ChatGPT in the Classroom: How Universities Have Responded

TL;DR

Universities globally have adopted four broad approaches to AI in academic work: permitted with disclosure (most common), prohibited entirely (specific courses), permitted with attribution, permitted without restriction (paired with assessment redesign). Scarfe 2024 detection finding (94% AI miss rate) shapes the policy debate. Assessment redesign increasingly seen as more important than detection.

The four approaches

1. Permitted with disclosure

The most common university approach globally. Key features:

• AI use permitted for limited purposes (brainstorming, grammar checking, idea generation)
• Students must disclose specific uses
• Original drafting and analytical content must be the student's own
• Sanctions apply for undisclosed use even when the use itself would have been permitted

Adopted by: most North American universities, most UK universities, most European universities. The default position in higher education.

2. Prohibited entirely

Adopted in specific course contexts where the assessment specifically targets capabilities AI can perform:

• Foundation-skills writing courses
• Mathematical reasoning courses with show-your-work assessment
• Language acquisition courses
• Some examination contexts

Sometimes applied institutionally as a baseline ("AI not permitted unless instructor explicitly authorises") with instructor discretion to relax.

3. Permitted with attribution

A minority approach. Allows substantial AI use provided the student properly attributes the AI contribution:

• "This paragraph drafted with ChatGPT, then revised"
• "Code generated with Claude, then tested and debugged"
• "Analysis structure proposed by AI, content original"

Less common; raises questions about what is being assessed. Tends to appear in advanced coursework where AI tools are part of the discipline's professional practice (some computer science, some research methods courses).

4. Permitted without restriction

Rare. Some institutions and instructors permit AI use without restriction, paired with assessment design that demonstrates understanding beyond text generation:

• Oral examinations
• Live problem-solving
• Project-based assessment with viva
• Applied work where the AI-generated component is part of the deliverable

This approach is increasingly discussed as the "assessment redesign" response to AI — make AI use largely irrelevant to grading by assessing what AI cannot do.

Country patterns

Anglophone Q1 — clear policies, broad detection deployment

UK, Australia, Canada, Ireland, New Zealand. Most universities have explicit AI policies (typically "permitted with disclosure"). Detection tools widely deployed including Turnitin AI detection.

The Guardian FOI investigation (June 2025) showed 5.1/1000 UK students caught using AI in 2023–24 — confirming that AI misconduct is being identified and processed at scale in jurisdictions with strong detection.

US — institutional variation

The fragmented US higher education system shows wide policy variation:

• Elite private institutions: typically "permitted with disclosure"
• State universities: varied, often "permitted with disclosure" with significant instructor discretion
• Community colleges: less developed policies
• For-profit institutions: variable

The US scores R_det=80 (second highest in dataset) reflecting near-universal Turnitin deployment. Disclosure requirements are institutional rather than federal.

Continental Europe — institutional infrastructure, evolving policies

Netherlands, Germany, France, Italy, Spain. Most universities have developed AI policies post-2023. Detection tool deployment is partial — Compilatio (French), PlagScan (German), Antiplagiat (Russian), JSA (Polish) provide language-specific alternatives to Turnitin.

The maxed D2 scores in Italy, France, Spain (D2=100 each) suggest the policy implementation has not yet brought demand signals down to Anglophone Q1 levels.

Asia — wide variation

Singapore: strong institutional infrastructure, clear policies. Japan: post-STAP institutional integrity culture extends to AI. South Korea: post-Hwang reform context with KRI integration. China, India, Pakistan: weaker institutional implementation.

Latin America, Middle East, Africa — policy lag

Most universities in these regions are still developing AI policies. The maxed D2 signals in Colombia, Argentina, Egypt, Iran, Saudi Arabia, etc. reflect substantial demand combined with limited institutional response.

The detection-assessment-redesign debate

The detection position

Some institutions invest heavily in detection tools (Turnitin AI detection, Originality.ai, GPTZero, Copyleaks). The bet is that detection capability will improve over time, deterring AI misuse.

Limitations:

• Scarfe 2024 shows current detection misses 94%
• Iterative AI improvement may outpace detection improvement
• False positives can produce unjust accusations (particularly affecting non-native English speakers)

The assessment redesign position

Other institutions emphasise redesigning assessment to make AI use largely irrelevant:

• Oral examinations
• Live problem-solving
• Project work with iterative review
• Demonstration of understanding through Q&A
• Work that requires specific context AI cannot provide

Limitations:

• Higher staff time per assessment (oral exams are expensive)
• Some disciplines harder to redesign than others
• Existing assessment frameworks slow to change

The integrated position

Most institutions are moving toward a combination:

• Detection as one tool, not the primary defence
• Disclosure as primary expectation
• Assessment redesign as ongoing strategy
• Clear penalties for undisclosed use

What AMI data shows about effectiveness

D2 scores cluster by R-Score

Countries with strong institutional response (Q1) have moderate D2 scores (UK 44, Australia 44, Canada 44, Ireland 31). Countries with weak response have maxed D2 (Egypt 100, Iran 100, Italy 100, France 100, etc.).

The correlation suggests institutional infrastructure does affect AI submission patterns — though the direction of causation is complex (do strong-response countries have lower demand, or does lower demand enable stronger response?).

Detection capability sub-component

The R_det sub-component (Detection tools) shows the strongest cross-country correlation with reported AI misconduct rates. Countries with high R_det (UK 90, Australia 85) report more confirmed cases per student than countries with low R_det.

This is partly the detection-incidence confound: strong detection produces more reported cases. But it also reflects that without detection, misconduct goes unmeasured.

What the next two years will show

Detection improvement

Detection capability is improving but slowly. The 94% Scarfe miss rate is a 2024 measurement; iterative AI tool improvement and detection tool improvement will both continue.

Assessment redesign adoption

The assessment redesign approach is increasingly discussed. Adoption rates will be measurable over 2026–2028. Whether the approach scales beyond specific high-resource institutions is the open question.

Policy maturation

Most universities have first-generation AI policies in place. Second-generation policies — more nuanced, more discipline-specific, more aligned with assessment redesign — are emerging.

Cross-country comparison

The AMI's D2 dimension will continue to track AI submission demand. Whether countries with strong institutional response can drive D2 down toward the Q1 Anglophone range (D2=31–44) will be visible over the next few versions.

Sources

• Scarfe, P., et al. (2024), University of Reading detection study
• The Guardian (June 2025), FOI investigation on UK AI misconduct
• Institutional AI policy documentation (various universities)
• AMI v1.5 D2 dimension methodology

Full methodology | Download dataset