EU AI Act transparency duties apply now; high-risk duties from December 2027. Check your exposure
Insights About us Careers
Contact us
Public Sector & Education

AI for Higher Education

Universities have the sector's most international student body and are being sold the tool that fails hardest on exactly those students.

Tier 1
Our depth here
Annex III
Admissions and evaluation
61.22%
Detector false positives

Universities carry the widest range of AI questions of any institution type — admissions, assessment, student support, research administration, and their own scholars studying the technology — and the most exposed one is academic integrity.

In one paragraph

AI for higher education covers student support and enquiry handling, retention and engagement analytics, research administration and grant management, admissions process support under high-risk obligations, accessibility and translation, and the academic integrity position an institution adopts.

Academic integrity, and the evidence universities need to see

This is the question institutions ask first, and the answer is unusually well-evidenced. It matters most in higher education because international students are a large share of most cohorts.

  • The published false positive rate is 61.22%. A study in Patterns tested seven widely-used detectors against TOEFL essays by non-native English writers and found that average rate.
  • All seven unanimously flagged around a fifth of them. Roughly 19.8% of the essays were called AI-generated by every detector tested, so consulting a second tool does not rescue the first.
  • Accuracy on US student essays was near-perfect. The failure is not noise; it tracks whether English was learned at home.
  • Misconduct proceedings carry visa consequences for international students. Which makes a systematically biased tool an unusually serious thing to deploy.
  • No detector publishes a confidence interval you can act on. A percentage score presented without an error rate invites staff to treat it as evidence, which it is not.
Worth knowing

The institutional decision is a policy, not a procurement

The useful output here is not a tool but a position: that a detector score is never sufficient evidence for a finding, that staff are told where the error concentrates, that no student is asked to prove authorship without other evidence, and that assessment is redesigned where integrity genuinely matters. Institutions that adopted this early have had far fewer contested cases than those that bought a detector and wrote the policy afterwards, and it costs nothing but the decision.

Student support and retention, where the value is real

Engagement data predicts disengagement well

Attendance, virtual learning environment access, submission patterns and library use together identify students disengaging weeks before they withdraw, which is when support can still work.

Design it as an offer, never as a flag on a record

The output should route a person to a conversation with a student, not create a label that follows them through their studies. Students who know they are scored behave differently, including by disengaging further.

Enquiry handling addresses a genuine service gap

University processes are hard to navigate — deadlines, regulations, funding, visas, appeals — and enquiry volume reflects that. The strict constraint is that anything with a regulatory or immigration consequence reaches a person.

Accessibility work is under-resourced and highly effective

Captioning, transcription, alternative formats and translation improve access for disabled students and international students simultaneously, and the cost has fallen sharply. See NLP services.

Admissions, research administration and the regulated edges

ApplicationRegulatory weightNote
Student enquiry handlingLowProcess navigation; escalate anything with a regulatory consequence
Engagement and retention analyticsModerateFrame as support offers; never a label on a record
Research administrationLowGrant document processing, compliance checking, reporting
Research output and collaboration analysisLowInternal analytics; handle individual metrics carefully
Accessibility and captioningLowDirectly inclusive and consistently underfunded
Admissions screening or rankingHighAnnex III; determining admission is explicitly high-risk
Automated evaluation of learning outcomesHighAnnex III; assessment that goes on a record
Remote proctoring behaviour monitoringHigh to prohibitedAnnex III where it monitors prohibited behaviour; emotion inference is prohibited outright
Worth knowing

Research administration is the underrated win

Universities spend a striking amount of academic time on grant applications, compliance documentation, ethics submissions and funder reporting — work that is high-volume, structured, and entirely unrelated to research quality. Document processing, obligation extraction from grant conditions and reporting assembly return academic time directly, with no student decision involved and no regulatory weight. It is rarely the project an institution arrives asking for, and it is frequently the one with the best return.

Process

How an engagement runs

The integrity position settled as policy, then the support and administration work.

Weeks 1 to 3

Scope and classification

Whether anything touches admission, evaluation or proctoring, and what Annex III and Article 27 require.

Weeks 4 to 7

Data assessment

Engagement data availability, research administration volumes, and student data governance.

Weeks 8 to 14

Build

Student support and enquiry handling, or research administration and accessibility.

Weeks 15 to 18

Trial

With support staff and academics on live work, measured on outcomes rather than adoption.

Ongoing

Operation

Retraining across cohorts, with outcome testing where anything touches admission or assessment.

Deliverables

What you receive

Students supported earlier, academic time returned, and no tool that misjudges international students.

01

Student engagement analytics

Disengagement identified weeks early, routed to a conversation rather than a label.

02

Student enquiry handling

Process navigation, with regulatory and immigration questions escalated to a person.

03

Research administration

Grant document processing, obligation extraction and funder reporting assembly.

04

Accessibility and captioning

Transcription, captioning and alternative formats across teaching material.

05

Translation

For international students and their families, across the material that matters.

06

Academic integrity position

A defensible institutional policy, with the detector error profile documented for staff.

Fit check

Is this the right starting point?

Worth being direct. There are situations in higher education where custom AI work is the wrong spend, and those are listed rather than buried.

Worth doing if

  • Student withdrawal is discovered too late for support to work.
  • Enquiry volume about process, deadlines and regulations is consuming support staff.
  • Academic time is going to grant administration and funder reporting.
  • Captioning and accessibility provision is rationed by cost.
  • You need an institutional position on AI integrity rather than a detection product.

Do something else if

  • You want detection tooling deployed into academic misconduct proceedings.
  • You want proctoring with emotion or affect inference.
  • You want admissions ranking without the Annex III obligations and outcome testing.
  • Engagement analytics are intended to produce a label on a student record.
Questions

Frequently asked questions

Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.

Should we buy an AI detection tool?

We would advise against it, and the reason is evidence rather than principle. A peer-reviewed study in Patterns found seven widely-used detectors produced an average false positive rate of 61.22% on TOEFL essays by non-native English writers, with all seven unanimously flagging around a fifth of them, while performing near-perfectly on US student essays. In a university with a substantial international cohort — which is most universities — that means a tool that systematically accuses those students, in proceedings that can carry visa consequences. The institutional decision worth making is a policy, not a purchase.

Then how do we handle academic integrity?

By adopting a position and redesigning assessment where integrity genuinely matters. The position: a detector score is never sufficient evidence for a finding, staff are told where the error concentrates, and no student is asked to prove authorship without other evidence. The redesign: process-visible work, oral components, and tasks anchored to material specific to your teaching. Institutions that did this early have had markedly fewer contested cases than those that bought a detector and wrote the policy afterwards.

Does the AI Act reach our admissions process?

Yes, directly. Annex III makes AI systems determining admission to educational institutions, evaluating learning outcomes, assigning people to institutions and monitoring prohibited behaviour during tests high-risk, with obligations from 2 December 2027. Public institutions additionally carry the Article 27 fundamental rights impact assessment before first deployment. In practice this means that if you intend to use AI in admissions, the harms analysis and the outcome testing shape what you build rather than documenting it afterwards, and that is a different project plan from the one usually proposed.

What is the best-value project we are not considering?

Research administration, almost always. Academics spend a striking amount of time on grant applications, ethics submissions, compliance documentation and funder reporting — high-volume, structured work with no bearing on research quality and no student decision involved. Document processing, extracting obligations from grant conditions and assembling funder reports returns academic time directly, carries no regulatory weight, and is rarely what an institution arrives asking for.

Can we predict which students will drop out?

You can identify disengagement well, and the framing matters more than the accuracy. Attendance, virtual learning environment access, submission patterns and library use together surface students disengaging weeks before withdrawal, which is when support still works. The design requirement is that the output routes a person to a conversation rather than creating a label on a record — students who know they are scored behave differently, sometimes by disengaging further, and a flag that follows someone through their studies is a harm the prediction does not justify.

Tell us what the problem looks like.

Thirty minutes, no charge, no deck. We will tell you whether this is an AI problem, a data problem, or a process problem — and we will say when the honest answer is to buy something rather than build it.