Jamaican students are being told to learn AI and, in the same semester, being penalised for using it. I have trained Jamaicans in AI every week for seven years and I sit on the National AI Task Force, so I hear from both sides: lecturers worried about integrity, and students accused on the strength of a percentage on a screen.
My position is simple. AI content detection in 2026 is not reliable enough to be the only basis for an academic or professional penalty. The reasons are technical, and anyone who sets or follows a policy should know them.
How detectors reach their score
A detector is itself a machine learning model. It reads a passage and estimates the probability that a language model wrote it, mainly from two statistical properties.
- Perplexity measures how predictable each word is. Language models choose likely next words, so their text scores as predictable. People choose odd words, make strange comparisons and break patterns.
- Burstiness measures how much sentence length and complexity vary. Human writing mixes short sentences with long ones and dense paragraphs with loose ones. Model output tends to be more even.
The score is a statistical guess. The detector does not check a database or find a fingerprint, and a "90% AI" result is a model's confidence, not a measurement of how much of the text a machine wrote.
Why the accuracy claims break down in real classrooms
Vendors report accuracy from tests on clean samples: text wholly written by a person or wholly generated. Real submissions are mixed. Students brainstorm with AI and then write, or draft and ask for edits, or write most of an essay and generate one section. Accuracy falls on exactly this mixed text.
The false positive, human writing flagged as AI, is the error that matters. Rates vary by tool and by the kind of text, and published studies range from under 1% to well above 10%. Even at the low end, a university that runs thousands of scripts a semester through a detector will flag some students who wrote every word themselves.
OpenAI withdrew its own AI text classifier in July 2023 because of its low accuracy. At launch the company said it caught 26% of AI-written text and wrongly labelled 9% of human text as AI.
The bias against second-language and Creole-influenced writers
A 2023 Stanford study by Liang and colleagues ran essays written for the TOEFL English exam through seven detectors. On average the detectors labelled 61% of these human-written essays as AI-generated, while essays by US eighth-graders were classified almost perfectly.
The cause is the same statistics the detectors rely on. Writers working in a second language use a smaller vocabulary and more regular sentence patterns, which lowers perplexity. Many Jamaican students write Standard English while thinking in Patois, and careful, formal, rule-following prose from those students is the kind of text these tools misread.
Detection will not catch up with generation
Detection is an adversarial problem. Each generation of language models is trained to produce more natural text, and the companies building them spend far more than detection vendors do. Paraphrasing defeats most detectors: thirty minutes of rewording AI output usually brings the score down. So detectors mostly catch the least careful users while missing the deliberate ones.
Watermarking is the more promising route. Google has released SynthID Text, which embeds a statistical signal in generated text that a matching tool can check. It only works when the generator adds the mark, and heavy editing weakens it, so it will not help with text from open models that carry no mark.
What students at UWI, UTech and other institutions should know
Policies differ between institutions, faculties and individual courses. Read the course outline, and if it says nothing about AI, email the lecturer and keep the reply.
Using AI to understand a concept, brainstorm, check grammar or find sources is generally treated differently from submitting generated text as your essay. The line blurs in between, which is why the course's own policy governs.
Write in Google Docs or Microsoft Word with version history on. A timeline showing an essay growing from notes to final draft over several days is stronger evidence than any detector score. Put your own reasoning and Jamaican examples into your work; they are hard to fake and easy to defend in a conversation.
What freelance writers and content creators should know
Some clients and platforms run submissions through detectors, and a false positive can cost a contract. Agree with each client at the start how AI use will be handled and what evidence settles a dispute, and keep your research notes and version history for every job.
Most professional writers use AI somewhere in their process in 2026. Say how you use it. Clients increasingly treat an AI-assisted draft the way they treat spell check; what they will not forgive is finding out afterwards.
What lecturers and institutions should change
At present every institution, faculty and often each lecturer sets a separate rule, which confuses students and produces inconsistent penalties. Institutions need one written policy that separates prohibited from permitted use, explained at the start of each semester. A total ban will be ignored in the way a calculator ban would be.
Faculty need training on what detectors can and cannot do before they judge a case. And assessment has to change: if a prompt pasted into ChatGPT can complete an assignment, redesign the assignment. Tasks that require reflection on a class discussion, local application or submitted drafts resist misuse and teach more.
Your students are using AI. What matters is whether you are teaching them to use it with integrity, skill and judgement.
Prompts that use AI without writing the work for you
I am writing an essay about [topic] for my [course] at [institution]. Do not write the essay.
Help me brainstorm arguments for and against [specific thesis].
List the points I should consider and the kinds of academic sources I should look for.
Review this paragraph from my assignment for grammar, clarity and logical flow.
Do not rewrite it. Point out specific problems and explain why each one is a problem.
Here is my paragraph: [paste your text]
Explain [complex concept] in simple terms, then give me two examples from Jamaica or the Caribbean.
I want to understand it well enough to explain it in my own words.
Act as a study partner. Ask me five questions about [topic], starting easy and getting harder.
After each answer, tell me whether I am right and explain any mistake.
What to do next
- Students: before your next assignment, turn on version history in Google Docs or Word and write in one file from first notes to final draft. Ask the lecturer by email what AI use the course allows, and save the answer.
- Students flagged by a detector: within the first week, ask in writing for the full detector report, the tool used and the threshold applied, then send your version history and offer to discuss the essay in person.
- Freelance writers: add one clause to your standard contract stating how you use AI and that a detector score alone is not grounds for rejection or non-payment.
- Lecturers: rewrite one assessment this term so it needs submitted drafts or a short oral follow-up, and treat any detector score as a prompt for a conversation, never as the finding.
- Deans and academic registrars: publish a single institution-wide AI policy before the next semester, with a defined appeal route for students accused on detector evidence.
Frequently Asked Questions
What should I do if I am accused of using AI because of a detector score?
Stay calm and ask, in writing, for the detector report, the name of the tool and the score threshold your institution uses. Send your version history, notes and sources, and ask for a meeting where you can explain your argument and answer questions about it. Point out that Turnitin's own guidance says its AI score should not be the sole basis for a misconduct finding. If the decision goes against you, use the faculty's formal academic integrity appeal process and keep copies of every message.
Has any AI company stopped offering detection because it did not work?
Yes. OpenAI launched an AI text classifier in January 2023 and withdrew it in July 2023 because of its low rate of accuracy. At launch the company reported that it correctly identified 26% of AI-written text and wrongly labelled 9% of human-written text as AI. When the company that builds the generator cannot reliably detect its own output, a third-party score deserves caution.
How should I read a Turnitin AI score?
Treat it as an estimate of how much of the text resembles model output, not as proof. Turnitin says low scores are less reliable and, since 2024, shows an asterisk instead of a number for scores between 1% and 19%. Its guidance to institutions is that the score should start a conversation with the student and not serve as the only evidence of misconduct.
Does using Grammarly or a translation tool count as AI use?
It depends on the policy, and many policies now treat them differently. Basic spelling and grammar correction is usually permitted, but Grammarly's generative features, which rewrite whole sentences, are treated as AI assistance by many institutions. Translating a draft you wrote in another language is also AI use under some policies. Disclose any tool that changed your wording, not just ones that generated text.
What does CXC say about AI in CSEC and CAPE School-Based Assessments?
The Caribbean Examinations Council sets the rules for School-Based Assessments, so check its current published guidance on AI rather than relying on what your school did last year. Teachers must be able to authenticate SBA work as the candidate's own, which makes drafts and teacher check-ins the practical evidence. If your school has no written position, ask the SBA teacher before you start, because the moderation process can question work late in the year.