Detection in Practice: How Institutions Handle Suspected AI Text

Most institutions that deal with large volumes of writing have now been through at least one round of AI detection policy. Some bought detectors and trusted the scores. Some bought them and then turned them off. Many wrote rules that never mention detection at all and instead tell writers what they must disclose and what they will be held responsible for. This lesson walks through what those institutions actually do, sector by sector, as of October 2026. Lesson 6, Accuracy and false positives, explains why the scores are shaky. This one is about what people do with them.

Schools: from the score to the conversation

Detectors spread fast in K-12 classrooms. The Center for Democracy and Technology surveyed a nationally representative sample of 460 public school teachers in grades 6 through 12 in December 2023. As Education Week reported in April 2024, 68 percent said they used AI detection tools, up substantially from the year before.

Policy came later. Idaho is a good example of where state law now sits. Senate Bill 1227, signed March 19, 2026 and in effect July 1, 2026, added a chapter on generative AI to Title 33 of the Idaho Code. Section 33-7004 requires every school district and public charter school to adopt a policy on generative AI use by students and employees. The policy must define appropriate and prohibited uses and include safeguards for student privacy, data security, accessibility and academic integrity. The law does not tell districts to buy a detector, and it does not tell them to avoid one. That choice is local.

Universities got there first, and several pulled back. Vanderbilt University disabled Turnitin's AI detection feature in August 2023 "for the foreseeable future." Its reasoning was arithmetic. Turnitin had claimed a 1 percent false positive rate at launch. Vanderbilt submitted 75,000 papers to Turnitin in 2022. One percent of that is about 750 students whose work could have been wrongly labeled. Vanderbilt also noted that Turnitin had not explained how the tool decides, and that detectors had been found more likely to flag writing by non-native English speakers. Lesson 6 covers the evidence behind that last point.

Schools that kept the feature usually kept Turnitin's own warning along with it. Idaho State University's faculty guide repeats Turnitin's statement that the AI writing indicator "should not be used as the sole basis for adverse actions against a student," and that Turnitin "does not make a determination of misconduct." The vendor is telling you the number is a lead.

The company that makes ChatGPT says something blunter. OpenAI's help page for educators says that in its experience detectors don't work, that its own research did not find them reliable enough, and that a detector it tried to train labeled Shakespeare and the Declaration of Independence as AI-generated. The same page says ChatGPT itself has no knowledge of what is AI-generated, so asking it "did you write this?" produces answers with no basis in fact. OpenAI suggests teachers ask students to share their AI conversations and log how they used the tools.

What a process-based check looks like

The schools that stopped leaning on scores did not stop caring about cheating. They moved the evidence from the finished text to the work that produced it. A check built that way has four parts, and you can use them in any setting where someone hands you writing and says it is theirs.

1. Drafts and notes. An outline, a messy first draft, research notes with page numbers, a list of sources with the dead ends crossed out. People who write leave a trail. Absence of a trail is not proof of anything, since plenty of people write in one sitting. But a trail, when it exists, is strong evidence.
2. Version history. Google Docs keeps earlier versions of a file and shows who changed it and when, and anyone with edit permission can browse them. A paper that grew over four evenings looks very different from one that appeared in a single paste. Vendors also sell writing-process recorders; lesson 9, Judging a claim, covers what they record and how writers can keep their own proof.
3. A conversation. Ask the writer to walk you through the piece. Why does the second section come before the third? Where did this number come from? What did you cut? Someone who did the work can usually answer, even if they are nervous.
4. An oral check. For high-stakes work, a short spoken explanation without notes. It tests understanding, which is the thing the assignment was supposed to build anyway.

None of these steps is new; they are the same checks used in plagiarism cases. What changed is that the finished text stopped being reliable evidence of who wrote it.

Newsrooms

The Associated Press published its generative AI standards on August 16, 2023, The rules are about what AP's own staff do and what comes in from outside. Any output from a generative AI tool "should be treated as unvetted source material." Staff may experiment with ChatGPT but do not use it to create publishable content. AP does not alter any elements of its photos, video or audio, so no generative AI to add or remove anything. For material arriving from others, AP tells journalists to exercise due caution and diligence, and suggests ordinary verification: identify the source of the original content, run a reverse image search, and check whether trusted outlets have similar reports. The lesson on ethics and disclosure in course 16 covers AP and the 2026 SPJ code in more depth.

Notice what is not in AP's list. No detector. The checks are about provenance and facts.

The clearest test of that approach came in August 2025. Press Gazette reported on August 21 that at least six publications, including Wired, Business Insider and SFGate, had run work by a freelancer using the name "Margaux Blanchard." Four took the work down at first, and a fifth followed. The person was caught by an editor at a small outlet called Dispatch, Jacob Furedi, who received a pitch about a decommissioned mining town called Gravemont. He could find nothing about the town. He asked for the public records the writer said she had used. They never came. Gravemont does not exist. Press Gazette could not verify many of the people quoted in the published stories and found no record of a journalist by that name.

Wired's own account is the part to remember. A Wired editor had run the story through two third-party AI detection tools, and both said the copy was likely to be human-generated. The red flags were elsewhere: the writer insisted on payment by PayPal or check after failing to provide enough information for Wired's payment system. Wired said the story "did not go through a proper fact-check process or get a top edit from a more senior editor." The detectors passed it. Ordinary checking would have caught it.

Scientific journals

Journals moved early. On January 27, 2023, Science editor-in-chief H. Holden Thorp published an editorial whose title states the position: "ChatGPT is fun, but not an author."

The International Committee of Medical Journal Editors (ICMJE), whose recommendations many medical journals adopt, now spells out the rules in detail. The January 2026 edition says journals should require authors to disclose at submission whether they used AI-assisted technologies such as large language models, chatbots or image creators. Authors should describe the use in the cover letter and in the paper: writing help in the acknowledgments, data collection, analysis or figure generation in the methods. Chatbots should not be listed as authors because they cannot be responsible for the accuracy, integrity and originality of the work. Humans stay responsible for anything AI produced, including making sure there is no plagiarism in AI-made text and images. Reviewers must keep manuscripts confidential, which may rule out uploading them to AI tools, and must ask the journal's permission before using AI to help with a review.

Springer Nature, which publishes the Nature journals, now sorts AI use by risk. Its policy has three levels. Green covers assistive use that adds no new intellectual content, such as polishing language or suggesting structure; disclosure is encouraged. Amber covers uses that may shape interpretation, like drafting explanatory summaries or extensive writing support; permitted with human oversight, verification and disclosure. Red is not permitted: fabricating data, citations or results, presenting AI-generated analyses or conclusions as human work, using a language model to generate core research reasoning without disclosure, listing AI as an author, or delegating peer review to a model.

Publishers also screen. On June 12, 2024, Springer Nature announced Geppetto, a tool built to catch AI-generated fake content, which it called a classic sign of paper mill activity. Geppetto splits a paper into sections, checks the consistency of the text in each, and scores each section on the probability that it was AI-generated. A higher score starts a human check by Springer Nature staff. The company said the tool had already flagged hundreds of fake papers soon after submission. Look for that design in any screening system you are told about. A score that opens a human review is a different thing from a score that closes a case.

Search engines

Google's position is that how a page was made matters less than whether it helps anyone. In a February 8, 2023 post, Google said its focus is "on the quality of content, rather than how content is produced," and compared the moment to an earlier wave of mass-produced pages written by people.

On March 5, 2024, Google announced three new spam policies, one of them aimed at what it calls scaled content abuse: generating many pages "for the primary purpose of manipulating Search rankings and not helping users," producing large amounts of unoriginal content "no matter how it's created." The current policy lists examples. Using generative AI tools to generate many pages without adding value for users is one. Scraping feeds and running them through synonym swaps or translation is another, along with stitching together content from other pages and setting up multiple sites to hide the scale. Sites that violate spam policies may rank lower or disappear from results, and site owners hit with a manual action get a notice in Search Console and can ask for reconsideration.

So Google does not try to answer "was this written by a machine?" It asks a question it can answer from the outside: is this site pumping out pages to game rankings?

Marketplaces and platforms

Amazon Kindle Direct Publishing

KDP requires self-publishers to tell Amazon about AI-generated content (text, images or translations) when they publish a new book or republish an edited one. It draws a line that many other rules now copy. AI-generated means an AI tool created the content, "even if you applied substantial edits afterwards." AI-assisted means you created it and used AI to edit, refine, error-check or brainstorm. Assisted use needs no disclosure. Either way, the publisher is responsible for making sure the content follows the guidelines, including intellectual property rules. In September 2023 KDP also lowered the number of new titles an account can create; KDP confirmed the limit to the Authors Guild as three per day. A volume cap does not detect anything. It just makes flooding the store harder.

Medium

Medium requires writers to label stories that use AI assistance. Its policy says AI-generated writing, "disclosed as such or not," is not allowed behind the paywall in its Partner Program. Undisclosed AI-generated writing gets "Network Only" distribution, meaning it goes only to the writer's own followers.

Wikipedia

English Wikipedia went furthest. After a March 2026 request for comment, its content guideline says using large language models to generate or rewrite article content is strictly prohibited, with two exceptions: suggesting spelling, punctuation and grammar fixes to your own writing, and translating articles under a separate guideline. Persistent use outside those exceptions can get an editor blocked.

The same guideline is careful about proof. It says some editors naturally write in a style that resembles model output, so sanctions require "evidence beyond basic stylistic or linguistic indications," and reviewers should look at the editor's full pattern of recent edits and whether they follow content policy. Wikipedia's fast-deletion rule for machine-written pages, criterion G15, applies only to unambiguous signs: text addressed to the user ("Here is your Wikipedia article on..."), references that do not exist or point to the wrong thing, and technical leftovers from chatbot interfaces. Writing style alone does not qualify. Neither does an author's own admission, since G15 is about what the page itself shows; those cases go through the slower deletion processes instead.

Employers

Hiring has the thinnest public record of any sector here. What is known comes mostly from surveys. One, commissioned by the resume service TopResume and run through Pollfish on May 15 and 16, 2025, asked 600 U.S. hiring managers about AI in applications. In it, 19.6 percent said they would reject a candidate with an AI-generated resume or cover letter, another 20 percent called heavy reliance on AI a red flag, and 52 percent said AI help with proofreading or drafting was acceptable if the final product sounded human. Treat it as one commercial survey.

What a cover letter can prove has changed, though. A polished letter used to show some writing skill. Now it shows that someone produced a polished letter. If you hire and writing matters to the job, the test has to move to places where you can watch it happen: a timed writing sample, a work exercise, or an interview question about something the candidate wrote.

Where the rules agree

Sector	Main rule	Role of detection	What settles a case
K-12 and universities	Local integrity policies; Idaho requires every district to adopt one	Common, sometimes disabled; vendors say not the sole basis	Drafts, version history, conversation, oral check
Newsrooms (AP model)	AI output is unvetted source material	Not part of AP's stated checks	Source tracing, records, identity of the writer
Medical and science journals	Disclose AI use; AI cannot be an author	Screening tools that trigger human review	Editor and staff investigation, data and citation checks
Google Search	No scaled content made to manipulate rankings	Not needed; judged on purpose and value	Site behavior at scale
Amazon KDP, Medium	Disclose AI-generated content; limits on paywall or volume	Not described in the policies	The publisher's own disclosure and account history
Wikipedia	No model-written article text	Style alone insufficient	Hard signs on the page and the editor's edit pattern

Read across the rows and you see the same design. The rules are written around disclosure and responsibility, which a person controls, instead of detection, which nobody controls well. Where a tool is used, it sorts the pile and a human decides.