We screened every answer and every peer review in TGLF’s certification “Malaria: new tools to turn the tide” to check for undeclared use of AI, plagiarism, and other Honor Code violations. We were surprised by who we had to send warnings to, and why.
How do we learn from the experiences of health workers about introducing a scientific innovation to a community?
New tools against malaria are arriving: two vaccines, monoclonal antibodies, a new kind of medicine, nets treated with two insecticides, spatial repellents, and ivermectin used in a new way.
Each one raises the same practical question for a health worker.
What do I say to a mother, a father or a village leader so that they understand what the tool does, what it cannot do, and that it is not here yet?
That is the premise of “Malaria: New tools to turn the tide”, a free certification from TGLF, in English and French.
Anyone who works for health in a community affected by malaria can join, with no science degree required.
Content strictly follows the WHO Global Technical Strategy for Malaria and other global standards.
Most of the course time goes to the learner’s own role in introducing something new, with less on the tools themselves.
The learning questions make this concrete.
- In Question 1, the learner describes one real time they introduced something new, down to what a person said in reply.
- Question 2 asks them to plan what they will listen for in one family they know.
- In Question 3 they write the exact words they would use, including what they would say if a parent asked whether the new tool could harm their child.
Each answer rests on one person’s real-world experience.
That is why we ask for them, and why a generated answer gives the colleague who reads it nothing to learn from.
What are the steps in peer learning?
Learners work with each other in three steps.
In the first three steps, they answer the learning questions.
In Step 4, they give feedback to three colleagues against a rubric.
In Step 5 they respond to the feedback they received and say what they will do with it.
The course tells them that colleagues cannot finish without their feedback.
TGLF has described the design reasoning in The integrity layer: using assessment to scaffold peer learning at scale.
Two points apply here.
- Peer review in TGLF courses is a learning mechanism, not a measurement instrument: applying criteria to a colleague’s work develops the reviewer’s own judgement. The value sits in the exchange, so integrity depends on whether the exchange is real, and accountability runs between peers.
- When one learner writes a thin review, the person harmed is the colleague who wrote a real account and received one word back.
The Honor Code that every learner agrees to before starting makes three commitments.
- Your work is your own.
- Give feedback to your colleagues, because they depend on you.
- If you use an AI tool, say how and why.
The course allows AI to find ideas, improve grammar, understand difficult concepts or overcome language barriers.
It does not allow AI to invent experiences or write the analysis.
What we did
Claude Cardot, our first agentic AI hire, reviewed each individual submission and peer review.
She applies explicit rules, so that two runs on the same data give the same result.
We screened 126 English and 303 French learners on their answers, and 326 reviewers on their feedback.
Three design choices shaped what happened next.
- Every finding carries the learner’s own evidence: what the learner wrote, what colleagues typically wrote, and what the activity asked for.
- Thresholds come from the cohort, so a review counts as thin only relative to what colleagues in the same course wrote.
- And findings that depend on interpretation go to a person.
Formatting that looks like unedited AI text measures the text and says nothing certain about the author, so those findings wait for a human reader.
The course platform now lets learners write in rich text, so headings, bullets and bold are ordinary formatting and are not counted as a sign of AI. The flag looks for “AI tells” such as em dashes, which the editor does not produce, and for connecting words such as “crucial” or “notably”.
Learners who declared AI use, as the Honor Code asks, were not warned for it.
In the submission step, we identified 6 English and 7 French learners, who stay on the record as context only.
What surprised us and why
79 learners were flagged at the submission step: 27 in English and 52 in French.
73 of them had a diligent progress record.
They were taking part, and in places they took a shortcut.
Three patterns emerged:
- Text that looks like unedited AI output appeared for 7 English learners (6 percent of those screened) and 8 French learners (3 percent). Nearly all had declared AI use. One French learner declared none while the text suggested otherwise.
- Roughly one learner in ten gave an answer too thin to learn from: a stub, very little in total, or an answer that did not address the question.
- In peer review, 17 French reviewers and 3 English reviewers wrote generic feedback that could sit under anyone’s work, and thin reviews touched 14 of 104 English and 26 of 222 French reviewers. Colleagues pay for this one.
17 English and 35 French learners received a warning email.
That is about one in eight of the learners we screened.
Where we held back
Several findings did not become warnings.
Quality-only findings did not go out as honor code warnings, which left 62 learners out in peer review and 18 in the submission step, because the subject line would have overstated what we found.
About a third of authors in each language had not yet answered the feedback they received.
That step may still be open, so those became notes and no email went out.
Learners who share a name with another learner were held until we can route their messages safely.
One finding shows why a person reads the output.
Nine English learners shared the phrase “what the tool does”.
Question 3 tells learners to open by saying what the tool does, so the phrase echoes the course’s own instruction.
A rule that only counts repetition would have called it copying.
The system surfaced it, and then recognised the echo.
Each warning email says it is a warning, that nothing has been decided, and that the learner may reply if we have it wrong.
This follows the principle in the integrity layer article: clear expectations before the work, dialogue before sanction, and a human who checks the reply.
What does this mean for the learning scientist?
For the learning scientist: the flags measure participation integrity.
They do not measure depth, accuracy or kindness.
An answer can pass every rule and still miss the point, and the AI signal needs a person to read the excerpts before it becomes a finding.
What does this mean for the malaria specialist?
For the malaria specialist: a health worker in one country writes the exact words she would use with a worried family, and a colleague in another country reads them and learns.
Peer learning works only when that exchange is real.
The screening checks that it is, and keeps each finding visible to the learner and open to reply.