

A detector that says "87% AI" sounds precise. It isn't a measurement of how much of the text was AI-written. It's the tool's confidence in its own guess, from a model with its own error rates on its own test data.

What the score looks likeWhat it actually means
"87% AI-generated"The tool's model leans strongly toward "AI-like patterns," not that 87% of the words came from AI
Highlighted sentencesPassages the tool found most predictable, not proof of their source
"Human-written"The tool didn't find strong AI-like patterns; the text could still be AI-assisted
Different scores from different toolsNormal; they use different methods and can disagree sharply on the same text

Run the same document through three detectors and you may get three very different answers. That alone tells you how much weight a single score deserves.

Should you use AI detectors in hiring?

We'd strongly advise against using them to reject candidates. The risks stack up:

  • You'll reject good people. Clear, conventional cover letters and non-native English speakers are disproportionately likely to be flagged.
  • You may create a discrimination problem. If a tool systematically disadvantages people by language background, that's a fairness and potentially a legal issue, depending on employment and anti-discrimination rules where you operate.
  • You're testing the wrong thing. Most of your employees will use AI tools at work. Whether a cover letter was polished with AI tells you little about whether someone can do the job.

Better approaches:

  1. Assume applicants may use AI and design around it. Ask for specifics only the applicant would know: a project they led, a decision they'd make differently.
  2. Use live or timed exercises for skills that matter, like writing, analysis or code, and let people explain their reasoning.
  3. Talk to people. A short conversation reveals far more than any detector.
  4. State your expectations. If you'd prefer candidates not use AI for a specific task, say so.

If you're using AI in your own hiring process, the same caution applies in reverse; our guide to AI resume screening covers fairness and human review.

Should training institutes and educators rely on them?

Educators face the hardest version of this problem, and many have concluded detector scores alone can't fairly be used to accuse a student. A false accusation of cheating is serious, and appeals based on a probability score are hard for anyone to resolve fairly.

More robust approaches include:

  • Assessment design: in-class writing, oral explanations, drafts submitted along the way, and assignments tied to personal experience or class-specific material.
  • Process evidence: version history in a document editor shows how a piece of work developed.
  • Clear rules: a written policy that says which AI uses are allowed (brainstorming, grammar checks) and which aren't (submitting generated work as your own).
  • Conversation first: if something looks off, ask the student to talk through their work before reaching any conclusion.

Coaching institutes and edtech businesses can find more on running AI responsibly in our guide to AI for coaching institutes and edtech.

What about checking freelancer or agency work?

Many businesses ask detectors a question they don't really need answered. What you usually care about is: is the work accurate, original, fit for purpose and not copied from someone else? And did the supplier honor what your contract says about AI use and confidentiality?

So:

  • Put AI expectations in the contract. If you need disclosure, or need certain material kept out of AI tools, write it down.
  • Check quality directly. Facts, originality (a plagiarism checker is more useful here than an AI detector), and whether it meets the brief.
  • Ask about process. A good supplier can explain how they work and where AI fits.

What's a better policy than "run it through a detector"?

For any team, the most practical approach is to stop treating AI use as something to catch and start treating it as something to govern:

  1. Decide which AI uses are fine for each kind of work: drafting, editing, research, translation.
  2. Decide what needs disclosure, to clients, customers or managers.
  3. Make people responsible for the output, whoever or whatever drafted it. If it's wrong, it's wrong.
  4. Train people on verifying AI output, especially facts and figures.
  5. Write it down in a short policy. Our AI usage policy guide gives you a structure, and our post on responsible AI for business covers the principles behind it.

If you do keep a detector in your toolkit, use it only as a prompt for a conversation, never as the deciding evidence.

Frequently Asked Questions

What is the most accurate AI detector?

Vendors publish their own accuracy claims, but results vary widely by text type, length, language and how much editing was done, and independent tests often find meaningful error rates. No detector is reliable enough to be the sole basis for a decision about a person.

Can AI detectors detect ChatGPT specifically?

They may pick up patterns typical of chatbot output in unedited text, but they generally can't confirm which tool produced something, and light editing often changes the result. Newer models also make detection harder over time.

Why was my own writing flagged as AI?

Probably because it was clear, conventional or formal, used common phrasing, was short, or was polished with a grammar tool. These are the same features detectors associate with AI. It says little about how you actually wrote it.

Do watermarks solve the detection problem?

Watermarking, where an AI provider embeds a hidden statistical signal in generated content, is being developed and used in some products, especially for images. But it only works if the provider used it, and text watermarks can be weakened by editing or paraphrasing. It's a helpful signal, not a complete solution.

Is it illegal to use AI detectors on job applicants?

Not inherently, but using any automated tool in hiring can raise fairness, transparency and discrimination concerns depending on the rules where you operate. If you use one at all, use it with human review and get advice on local employment law. This is not legal advice.

Want an AI policy that's fair and actually workable?

If you're figuring out how your team, applicants or students should use AI, and how to check the work without false accusations, talk to us. We help businesses set up practical AI workflows with sensible guardrails. See how we approach AI automation.