Piauí is using AI to score practice essays. Students deserve the marking rubric

THE SIGNAL IN ONE SENTENCE

A score can feel like an answer even when it is only a prediction. That is the strange power of putting a number beneath a student's essay. The number arrives clean. The writing process was not. A student made choices about argument, evidence, structure, grammar and what kind of future could solve the problem in front of them. The score compresses all of that into something that fits inside a circle. Now add artificial intelligence and the circle arrives faster. Piauí's state education department says its Canal Educação platform uses an AI tool to help correct practice essays for Brazil's National High School Examination, known as Enem. In an October 11 report, the department described the tool as one piece of a wider public-school preparation programme that also includes in-person review sessions, writing laboratories, mentoring, simulations and digital lessons. A student quoted by the department said he uses the tool to correct texts and simulate his likely score. That distinction matters. The tool is not grading the official Enem exam. It is giving practice feedback before the exam, which is scheduled for November 8 and 15. The official essay is evaluated by human reviewers under rules published by Brazil's National Institute for Educational Studies and Research, Inep. The plain signal is that an AI score can be useful practice, but students deserve to see how the machine reached it and how often that reasoning agrees with qualified teachers. The official Enem rubric is not a secret. Inep says each essay is evaluated across five competencies: formal written Portuguese, understanding the prompt and argumentative format, selection and organization of evidence, use of linguistic connections and a proposed intervention that respects human rights. Each competency can receive up to 200 points. Two evaluators score the official essay, and the final result begins with the average of their totals. That structure gives Piauí a ready-made transparency test. An AI practice tool should not simply flash 760 and send a teenager back to the blank page. It should show the predicted score for each competency, identify the passage that influenced the result, explain what the student could revise and state how confident the system is. The feedback should sound like a teacher pointing to the work, not a slot machine announcing a number. The first requirement is a visible rubric. Students should be able to open the same five competencies the tool claims to simulate. Every suggestion should connect to one of them. If the model flags cohesion, it should show where the relationship between ideas became unclear. If it lowers the argument score, it should identify which claim lacks evidence or how the text drifted from the prompt. A generic comment such as improve your conclusion is not useful merely because it appeared instantly. The tool should also distinguish rules from judgment. Spelling and punctuation can often be pointed out with relatively direct evidence. Evaluating whether an argument is coherent, relevant and well developed requires interpretation. A proposed intervention may be specific without being wise. A sentence can be stylistically unusual without being wrong. Students need to know which kind of decision the system made. The second requirement is a public agreement test. The state should compare the AI tool with experienced Portuguese-language teachers on a representative set of essays. Results should be reported by competency, score range, writing style and student group. How often does the predicted total land close to the teachers' scores? Where does it disagree? Does it consistently mark certain styles higher or lower? Does it handle unconventional but valid arguments? What happens with handwriting converted through image recognition? Does performance change for students with dyslexia or other learning differences? An average difference is not enough. A tool could be close on most essays and catastrophically wrong on a smaller group. It could predict total scores reasonably while misunderstanding the reason a human evaluator awarded them. It could reward formulaic writing because the pattern resembles high-scoring examples in its training data. The useful report includes the distribution of errors and examples of serious disagreement, with student identities removed. This is calibration: checking whether the confidence and scores produced by a system match what happens when qualified people review the same work. Calibration should be repeated whenever the model, rubric, prompt format or scanning process changes. The third requirement is a correction path. Practice is supposed to let a student learn from a mistake. That means the tool must be allowed to be wrong too. A student should be able to flag feedback, ask a teacher to review it and see whether the predicted score changes. Teachers should be able to override a suggestion and record why. Repeated disputes should become test cases for the next evaluation cycle. This is not an appeal process with lawyers and tiny clocks. It is a feedback loop with dignity. If the machine tells a student that a strong paragraph is weak, the student should not have to choose between trusting the number and trusting their own voice. A teacher should be able to make the disagreement legible. Piauí's broader Writing Laboratory programme already places technology alongside pedagogical support. Official 2026 programme material says essays are corrected on the Canal Educação platform with AI assistance under Enem criteria and validated by Portuguese-language teachers. That human validation is an important boundary. It should be visible inside the student experience, not buried in programme language. Which feedback came from the model? Which came from a teacher? Was the score reviewed? Did the teacher agree? If not, what changed? The interface should preserve those distinctions. The fourth requirement is evidence of learning. Piauí reports substantial reach across its wider preparation effort. The October 11 announcement says Pré-Enem na Estrada had reached more than 22,000 students, the Esquenta Enem programme held 30 reviews with 1,991 participants and the Writing Laboratory included 43 classes and 1,350 students. Those are programme counts. They are not evidence that the AI tool improved writing. An earlier state report said the platform's AI correction system had evaluated more than 81,000 essays during the prior cycle. That is a large volume of activity. It still does not tell us whether students became better writers, whether predictions matched final Enem results or whether the tool helped some students more than others. Usage is not learning. The state could measure improvement without turning every classroom into a laboratory experiment. It could compare a student's first and later drafts by competency, record which feedback they acted on and ask teachers whether the revisions improved the argument. It could compare results for students using AI feedback alone, teacher feedback alone and both together, while controlling for obvious differences in participation. It could measure whether students become less dependent on formulaic templates over time. It could also ask the students. Did the feedback make sense? Did it change how they revised? Did it make them more confident or simply more anxious about a predicted number? Did they start writing toward what they thought the algorithm wanted? The point of practice is not to make the simulation accurate enough to feel official. It is to help the writer think better before the official exam begins. The fifth requirement is protection for student work. An essay can contain personal experiences, political beliefs, family details, health information and opinions a teenager never expected to become training data. The current October announcement does not name the AI vendor, describe where essays are stored, state how long they are retained or say whether student writing is used to improve another model. Those questions should be answered in plain Portuguese before a student submits a text. The platform should collect only what it needs. It should separate identity from evaluation data where possible, restrict staff and vendor access, publish a retention period and provide a route to delete practice work. Student essays should not become a reusable commercial dataset by default. Schools should also avoid making an opaque tool the only route to feedback. Students with unreliable internet, limited device access or accessibility needs must receive an equivalent path. Inep publishes specialized guidance for participants with dyslexia, autism, deafness or hearing impairment. A practice system should be tested with the same seriousness instead of assuming one interface and one language-processing pattern fit everyone. The sixth requirement is a boundary around the score. The predicted number should never determine admission, placement, discipline, access to mentoring or which student a teacher decides is promising. It is a practice signal. Piauí's own programme is strongest when the AI tool sits among teachers, writing labs, mentoring and revision. The danger begins when convenience turns the prediction into an invisible gatekeeper. Schools should publish a simple rule: no consequential decision is made from an automated essay score alone. They should also show students the gap between prediction and official evaluation. If the tool says 880 and the human reviewers later produce a different result, that difference is not an embarrassment to hide. It is evidence that can improve the tool and calibrate student expectations. There are practical benefits here. Fast feedback can let a student revise more often. Teachers can use aggregated error patterns to plan lessons. A writing programme can identify where students struggle across the five competencies. Rural students can receive another layer of support between in-person sessions. A clear rubric can turn a mysterious score into a checklist for the next draft. None of that requires pretending the model is an examiner. In fact, the system becomes more useful when it admits what it is: a practice partner that predicts, explains, sometimes disagrees and always leaves the final educational judgment with people. The state has not published the vendor, model architecture, validation sample, agreement rates, subgroup results, privacy terms or error analysis for the Canal Educação tool. It has not shown that predicted scores improve final Enem performance. Those are real gaps. They are also fixable. Piauí could publish a one-page model card for students, a competency-by-competency validation report, examples of corrected disagreements, data-retention rules and a yearly learning-outcome review. It could let teachers and students help test each new version before it reaches the entire platform. The goal is not to slow feedback until it becomes useless. It is to keep speed from impersonating certainty. Students already have a public rubric for the exam that matters. The tool practicing beside them should be at least as willing to show its work.