What Is AI Detection and How Does It Work?

AI detection is software that estimates whether a passage resembles patterns associated with AI-generated writing. It does not read an authorship history, identify the person who wrote the text, or prove that an AI system produced it. A detector score is an uncertain signal that should be interpreted with the text, its context, and evidence of the writing process.

Different products use different methods. OpenAI’s retired classifier was a language model fine-tuned on paired human-written and AI-generated responses to the same prompts. Turnitin describes its AI Writing Report as identifying qualifying prose that its model judges could have been generated by an LLM or generated and modified by an AI paraphraser or bypasser. These descriptions apply to those products and should not be generalized to every detector.

What AI detection means

A detector receives text and returns one or more of the following:

A classification such as likely human or likely AI-generated.
A probability, confidence value, or percentage.
Highlighted passages that the system considers more likely to be AI-generated.
Warnings that the input is too short, unsupported, or outside the product’s intended content type.

The output describes how the text resembles patterns in the detector’s training data or rules. It does not establish who typed the words, which tools were used, or whether a person edited an AI draft.

How an AI text detector works

1. The system is trained or designed to recognize signals

A statistical detector learns differences between examples labeled as human-written and AI-generated, or combines language features with other signals. OpenAI said its 2023 experimental classifier was fine-tuned on paired human and AI responses to the same prompts. It divided examples into prompts and responses, generated model responses for those prompts, and adjusted its confidence threshold to limit false positives. This was one implementation, not a disclosure of every vendor’s design.

2. The submitted passage is analyzed

The detector may examine word choices, sentence structure, repetition, predictability, transitions, and other features learned during development. Vendors generally do not expose every feature or model detail. A passage can therefore receive a score without the user being able to reproduce the calculation.

3. The result is converted into a report

Some tools show a score for the document; others mark spans or report categories. Turnitin’s AI percentage is separate from its similarity score: similarity concerns overlapping text, while the AI report concerns prose its model judges may have been generated or modified by AI. Do not treat those two numbers as interchangeable.

4. A human interprets the result

The report is an investigative signal. It is not an authorship record, a confession, or a standalone finding of misconduct. For consequential decisions, review drafts, revision history, source notes, conversations about the work, and the applicable policy.

What an AI detection score means

Output	What it can tell you	What it cannot prove
“Likely AI” or a high percentage	The passage resembles patterns in that product’s AI examples.	That an AI system wrote it, which system was used, or that the writer violated a rule.
“Likely human” or a low percentage	The passage did not match the detector’s AI patterns strongly.	That no AI assistance was used.
Highlighted sentences	Regions the model considers more suspicious or characteristic.	That those sentences have a different author.
No result	The input may be too short, unsupported, or outside the product’s qualifying format.	That the content is human-written.

A score is conditional on the product, model version, language, input length, content type, and threshold. Record those details whenever a result matters.

Why AI detectors produce false positives and false negatives

A false positive is human-written text labeled as AI-generated. A false negative is AI-generated text that receives a human-like result. Both occur because the detector infers from patterns rather than observing authorship.

OpenAI reported that its retired classifier marked 26% of AI-written English challenge-set text as likely AI-written and incorrectly marked 9% of human-written English challenge-set text as AI-written. Those figures describe that classifier and test set; they are not universal error rates. OpenAI discontinued the classifier on July 20, 2023 because of its low accuracy.

False positives can be more likely when writing is short, formulaic, highly edited, translated, or produced by a writer whose style resembles training examples. False negatives can occur after paraphrasing, human editing, or other transformations. A detector’s confidence does not remove these uncertainties.

Length, language, and content limits

Coverage differs by product. OpenAI described its retired classifier as very unreliable below 1,000 characters, significantly worse outside English, and unreliable on code. Do not transfer that threshold to another detector.

Turnitin’s current AI Writing Report guide specifies at least 300 words of qualifying long-form prose, a maximum of 30,000 words, and a file size below 100 MB. It lists English, Spanish, Japanese, and Arabic support. The guide says poetry, scripts, code, bullet points, tables, and annotated bibliographies are not reliably detected as qualifying prose. Its English detector includes AI-paraphrasing and bypasser detection; the Spanish and Japanese versions do not.

Turnitin also says results above 0% and below 20% are not shown as a precise percentage and use an asterisk because its testing found a higher incidence of false positives in that range. If a report was generated before July 8, 2024, users may see a numeric score below 20%. This is Turnitin guidance, not a general cutoff for AI detection.

AI detection versus content provenance

Detection and provenance answer different questions:

Detection: Does the wording resemble text associated with AI generation?
Provenance: Is there metadata, a signature, or an embedded signal about where the content came from?

Provenance can provide origin information without inferring it from prose. OpenAI has discussed cryptographically signed metadata and text-watermarking research, while noting that watermark false positives can accumulate at large scale. Metadata can also be removed when content is copied or transformed. The absence of a provenance signal does not prove human authorship.

Do AI detectors work?

They can provide a useful prompt for review, but no detector score should be treated as proof. OpenAI’s educator guidance answers the question “Do AI detectors work?” with “In short, not in our experience.” Turnitin’s report guide warns that its model may misidentify human-written, AI-generated, and AI-paraphrased text and should not be the sole basis for adverse action against a student.

A 2023 study that evaluated 12 publicly available tools and two commercial systems concluded that the tested tools were not accurate or reliable overall, and that obfuscation worsened results. That study is historical evidence about its test set, not a current ranking of every detector.

A responsible workflow for using a detector

Check eligibility. Confirm the language, content type, word count, file size, and product version meet the detector’s requirements.
Preserve the original. Keep the submitted file unchanged and record the date, product, model or report version, and settings.
Run the report as a screening step. Do not alter the text repeatedly just to chase a lower score.
Review highlighted passages in context. Look for ordinary explanations such as quotations, templates, technical phrasing, translation, or consistent personal style.
Gather process evidence. Review drafts, version history, research notes, source records, and relevant AI conversations.
Invite an explanation. Ask the writer to explain sources, revisions, and decisions in a constructive conversation.
Apply policy consistently. Use the institution’s rules and human judgment; never make a consequential decision from the score alone.

Practical example: interpreting a report

Suppose a report marks 18% of a qualifying English essay and highlights two paragraphs. The correct conclusion is that the product found some text resembling its AI patterns. It is not correct to conclude that 18% of the essay was written by AI or that the writer committed misconduct. Check whether the highlighted passages are quotations, standard definitions, heavily edited sections, or short formulaic paragraphs, then review the writing process.

Common mistakes and troubleshooting

Problem	Likely cause	Fix
The result says the document is unsupported.	The file, language, content type, or word count is outside the product’s requirements.	Check the named vendor’s current guide; submit qualifying long-form prose in a supported language.
A short passage receives an extreme score.	Short inputs provide too little evidence and are unstable.	Do not generalize from the result. Use a longer qualifying sample or gather process evidence.
Human writing is flagged.	False positive, formulaic phrasing, translation, or strong editing.	Review the highlighted text and drafts; do not treat the score as proof.
AI-assisted writing receives a human-like score.	Editing, paraphrasing, or model differences changed the detectable patterns.	Do not treat a low score as proof that no AI assistance occurred.
The AI score and similarity score conflict.	They measure different things.	Interpret AI reporting and text-overlap reporting separately.
A result changed after rerunning.	The vendor changed its model, threshold, or report behavior.	Record report dates and versions; compare like with like.
Code, tables, or bullet points are flagged.	The detector is designed for prose and may not reliably support those formats.	Follow the product’s content limitations and use human review.

Performance, reliability, and cost considerations

Performance: Longer qualifying prose may provide more signal, but length does not make a score proof.
Reliability: Results are product-specific and can change with language, genre, editing, paraphrasing, and model updates.
Reproducibility: Save the exact input and report metadata so another reviewer can understand what was evaluated.
Cost: Check the vendor’s current limits and plan terms. A free or inexpensive report is not evidence of accuracy.
Governance: For academic or employment decisions, define when a report triggers review, who reviews it, and what additional evidence is required.

Or skip the browser setup

If you need to archive a detector report or capture a results page for review, ScreenshotNeo provides a website screenshot API. The DIY approach is to launch a browser, load the report, wait for it to render, remove overlays, and save an image. ScreenshotNeo handles that capture in one request.

See the ScreenshotNeo documentation for all options.

cURL

curl -G "https://api.screenshotneo.com/v1/shot" \
  -d access_key=YOUR_API_KEY \
  --data-urlencode url=https://example.com/report \
  -o report.webp

Python

import requests

r = requests.get(
    "https://api.screenshotneo.com/v1/shot",
    params={"access_key": "YOUR_API_KEY", "url": "https://example.com/report"},
    timeout=90,
)
r.raise_for_status()
open("report.webp", "wb").write(r.content)

Node.js

const q = new URLSearchParams({
  access_key: 'YOUR_API_KEY',
  url: 'https://example.com/report'
});
const res = await fetch(`https://api.screenshotneo.com/v1/shot?${q}`);
if (!res.ok) throw new Error(`HTTP ${res.status}`);
const fs = await import('node:fs/promises');
await fs.writeFile('report.webp', Buffer.from(await res.arrayBuffer()));

Cookie banners, newsletter popups, and chat widgets are removed before the shot. Bot checks, blank pages, failed loads, timeouts, and cache hits are never billed; response headers identify the page verdict and whether the shot was billed. An MCP server lets AI agents use take_screenshot, get_page_info, and capture_pdf. You get 1,000 screenshots a month free with no card; paid plans start at $5 for 3,000 shots. Create a free ScreenshotNeo account.

FAQ

Can ChatGPT tell me whether a text was AI-written?

No. An AI system’s answer about authorship is not verification. Use the document, its history, and independent process evidence.

Does a high score mean the writer cheated?

No. It means the passage matched the detector’s patterns. A high score requires human review and context.

Can editing make AI text undetectable?

Editing or paraphrasing can change a detector’s result, but neither a lower score nor a higher score proves origin.

Should every organization ban AI detectors?

Policy depends on the use case. If a detector is used, define it as a screening aid, document its limitations, and require human judgment for consequential decisions.

What should I save with a detector result?

Save the original text, report export or screenshot, product name, date, language, content type, model or report version, and the policy used to interpret it.