AI grading for exams and essays, reviewed by the teacher

Here is an assisted grading from start to finish, on a fictional open question: the rubric locked before the exam, a participant's answer, the Evalmee AI suggestion criterion by criterion, then the grader's decision. That decision, not the suggestion, is what enters the grade.

The question, out of 6 points: "Explain the difference between formative and summative assessment, with one example of each." Before the exam, the grader attached three criteria to the question, 2 points each, and locked the rubric: the definition of formative assessmentFormative assessmentAn assessment carried out during learning, whose function is to inform learner and teacher about what is currently being acquired. Its useful output is the feedback, not the mark: it is frequent, low-stakes, and serves to correct course while there is still time to do so. The term comes from Michael Scriven's work in 1967, taken up by Benjamin Bloom, which sets the assessment that forms against the assessment that takes stock. In practice a formative assessment is recognised by three traits: fast, precise feedback on every answer, a result that carries little or no weight in the final mark, and a teaching decision taken from the errors observed. Example: a ten-question quiz at the end of every session, marked automatically, with the reason for the correct answer shown at once; the teacher reads the per-question statistics and returns, in the next class, to the notion that half the group missed. The testing effect strengthens the case: testing yourself consolidates memory more than re-reading does. The same question can serve in a formative or a summative setting; what decides is the use made of the result, not the format.See the full definition, the definition of summative assessmentSummative assessmentAn assessment that comes at the end of a sequence, a module or a year to take stock of what has been learned and produce a mark that counts. It fixes a result rather than accompanying progress, and differs from formative assessment in its function, not in the format of the questions asked. The word comes from sum: the mark summarises what has been learned, and it serves to decide, to pass a module, to rank, to admit, to award a title. When that decision conditions a degree or a certification, the more precise term is certification assessment. Example: the final exam of a contract law course, two hours, one case study and ten short questions, marked out of 100 and weighing sixty per cent of the module mark, the rest coming from continuous assessment. Because its mark has consequences, a summative assessment demands more than a formative one: a marking scheme published in advance, identical sitting conditions for everyone, identification of the candidate, a traceable marking process and a route for appeals. An item analysis after the exam remains useful: a question that the strongest candidates miss more often than the others points to an ambiguous wording or a wrong answer key.See the full definition, and the relevance of the examples.

The participant's answer: "Formative assessment happens during learning. It is used to spot what the learner has not yet understood, so it can be revisited. Summative assessment is the one that counts towards the average, like the end-of-semester exam."

After a click on Generate feedback, Evalmee AI suggests 4 points out of 6 and this feedback: "The two definitions are told apart by their timing and by the use made of the mark. An example of formative assessment is missing; an ungraded end-of-class quiz would be one."

The grader reads it over. They keep the feedback and the point withheld on the examples, but restore the 2 points for the summative definition: "the one that counts towards the average" is the wording used in class. They click Approve. The answer is worth 5 points out of 6.

You define the criteria and the scoring before the exam. After the exam, you lock the rubric of the question, the AI generates suggested points and feedback, you read them over and you click Approve. The grade only moves at that last step.

The order matters: the rubricRubricA grid crossing assessment criteria with performance levels described in full sentences, from the weakest to the strongest. The marker places the paper in a cell rather than awarding a figure: the mark becomes the consequence of a description shared in advance.See the full definition is locked before the AI steps in, so the scoringMarking schemeThe rule that converts a candidate's answers into a mark. It sets the number of points per question, any penalties, and the way partial credit is awarded. Published before the exam, it constitutes the examiner's commitment as to what will be rewarded.See the full definition is settled before any submission has been seen and applies to every answer.

The assistant is called Evalmee AI. It relies on OpenAI, processing takes place in the European Union, and the data sent is anonymized. Client personal data is never shared with an AI provider, nor used to train models.

The AI suggests points and writes feedback on written answers. It does not grade closed questions, defines no criteria, validates no submission and does not detect plagiarism. Each of these four limits matches a design choice, not a temporary gap.

A product page that promises fully automatic grading describes an exam few institutions would agree to defend before a board. So here is the exact line. The comparison of online exam platforms puts the same question to the six other French vendors, with the public page each answer was read on.

Before the exam, the scoring is set and the rubric locked. After the exam, it can still be changed from the Paper tab, and every grade is recalculated. In both cases, one exam is never graded under two scoring schemes.

A locked rubric cannot be changed without consequence: if you unlock it while answers have already been graded, that grading is reset. The paper follows the same rule. Once the exam is open, the questions seen by participants no longer change, so that everyone sat the same paper.

A scoring mistake can be fixed afterwards. From the Paper tab, the Edit button lets you revisit the correct choices, the weightings, the grading comment and the tags, then run Recalculate grades. A right answer ticked wrongly is fixed that way in two minutes, without reworking submissions one by one. A question to discard comes down to a weighting reset to zero, followed by the same recalculation.

On choice questions, the scoring is explicit and configurable, even after participants have sat the exam. The discordance mode, used in health-studies entrance exams, grades both the choices ticked and the choices left blank. To see what a penalty does to a paper before settling on one, the test grade calculator applies the same arithmetic question by question. To report the grade on a scale other than the point total, the score converter does the arithmetic and prints the full conversion chart.