AI Essay Grading as High-Risk Under the EU AI Act

How the classification logic works

The EU AI Act (Regulation (EU) 2024/1689, dated 13 June 2024 and published in the Official Journal on 12 July 2024) classifies stand-alone high-risk systems by intended purpose. A system is high-risk if, given its intended purpose, it poses a high risk of harm to health, safety or fundamental rights, taking into account severity and probability of harm, and it is used in one of the pre-defined areas the Regulation sets out. Education is one of those areas, under the Annex III education category.

The education recital says AI systems used in education or vocational training should be classified as high-risk. It lists determining access or admission, assigning persons to institutions or programmes, evaluating learning outcomes of persons, assessing the appropriate level of education for an individual, and monitoring and detecting prohibited behaviour of students during tests. The stated rationale is that such systems may determine a person's educational and professional course of life and so affect their ability to secure a livelihood.

For essay grading, the working questions are therefore practical. Is the system stand-alone? Does its documented purpose fall in the education area? Does the output evaluate a learner's outcome, or materially influence the level of education a person receives or can access? Because classification follows intended purpose, the answer depends on how the product is specified and used, not only on the fact that it reads essays. As of November 2026, no amendment, delay or repeal of the Regulation has been identified.

Where common essay-related uses are likely to fall

The table gives a starting position for internal triage, not a conclusion. Each use needs its own recorded reasoning against the recital categories.

• Summative grading that determines a mark or result: Most directly matches evaluating learning outcomes. Treat as high-risk unless counsel concludes otherwise on documented grounds.
• Scoring that feeds placement, progression or access decisions: Also engages the recital language on assessing the appropriate level of education and materially influencing what a person can access.
• Formative feedback with no effect on results: Not named in the recital. Position depends on intended purpose, and the line to grading is an interpretation question.
• Integrity or test-monitoring checks: Monitoring and detecting prohibited behaviour during tests is named. Whether other plagiarism checks fall in scope is a question for counsel.

Intended purpose, Article 6(3) and the formative line

The difference between grading that influences outcomes and feedback that does not is the practical core of the decision. The recital lists evaluating learning outcomes separately from materially influencing the level of education a person will receive or can access. Whether a given scoring feature falls under the first, the second or neither is an interpretation question, and a score shown to a student as a suggestion may still become a result once a teacher accepts it without review.

Because classification follows intended purpose, the instructions for use and product configuration should match the documented purpose. A tool configured in summative mode while documented as formative creates a gap between the classification on paper and the behavior in production. Keep the mode, the stated purpose and the deployment settings aligned.

If your organisation intends to rely on an Article 6(3) exception, record which exception, the facts supporting it and who approved it. A claimed exception is a position to be defended to a regulator, so it needs a written basis that counsel has reviewed. The same applies to any conclusion that a use is lower-risk.

What high-risk status means for providers and deployers

Providers of high-risk systems must ensure the systems meet the requirements of the relevant Section of the Regulation. Two of those requirements bear directly on runtime behavior.

Human oversight is a design requirement. High-risk systems must be designed and developed, including with appropriate human-machine interface tools, so that natural persons can effectively oversee them while in use. The aim is to prevent or minimise risks to health, safety or fundamental rights, including under reasonably foreseeable misuse. For essay grading, this suggests an interface where a reviewer can see a score and its basis, and can override or escalate it before it affects an outcome.

Logging is the second. Providers must keep automatically generated logs referred to in Article 12(1) to the extent those logs are under their control. In a hosted or API deployment, that qualifier matters. Decide in advance which party controls model inputs and outputs, grader decisions and overrides, and write down the split.

Deployers are expected to take appropriate technical and organisational measures to use high-risk systems in accordance with the instructions for use, with further obligations on monitoring how the system functions and on record-keeping, as appropriate. An institution that deploys a vendor's grader therefore needs its own evidence of use within instructions, not only the vendor's documentation.

• Oversight design: Evidence that reviewers can see, override or escalate a score before it takes effect.
• Automatically generated logs: Records of interactions and decisions, with documented custody between provider and deployer.
• Use per instructions: Configuration records showing the system runs in its documented mode and purpose.
• Monitoring and record-keeping: Ongoing deployer records of how the system functions in use.

Where runtime enforcement and evidence fit

Classification is a documentation exercise, but oversight and logging are only credible if they hold at runtime. A regulator or auditor is likely to ask what the system was allowed to do and what it actually did, not only what the policy document says.

Trussed enforces policy in the execution path of AI interactions and tool calls. Policy can allow, redact, block or escalate an interaction for human approval, enforced at runtime, not reviewed after the fact. In an essay-grading context, that means a rule can route a score or action to a named human approver before it proceeds, and the decision is recorded. The enforcement produces continuous audit-ready evidence, which Trussed maps to EU AI Act requirements among other frameworks. Deployment is self-managed in the customer cloud, with configurable retention and data residency, which bears on the log custody question above.

This does not amount to compliance on its own. Classification, intended-purpose documentation, provider and deployer role allocation, and legal review remain the organisation's responsibility. Enforcement and evidence address the part of the obligations that depends on observable runtime behavior.