What did the Caine Prize check, and what did it find?

It says it held interviews with the author, the judging panel and the publication that first published the story, had an independent investigator run a metadata analysis of the author's drafts, and received documentation from the author and the publisher. It says 4 separate Word documents were created as drafts, which it puts at up to 27 hours of writing and editing. It reports that the review "found no evidence that the story was produced using generative AI", and that the award stands.

Does "no evidence" mean no AI was used?

Not by itself. The prize's finding is "no evidence". In our reading that is not the same as a finding that no AI tool was used at any stage, and the release does not claim it. The pages we read do not publish the review's full report or method, so we cannot say more. We take no side on the story.

What should I do if a detection tool gives a score on a contractor's or applicant's work?

In our reading, do not act on the score alone. Ask for dated drafts, version history and file details, talk to the person, and follow a written rule you set before any dispute. The prize's policy asks permission before work is run through a detector and has a step where the matter is assessed using human judgment before the decision is communicated. We are not lawyers: if a contract, a job or a payment depends on the outcome, ask your counsel or HR.

What is the Caine Prize changing for 2027?

It says submissions open on 23 November and lists five measures: an AI declaration from publishers for each submitted story, an AI declaration from shortlisted writers, guidance for judges on raising concerns about AI use, a published process for reviewing future concerns, and an AI Use Policy. The policy says AI-generated content, even if later edited or rewritten, is not eligible for submission. It also allows assistive tools such as spelling and grammar support, where they do not generate the story. It applies to submissions from 23 November 2026.

Caine Prize Says AI Detection Alone Is Not Reliable

Q: Can an AI detection tool prove that a text was written by AI?

The Caine Prize says AI detection tools alone are "not a reliable basis for determining authorship". Its AI Use Policy says they may be used "only as an initial indicator" and not as conclusive evidence. The pages we read give no error rates, and we did not test any tool. In our reading, a score is a reason to ask questions, not a conclusion.

A stack of translucent glass sheets at an angle, each with blank grey bars standing for lines of text, and a thin vertical line of red light with glowing dots running down through the stack. A round magnifying glass with a dark metal rim rests on the top sheet, on a near-black glossy surface.

The short version

The Caine Prize says an article alleged that its winning story may have been written using generative AI. The prize opened a review on 11 October 2026. In a release dated 11 November 2026 it says the review found no evidence of that, and that AI detection tools alone are not a reliable basis for determining authorship.

In plain words

How would you show how a piece of writing was made? One literary prize faced that question when, the prize says, an article alleged that its winning story may have been written using generative AI (software that writes text from a typed request).

The Caine Prize says it found no evidence of that. It also says that AI detection tools alone are not a reliable basis for determining authorship. In our reading, the lesson for a business is to ask how the work was made before acting on a detector score.

Status as of November 12, 2026. This covers one literary prize's public statement and its AI Use Policy. The facts are the prize's, the recommendations are ours, and it is not legal advice: we are not lawyers.

Why is a prize review in our AI Failures hub, where every case is paired with a control (a safeguard)? We do not say that any tool, person or publication was wrong. The risk, in our reading, is a business acting on a detector score or a suspicion alone. The control is a written review process, backed by drafts, version history and other evidence, with a named person responsible for the decision. Our Nov 11 piece, on a Bombay High Court order, was about checking what a document cites. This one is about evidence of how something was made.

What did the Caine Prize say happened?

The Caine Prize for African Writing is a literary prize. On 11 November 2026 its Board of Trustees published a press release about its 2026 winning story. The release says a review opened on 11 October 2026, after an article "alleged the story may have been written using generative AI". (Our gloss: generative AI means software that creates text, images, audio or video from a typed request.)

The Board calls it an "internal review" and says the award stands. We have not read the article, the story or any review report, and we take no side. We cover only what the prize itself says.

Date as the prize prints it	What the prize says
April 2025 to June 2025	The writer's process "started in April 2025" and was "completed in June 2025".
11 October 2026	The review opened, "after an article" alleged the story "may have been written using generative AI". The release gives no date for the article.
1 November 2026	The prize's AI Use Policy is accepted and takes effect. This is ten days before the release (our arithmetic). The release lists the policy among measures introduced "alongside this decision"; neither page says whether the policy applied to the review.
11 November 2026, noon, London	The Board's press release: the review "found no evidence that the story was produced using generative AI", and the award stands. From 11 October that is 31 days (our arithmetic).

What did the prize check?

The release lists three kinds of evidence.

Interviews with the author, the 2026 judging panel and the publication where the story was first published.
Metadata analysis of the author's drafts. Metadata is the hidden information a file keeps about itself, such as when it was created and changed (our plain-words gloss). The prize says an independent investigator did this analysis.
Documentation from the author and the publisher "showing the story's development over time".

The prize also gives numbers. It says 4 separate Word documents were created as drafts, which it puts at "up to 27 hours of writing/editing". The release does not say who counted the hours or how, and it does not say what the metadata showed. The prize calls it an internal review and says the independent part was the metadata analysis.

What does "no evidence" mean?

The prize's sentence is: "The review found no evidence that the story was produced using generative AI." That is the prize's finding, in its words. The release adds that the winner remains the winner and that "The Board considers the matter closed."

In our reading, "no evidence" is not the same as a finding that no AI tool was used at any stage. The release does not claim the second. The pages we read do not publish the full review or its method, so we cannot check either.

What does the prize say about AI detection tools?

One sentence in the release is useful beyond literature: "AI detection tools alone are not a reliable basis for determining authorship." An AI detection tool is software that scores how likely a text is to have been produced by AI.

Think of it as a metal detector at a gate (a comparison, not the prize's). A beep is a reason to look in the bag. It is not a verdict about the person. This sentence is the prize's statement, not a test result. The pages we read give no error rates for any tool, and we did not test one. They do not name any detection tool or score.

The prize's AI Use Policy goes further and, in our reading, is the more useful of the two pages for a company. It was accepted on 1 November 2026. It says detection tools may be used "only as an initial indicator" and not as conclusive evidence. It says permission to run work through a detection tool must be sought from the writer and publisher first. And it sets a fair process of five steps, in the policy's words, lightly shortened:

Record the concern and the basis on which it arose.
Contact the submitter and invite an explanation and supporting information.
Consult relevant stakeholders where appropriate and lawful.
Assess the matter using human judgment, taking account of the limitations of detection tools.
Communicate the decision and any consequences clearly.

This is a rule for a literary prize, not a template for companies. The policy also sets the charity's own workplace rules. They do not allow AI to make final decisions about recruitment, employment, funding, eligibility, judging or disciplinary matters without meaningful human review, and they say a named person remains responsible for the task and the final output. These apply to the charity's staff, trustees and contractors, not to every company.

What is the prize changing for 2027?

The release says submissions open on 23 November and lists five measures:

an AI declaration for all stories that publishers submit;
an AI declaration for all shortlisted writers;
guidance for judges on assessing and raising concerns about AI use;
a published process for reviewing future concerns;
an AI Use Policy, "developed for guidance".

The policy adds a rule on eligibility: content generated by AI, "even if later edited or rewritten", is not eligible for submission. It also says any material use of AI must be disclosed when a work is submitted. Both apply to submissions from 23 November 2026. The policy still allows assistive tools for accessibility, transcription, spelling, grammar or formatting, where they do not generate the story or replace the writer's creative contribution.

What does the release leave out?

Neither the press page nor the policy states:

which detection tool, if any, produced a score behind the allegation, or any score;
any error rate for a detection tool, or any test the prize ran;
the full review report or its detailed method;
the date of the article, or its content beyond the release's one-clause description of the allegation.

What can you do this week?

Suppose you hear that a tool gave a contractor's, applicant's or supplier's work an "AI-made" score. Neither document is advice for companies. The steps below start from what the prize did and wrote; anything marked "in our reading" is ours.

Do this week

Three steps for an owner or manager who might be handed a detector score. This is not legal advice.

In our reading, treat a detector score or an accusation as a reason to ask questions, not as a conclusion. The prize's policy says detection tools may be used "only as an initial indicator", asks permission first, and puts a human-judgment step before the decision is communicated.
In our reading, ask the person to show how the work was made: dated drafts, the document's version history, file details such as creation and edit times, and a conversation about their choices. The prize looked at interviews, draft metadata and documentation from the author and the publisher. These can support a review; they do not rule out every use of AI. Ask only for what you need: collecting someone's files and metadata may carry privacy duties, so ask your counsel first.
In our reading, before any dispute, write down your own AI-use rule (what is allowed, what must be declared, who decides) and your review steps (who is told, what evidence counts, how the person can respond). We are not lawyers. If a contract, a job, a grade or a payment depends on the outcome, ask your counsel or HR.

Book an AI discovery call

How do we know?

How we know

On November 12, 2026 we read the Caine Prize's Press page (the release dated 11 November 2026, noon, London) and the AI Use Policy PDF that it links. Every quote above was matched word for word against those two pages. We did not read the article that started the review, the story or any review report (the pages we read publish none), and we ran no detection tool. This page covers only what the prize says. We took no facts from news reports. The release names several people; we leave every name out, because the facts do not need them. The Press page can change.

Caine Prize Says Its Review Found No Evidence a Winning Story Was Produced Using Generative AI