Judging a Claim That Something Was Written by AI

"This was written by AI" is a claim about a fact: who or what produced a piece of text. It deserves the same treatment as any other factual claim. Ask what the evidence is, how good it is, and what happens to someone if the claim is wrong. The earlier lessons in this course explain how detectors work and fail. This one is about judgment: how to weigh a claim when it lands on your desk, how to protect yourself if it is aimed at you, and where the honest lines are.

Start with the stakes and the source

Before you look at the text, look at the situation. Three questions sort most cases.

Who is making the claim, and what do they gain? A teacher who suspects a student, a rival who wants a competitor's book pulled, a campaign that wants an opponent's op-ed dismissed and a commenter who wants other comments ignored are all making the same kind of claim with very different motives.

What decision rides on it? A private conversation needs little evidence. A failing grade needs more. A firing, an expulsion, a retraction or a public accusation needs a great deal. Match the evidence you require to the harm a mistake would cause.

Does authorship even matter here? Sometimes it does not. A public comment that makes a sound point about a road project makes the same point whoever typed it. A fake review is a problem because the reviewer does not exist. A student essay matters because the assignment was meant to show the student's own thinking. Be clear about which harm you are trying to prevent, because that tells you which evidence is relevant.

A step-by-step method

1. Write the claim down precisely. "The whole thing was generated." "Parts were generated and pasted in." "AI was used without the disclosure this outlet requires." Each is a different claim with different evidence.
2. Check the facts in the text. This is the most productive step and the one people skip. Look up every citation, quote, case, statistic and name. Invented sources are among the best-documented failures of language models (see Hallucination, grounding and reasoning and lesson 5, The tells). A fabricated source is a problem whoever wrote it, and it is far stronger evidence than any style pattern.
3. Look for hard signs. Text addressed to a chatbot user ("Certainly! Here is a 500-word essay on..."), a knowledge-cutoff disclaimer, an unfilled placeholder like "[Name]," or interface leftovers. Wikipedia's fast-deletion rule for machine-written pages, criterion G15, names signs of this kind, including citation artifacts such as "oaicite" and "turn0search0" that some chat interfaces leave behind.
4. Check identity and provenance. Does the named author exist? Is the commenter a real resident? Did the freelancer's sources and places check out? In lesson 7, the fake freelancer that fooled Wired was caught this way after two detectors called the copy human.
5. Ask for process evidence. Drafts, notes, version history, source files, and if AI was used, the chat logs. Ask politely and without presuming guilt. Many people have this material and have never been asked for it.
6. Talk to the writer. Ask them to explain the piece: why it is organized this way, where a fact came from, what they would change. Listen for understanding. Polish proves little.
7. Use a detector only as a lead, if at all. If you run one, know its claimed false positive rate and what that means at your volume (lesson 6, Accuracy and false positives, works through the arithmetic). Never treat a score as a finding. Never paste the text into a chatbot and ask whether it wrote it: OpenAI says ChatGPT has no knowledge of what content is AI-generated and its answers to that question have no basis in fact.
8. Weigh the whole record and decide in proportion. Strong evidence on one side and none on the other is a decision. Weak signals pointing both ways are not. When in doubt at high stakes, the absence of proof should protect the person accused.
9. Write down what you checked. If the decision is challenged, a record of the steps above protects everyone, including you.

What evidence actually settles authorship

Evidence	Strength	Why
The writer's own account, with the chat logs or shared conversation	Strong	Direct record of what the tool produced. OpenAI suggests teachers ask students to share their ChatGPT conversations.
Fabricated citations, quotes or facts	Strong for the misconduct, moderate for the method	Proves the work is unreliable. A person can invent sources too, but invented sources that look plausible are a known model failure.
Leftover chatbot text or interface artifacts	Strong	Hard to explain any other way. These are the signs Wikipedia's G15 relies on.
Version history showing gradual work, or a single paste	Strong either way, with care	Shows how the document grew. A single paste can also mean the writer drafted in another program, so ask.
Drafts, notes and sources	Moderate to strong	Hard to fake in volume after the fact.
A conversation or oral explanation	Moderate	Shows understanding. Nervous people do badly, so it cuts one way more reliably than the other.
Detector score	Weak	Error rates vary by tool, text length and writer, and paraphrasing defeats most detectors (lessons 3 and 6).
Style and vocabulary tells	Weak	Many human writers share them. Wikipedia's guideline requires evidence "beyond basic stylistic or linguistic indications" before sanctions.

Wikipedia's rules show how careful a high-volume institution has to be. Its guideline bans model-written article text, yet it warns that some editors naturally write in a style that resembles model output and tells reviewers to look at an editor's whole pattern of recent edits. And G15 is limited to what the page itself shows; an author's admission does not qualify for that fast track, and those pages go through slower deletion processes.

For officials: public comments and constituent mail

Agencies, boards and legislators have the hardest version of this problem: thousands of messages, little staff time, and a duty to hear the public.

The research is sobering about reading for style. In October 2019, researcher Max Weiss built a bot using publicly available AI methods and submitted 1,001 generated comments to Medicaid.gov on the Section 1115 Idaho Medicaid Reform Waiver over four days. They made up 55.3 percent of all 1,810 comments. He then withdrew them. When people were asked to sort a sample of the comments into human and bot, they got 49.63 percent right, no better than guessing. Weiss, who published in Technology Science in December 2019, estimated the cost of such a campaign at under $100. He used the smallest version of OpenAI's GPT-2, with 124 million parameters.

In 2020 Sarah Kreps and Douglas Kriner of Cornell sent more than 32,000 emails to about 7,000 state legislators, some written by students and some generated with GPT-3. Legislators responded to 17.3 percent of the human emails and 15.4 percent of the AI ones, a small difference. On gun control and health policy the rates were virtually identical; on education the AI emails did better. Afterward, legislators said they spotted fakes by missing local detail, like geography, and some in small districts could tell by the name alone.

The largest documented flood of fake comments was not mainly about writing style. The New York Attorney General's May 2021 report found that nearly 18 million of the more than 22 million comments in the FCC's 2017 net neutrality proceeding were fake. More than 8.5 million impersonated real people, submitted by lead generation firms in a campaign funded by the broadband industry. Another 9.3 million, supporting net neutrality, used fictitious identities, and most of those came from a single 19-year-old college student using automated software. Fraud on both sides, and the core of it was identity.

So the practical checks for comment review are about people and substance:

• Count distinct, verifiable commenters. Submissions can be multiplied at no cost. Verify a random sample by contacting them.
• Watch for impersonation: names and addresses of real people who say they never commented.
• Weigh arguments on their merits and new facts on their accuracy. A comment that adds a fact you can verify is useful whoever wrote it.
• Do not throw out a comment because it "sounds like AI." People use these tools to write to their government, including people who are not confident writers. Discarding them on style alone silences real residents.
• Be candid about volume. A thousand near-identical messages show organized effort, which is legitimate. They do not show a thousand independent opinions, and the record should say so.
• Use in-person hearings, surveys and local detail as checks. Kriner pointed to district polling and town meetings as sources that are harder to fake.

Protecting your own writing

Anyone who writes for school, work or publication can be accused, and the best defense is a record made before anyone asks.

• Write where history is kept. Google Docs records earlier versions and who changed what and when. Anyone with edit permission on the file can browse them, so keep edit access to your own work. If you draft in another program, save dated copies as you go.
• Keep the mess. Outlines, notes, interview recordings, photos of handwritten pages, browser bookmarks and a list of sources you checked. Delete nothing until the piece is long settled.
• If you used AI, keep the logs. A saved conversation shows exactly what the tool did and did not contribute. That is often the fastest way to end a dispute.
• Consider a process recorder for high-stakes work. Grammarly says its Authorship feature records whether sentences were typed, pasted from a website or suggested by Grammarly, and can replay a document's creation for a teacher or editor. That is the vendor's description. Any recorder only proves what happened inside it.
• Know the disclosure rules before you start. The Copyright Office expects registrants to disclose AI-generated material and exclude it from the claim. ICMJE journals want AI use described in the cover letter and the paper. Amazon KDP and Medium have their own lines, covered in lesson 7. Under Article 50(4) of the EU AI Act, AI-generated text published to inform the public needs disclosure unless a person with editorial responsibility reviewed it (lesson 8).

If you are accused, ask what the evidence is, in writing. Offer your drafts and history. Ask for a conversation about the content. If the only evidence is a detector score, say so plainly and point to the vendor's own guidance; Turnitin, for one, says its indicator should not be the sole basis for adverse action (lesson 7).

Disclosure

Disclosure solves most of what detection cannot. A reader who knows how a piece was made can judge it on that basis. The rules described in lessons 7 and 8 converge on the same idea: say what the tool did, say what a person checked, and take responsibility for the result. Course 16's lesson on ethics and disclosure gives sample wording. A vague line like "AI may have been used" tells the reader nothing and invites suspicion of the whole piece.

The ethics of humanizers and of accusing

A humanizer is a tool sold to make AI text pass detectors. Research shows why they work. Kalpesh Krishna and colleagues (arXiv, 2023) built a paraphrasing model called DIPPER and showed it evaded several detectors, cutting DetectGPT's detection accuracy from 70.3 percent to 4.6 percent at a fixed 1 percent false positive rate. Vinu Sankar Sadasivan and colleagues (arXiv, 2023) tested watermarking, trained classifiers, zero-shot methods and retrieval-based detection against recursive paraphrasing and found it can significantly reduce detection rates while only slightly degrading text quality in many cases. They also showed the reverse attack: an attacker can spoof a watermark so that human-written text gets flagged as AI-generated. Keep that second finding in mind the next time a mark or a score is offered as proof against someone.

What makes a humanizer right or wrong to use is the purpose. Running text through a humanizer in a setting that forbids AI, or requires disclosure, is concealment: the point is to make the reader believe something false about how the work was made. In a setting with no rule and no expectation, rewording machine text is just editing. If you would be uncomfortable telling your teacher, editor or client what you did, you already know which case you are in.

Accusing someone carries its own duties. Weixin Liang and colleagues (Patterns, July 2023) found that widely used detectors consistently misclassified writing by non-native English speakers as AI-generated while classifying native writers' samples accurately. Vanderbilt pointed to that bias when it turned off Turnitin's detector. The people most likely to be falsely flagged are often the people with the least power to fight back. So the person making the claim carries the burden. Ask before you accuse. Accuse in private before you accuse in public. And if the evidence turns out to be wrong, say so as clearly as you made the claim.

The strongest sources to follow

• Primary rules. The EU Commission's Code of Practice page for Article 50, the FTC rule at 16 CFR Part 465, the Copyright Office AI page and your own state's statutes. Lesson 8 links them.
• Independent evaluations over vendor claims. Peer-reviewed and preprint studies like Liang et al., Krishna et al. and Sadasivan et al., which test detectors against real writers and real evasion.
• Institutions that publish their reasoning. Vanderbilt's 2023 explanation, Wikipedia's guideline and deletion criteria, ICMJE's recommendations and Nature Portfolio's risk levels show how careful organizations weigh evidence.
• Vendors' own cautions. When a detector company says its score should not be the sole basis for action, that is the most honest line on its website. OpenAI's help page on detection is in the same category.
• Studies of public input. Weiss 2019, Kreps and Kriner 2023 and the New York Attorney General's 2021 report, for anyone who reviews comments or constituent mail.