Europe Delays Education AI Grading Rules to 2027

Europe Just Gave Schools 16 More Months Before AI Grading Rules Apply

On 27 July, Regulation (EU) 2026/1744 entered into force and moved a date that thousands of European schools and edtech vendors had circled. The AI Act's obligations for stand-alone high-risk systems, the category covering software that grades student work, screens applicants, and watches them sit exams, were due to apply on 2 August 2026. They now apply on 2 December 2027.

Sixteen months of runway is the news. What makes the runway useful is that the same week produced unusually direct evidence on what schools should build with it.

What moved, and what stayed put

The regulation, known as the Digital Omnibus on AI, was published in the Official Journal on 24 July and took effect three days later. Until that publication the original schedule was still operative law. A delay that had been agreed politically became one a compliance officer can plan around.

The new date applies to stand-alone systems listed in Annex III. High-risk systems embedded in products already regulated under Annex I get longer still, until 2 August 2028.

One set of duties did not move at all. The Article 50 transparency obligations still apply from 2 August 2026: telling people when they are interacting with an AI system, marking synthetic audio, images, video and text in machine-readable form, disclosing deepfakes, and notifying people subject to emotion recognition or biometric categorisation. Systems already on the market before that date have until 2 December 2026 to meet the content-marking duty.

That split matters more than it sounds. An institution that read the Omnibus headline and shelved its entire AI Act programme still has a live obligation in a matter of days if it runs a student-facing chatbot or generates synthetic teaching material. The high-risk regime was postponed; transparency was left exactly where it was.

The four classroom systems Europe calls high-risk

"Education" is a uselessly broad label, and most coverage of the delay stops there. The Annex III text is narrower and far more useful. Point 3 names four things: systems that determine access, admission, or assignment to educational and vocational training institutions at any level; systems that evaluate learning outcomes, including where those outcomes steer what a learner does next; systems that assess the appropriate level of education a person will receive or can access; and systems that monitor and detect prohibited behaviour during tests.

The second one is the wide net. An automated marker sits inside it without argument. So does an adaptive-learning product whose scoring changes which module a student sees tomorrow, which sweeps in a great deal of software institutions currently file under "courseware" rather than "assessment." Classification follows what the system decides, not what the vendor calls it on the invoice.

Once a system lands in Annex III, five families of obligation attach: risk management, data quality, human oversight, logging, and technical documentation. Three of those are paperwork. Two of them, human oversight and logging, are architecture, and you cannot staple them onto a system that was never built to pause or to keep a record.

For vendors selling education AI tools into Europe, or to any institution with European students, that classification travels with the product rather than with the buyer. A school running a marking pilot alongside a proctoring tool is operating two Annex III systems in December 2027, whatever the internal project page calls them.

The evidence that arrived the same week

Much of the public argument about AI in schools still runs on a single question, will AI replace teachers, while the systems actually landing in classrooms raise a narrower one: what happens when software gets write access to a gradebook. On that narrower question, some of the most useful evidence yet on how AI is affecting education arrived within five days of the regulation.

James Zumel Dumlao and colleagues tested what they call the GenAI substitution hypothesis. If students are offloading cognitive work to chatbots, grades should climb faster in courses built on take-home problem sets and essays than in courses assessed by in-class exams. The team pulled syllabus and administrative data from a large US university covering 2016 to 2025, 138,386 students in total, classified each course's exposure with a human-validated extraction pipeline, and ran a differences-in-differences comparison across ChatGPT's release while modelling pandemic effects both as persistent and as transient.

They found no significant differential effect on grades, overall or among students who had previously been performing worse. Self-reported understanding showed nothing either. That sample size is what gives the null result its weight: an institution writing AI policy this quarter is probably solving for a pattern its own gradebook would not show.

Then there is the chart from Brown that circulated in July. In Roberto Serrano's welfare economics course, the take-home midterm averaged 96 percent with nearly the entire class bunched at the ceiling. The in-person final averaged 48.6 percent, the lowest in the course's history. The course had drawn far more students than it usually does, and a substantial share of them dropped, skipped the final, or failed it.

Both findings are true and they do not contradict each other. The Brown figures describe one course's assessment design collapsing under load. The university-wide study describes what happens across an entire institution over a decade. Aggregate panic is not showing up in aggregate grades, while individual assessment designs are exposed. Leo Schumann, writing about the Brown case in Inside Higher Ed, reduced the diagnosis to one line: "We made grades the point. AI just made them cheap."

Kai Yao's audit of published AI assessment guidance at 30 universities finds the same gap from the policy side. Four open-weight models scored each institution's guidance against a pre-specified rubric, with results averaged to damp down any single model's bias. Policies turn out to be good at drawing boundaries, what a student may and may not delegate, and weak on what sits underneath: what a learner must still demonstrate, what evidence the institution will preserve, and how either of those connects to the credential the degree is meant to certify. Permission categories are quick to write; evidence standards take institutional judgement.

Sixteen months, and what to build in them

One study this week measured an intervention rather than an attitude. Matey Yordanov, Mikhail Bychkov and Andrei Kuchma ran a 16-week term with A-level students in mathematics, further mathematics, biology, chemistry and physics. One group's scripts were marked by humans, the other by an automated handwritten-assessment platform. The automated group finished ahead on final mock marks by a margin the authors report as statistically significant.

The margin is less instructive than the mechanism. Marking turnaround fell from 11.2 days to under a tenth of a day. Eleven days means a student corrects a misconception a fortnight after forming it, or never corrects it at all. Under a tenth of a day means the correction lands while the reasoning is still in working memory, and it is the only reason the same teaching staff could set several times as much practice over the term. The software did not teach better; it deleted a queue.

Hold that result to its actual size: 142 students at a single institution, one term, quasi-experimental rather than randomised, and the outcome measured was mock exam marks, not the final examinations themselves. It is a strong signal about latency in formative assessment, not a general law about automated marking.

Schumann reaches a compatible prescription from the opposite direction: tell students plainly that the struggle is the mechanism, lower the stakes by assessing more often in smaller pieces, and ask for visible thinking, drafts, revisions and conversations, in place of polished final artefacts. That advice has always been sound and always been unaffordable, because frequent assessment means frequent marking. The A-level result is what puts a price on it.

Which surfaces the part institutions tend to meet late. An automated marker is software with write access to a student record. That is precisely the shape the December 2027 human-oversight and logging duties are aimed at, and it is also the shape of an ordinary Tuesday problem: a student challenges a mark, and somebody has to show what the system actually did.

Both are workflow design questions before they are compliance questions. On AgentPMT, an educational agent that marks scripts can be wired with a human-in-the-loop checkpoint at any step of its workflow, so the agent pauses before a grade is committed while a teacher approves or rejects it from the mobile app with biometric auth, then the run continues. The audit feed records every action that agent took in real time, filterable down to the full request and response behind each step, which is what an appeals process needs and what most deployments still cannot produce.

An institution that designs in those two properties now collects the appeals benefit sixteen months before the legal one. Threading an approval gate through a live marking pipeline in late 2027 is a different and much worse project.

The cost question the study leaves open

The A-level paper reports its outcome and describes the intervention as low-cost. It does not publish a per-script figure, which is a gap in the evidence, and not one anyone else can fill on its behalf. Several times as much practice means several times as much inference, and an institution that cannot say what a marked script costs cannot decide whether to scale the pilot that worked. That is how promising pilots stall.

Right-sizing is a per-task question, not a per-institution one. Marking a multi-step mathematics proof and marking a one-line biology definition are different jobs and do not need the same model behind them. AgentPMT's budget system sets hard spend caps per agent connection, with tool, workflow, and credential restrictions attached, so a marking agent cannot run up a term's bill unnoticed. The test bench itemises what each run actually spent, down to the token and the tool call, then lets you swap the model, re-run the same scripts, and compare cost against output quality before anything scales. Because the model gateway is swappable and the platform stays agent-agnostic, changing the model behind a marking workflow does not mean rebuilding the workflow, which matters when the system carries documentation duties and models turn over faster than documentation does.

Cost is staff time as well as tokens. The case for automation in schools is usually argued on teacher hours, yet a large share of educators report no AI training at all, and the institutions expanding fastest are drawn from that same population. Buying the marking system costs less than teaching people to run it.

Scott Latham of the University of Massachusetts at Lowell, who built an index scoring institutions on AI readiness across classroom use, campus life, operations, governance, research and workforce preparation, is blunt about where the argument has settled: "There is no endgame where five years from now there are faculty across the academy that say, 'Wow, it's good we didn't buy into AI. It's gone.'" He is equally blunt about the other failure mode, warning institutions not to go hog wild and push AI into every corner of campus.

What the delay actually decides

December 2027 will ask for the same things it was always going to ask for: oversight a human can exercise, logs that survive a challenge, documentation that matches the system in production. The delay changes when that has to be working, and it hands over sixteen months in which the evidence for what to build got considerably better than it was in June.

For institutions using AI in education today, the practical question is narrower than the policy debate suggests. Adoption already happened, and a student appeal will land long before the deadline does. What decides that appeal is whether the system that produced the mark can pause for a human and account for itself afterwards. Those are the properties the regulation will require in sixteen months, and the ones that make a mark defensible next term.

Sources

• The Digital Omnibus on AI enters into force today, Lewis Silkin
• EU Digital Omnibus on AI Enters Into Force, Hunton Andrews Kurth
• Annex III: High-Risk AI Systems Referred to in Article 6(2), EU Artificial Intelligence Act
• Generative AI Availability, Grades, and Student Satisfaction at a Large University, arXiv
• That Chart From Brown Isn't Really About Cheating, Inside Higher Ed
• What Does the Credential Still Certify? Cognitive Stewardship for AI-Mediated Education, arXiv
• The Effect of High-Frequency, Automatically-marked Formative Assessments on Student Outcomes in A-Level Sciences, arXiv
• Why One Professor Abandoned the AI Resistance, Inside Higher Ed
• AI embraced by more students and educators, Instructure finds, K-12 Dive