Point 3(b) of Annex III classifies as high-risk AI systems intended to be used to evaluate learning outcomes in educational and vocationaltraining institutions at all levels. The evaluation of learning outcomes refers to the process of appraising knowledge, know-how, information, values, skills and competences-acquired in formal and non-formal, educational and vocational training settings-against relevant standards (learning outcomes, validation).
The outcome of this evaluation process typically involves a quantified or qualitative assessment of the learner’s achievement against predefined learning objectives, standards, curriculum goals, or competency frameworks. The resulting evaluation may be used for a range of purposes, such as issuing grades or certificates, awarding a qualification identifying gaps in mastery, informing accreditation or qualification decisions, and automatically steering or personalising subsequent instructional pathways (e.g., recommending remedial modules, adaptive learning resources, or progression to more advanced content).
A distinction should be made between ‘summative’ evaluation and ‘formative’ evaluation, in view of their different objective and impact on a person’s learning prospectus. Summative evaluation refers to the assessment of student learning, which leads to a grade, evaluation, or qualification, or at the end of an formal instructional unit (a course, programme, or educational pathway, such as a term, semester, or a school year), and can thus have a significant impact on the learner’s educational trajectory, including by steering his or her learning process. Formative evaluation, by contrast, is an ongoing process of assessing student learning and progress, and has the objective of improving performance, rather than leading to a final grade, evaluation, or qualification. Formative evaluation is, in most cases , conducted continuously throughout the implementation phase of the academic project or programme and does not lead to a summative evaluation.Its primary purpose is to support learning through activities such as exercises, providing feedback and guidance to help students progress.
Considering the impact on the student ’ s educational trajectory, only AI systems that make an evaluation towards a final decision affecting educational and professional pathways based on a summative assessment fall under point 3 (b) of Annex III. This means that AI systems that have a significant influence on a student’s educational path, such as AI-powered grading systems that determine a student’s final grade or intermediate grades that are considered in the final evaluation or academic standing, should be classified as high-risk. This follows from the use of the terms ‘evaluate learning outcomes’ used in that provision.
By contrast, AI systems that provide feedback and learning support as part of the ongoing educational process should not be classified as high-risk. This category includes adaptive learning systems that adjust the difficulty level of course materials based on a student’s performance, intelligent tutoring systems that provide personalised feedback and guidance to students, and learning analytics platforms that track student engagement and progress and provide insights to teachers and educators . Such systems generally do not make final decisions affecting the educational pathways of a person, but rather support the learning process and provide feedback to students in a continuous manner.
In addition, AI systems used by students on their own volition to support non-formal and informal learning, without specifically being required to do so by the educational or vocational training institution at which they are enrolled , should not be classified as high-risk. Such systems should only be classified as high-risk if they lead to the attainment of a credential, certification, or a form of validation recognised by the relevant public authorities, since they would then have an impact on student’s educational and professional trajectory, unless one of the exceptions in Article 6(3) AI Act applies.
a) Practical examples of AI systems falling within the high-risk use case of point 3(b)
- AI-enabled grading and feedback systems intended to be used to evaluate students ’ assignments, such as tests, quizzes, and exams which count towards a final evaluation. Such AI systems have an impact on the summative evaluation of students by analysing the results of assignments and proposing grades.
- AI-enabled personalised learning assistant intended to be used to provide recommendations and feedback to students on their progress by generating reports at the end of each unit, which include grades, areas of strength and weakness, and recommendations for improvement . Although such AI systems support formative evaluation by providing ongoing feedback and guidance to students to help them improve their understanding of course material, their output is also used by the teacher to inform decisions about student ’s grades at the end of the academic period and therefore have an impact on summative evaluation.
b) Practical examples of AI systems falling outside the high-risk use case of point 3(b)
- AI-enabled language learning software applications intended to be used by students to support non-formal and informal learning by providing instant feedback and correction s on identified errors, but which does not lead to the attainment of a credential, certification, or form of accredited validation. Such AI systems are used by students on their own volition, without being required to do so by an educational institution.
- AI-enabled neurodiverse learning companions intended to be used to support students with neurodivergence, such as autism, dyslexia, and ADHD, which uses machine learning algorithms to create personalised learning pathways that cater to the individual needs and learning styles of each student by providing real-time feedback and adjustments to the learning materials, pace, and format to help such students stay engaged and motivated. For example, such an AI system may provide a student with dyslexia text-to-speech functionality, font size and colour adjustments, and multisensory learning materials to help that student better understand and retain information. Such a system may also provide accommodations such as extra time to complete assignments, breaks, and stress-reducing exercises to help students with anxiety or sensory processing issues. Such a system’s output is used to inform the teacher’s instruction and provide additional support to students with neurodivergence, but it is notintended to beused to determine grades or to assess student learning.
- AI-enabled pronunciation feedback tool intended to be used to provide feedback to students on their language pronunciation or fluency by analysing the student’s speech and providing suggestions for improvement, such as correcting intonation, rhythm, and accent. The feedback provided by the system is used by the students to improve their pronunciation skills; it is not used by educators to determine grades counting towards a final evaluation or assess language proficiency. Same example can be applied for an AI tool on reading skills providing feedback intended to be used by learners or students only.
c) Practical examples of AI systems falling within the use case but exempted by the filter mechanism in Article 6(3)
- AI-enabled exam quality checker intended to be used to check an exam prepared by a teacher for errors, such as grammatical issues, ambiguous wording, or inconsistencies with the rubric, and to suggests revisions for the teacher’s consideration. However,since the AI system does not alter the core content of the exam, nor does it change the level of difficulty or the educational intent behind the questions , and t he teacher remains responsible for any final decisions on the exam’s content, the AI system’s role should be considered strictly to enhance the clarity and accuracy of the exam, ensuring that it is error-free and aligns with the intended learning outcomes. Such a system would therefore fall under the exception for systems intended to improve the result of a previously completed human activity listed in in Article 6(3)(b) AI Act, namely, the preparation of the exam , by refining its quality without altering the substance or impact of the teacher’s original educational decisions.
- AI-enabled grade calculator intended to be used by an educational institution to process data related to assessment and test grades obtained by students during the academic period which determines the final grade of the students’ learning outcomesfor the year. Such an AI system is designed to perform a simple calculation of theaverage grade, taking into account the grades and weights of each assessment, and test the output of the system as a numerical value representing the average grade, which is then used by the teacher to inform their final evaluation of the student’s performance. Such a system would therefore fall under the exception for systems intended to perform a narrow procedural task listed in Article 6(3)(a) AI Act.
- AI-enabled assessment review tool intended to be used to detect decision-making patterns or deviations from prior decision-making patterns, with the goal of identifying potential inconsistencies or anomalies in the assessment decisions made by instructors. Such an AI system analyses the grading patterns of instructors over time, and flags any deviations from the expected pattern. For example, if an instructor has consistently given high grades to students who complete a certain assignment, but suddenly gives low grades to a group o f studen ts who complete the same assignment in the same manner, the AI system will flag this deviation as a potential anomaly. The output of the AI system is then reviewed by a human assessor, who investigates the flagged deviations and determines whether they arejustified or not. The human assessor may decide to adjust the grades or take other corrective action, but the AI system’s output is not used to automatically change the assessment decisions. Such a system would therefore fall under the exception for syste ms intended to detect decision-making patterns or deviations from prior decision-making patterns listed in Article 6(3)(c) AI Act , since it is not meant to replace or influence a previously completed human assessment.