Guardian FOI Investigation: 7,000 UK Students Caught Using AI

TL;DR

The Guardian Freedom of Information investigation (June 2025) showed nearly 7,000 UK university students were formally caught using AI tools in 2023–24 — 5.1 per 1,000 students. The data is the largest single confirmed-case dataset on AI misconduct globally and contributes to the AMI's D2 dimension methodology.

What the Guardian did

In early 2025, The Guardian submitted Freedom of Information requests to UK universities asking for statistics on formal academic misconduct cases involving AI tools for the 2023–24 academic year. The investigation:

• Scope: most UK universities responded
• Output: nearly 7,000 confirmed cases across the responding institutions
• Rate: 5.1 per 1,000 students (formally caught and processed)
• Categories: ChatGPT, Bard, Claude, Gemini, and other AI tool use treated as academic misconduct

The investigation was published in June 2025 as a long-form piece with institution-level breakdowns and editorial analysis.

Why the data matters

First systematic national-level confirmed-case dataset

Before the Guardian investigation, AI misconduct data was fragmented — individual institutions reported numbers, anecdotal evidence circulated, but no systematic national picture existed.

The Guardian data provided:

• A national-level confirmed-case rate for the first time
• Institution-level breakdowns enabling cross-institutional comparison
• A baseline for measuring whether incidence is growing or falling

Detection-incidence reasoning

The 5.1/1000 confirmed rate combined with the Scarfe et al. (2024) detection finding (94% of AI submissions undetected) suggests:

• 5.1/1000 detected × (1/0.06) detection factor = 85/1000 true rate estimate
• This implies an ~8.5% true incidence rate

The AMI methodology applies similar detection corrections in calculating D2 scores from observed signals.

Institution-level variation

The Guardian data showed substantial variation across UK universities:

• Some institutions reported very high rates (e.g. above 10/1000)
• Others reported very low rates (below 1/1000)

The variation likely reflects differences in:

• Detection investment and deployment
• Institutional reporting practices
• Specific course types and assessment design
• Student population composition

The variation is itself informative — it suggests detection capability is the principal determinant of reported rates, with actual incidence relatively more uniform across institutions.

Russell Group FOI follow-up

Times Higher Education published a parallel investigation focused specifically on UK Russell Group universities. The Russell Group data:

• Showed lower per-student rates than the broader UK pattern [verify specific Russell Group rate]
• Reflected stronger detection infrastructure at research-intensive institutions
• Provided cross-institutional comparison among peer institutions

The Russell Group pattern is consistent with the broader hypothesis: stronger detection capability produces more reported cases (in jurisdictions where detection drives reporting), while actual incidence varies less.

How the AMI uses the data

Direct UK D2 contribution

The AMI's D2 dimension for the UK is partially anchored on the Guardian confirmed-case rate. The UK D2 score of 44 reflects:

• Moderate Google Trends signal
• Confirmed-case rate from FOI
• Detection capability adjustment

The UK D2 is lower than the maxed-D2 countries (Colombia, Argentina, etc.) partly because of the confirmed-case data providing actual rates rather than relying on demand signals alone.

Cross-country calibration

The Guardian data provides a calibration point for D2 across the dataset. Countries without confirmed-case data are scored on demand signals (Google Trends), but the Guardian baseline lets the methodology estimate the demand-to-incidence translation more confidently.

Methodology improvement

Future AMI versions will incorporate similar FOI data as it becomes available in other jurisdictions:

• Australia (FOI submissions reportedly in progress)
• US (state-level FOI variability)
• Other Anglophone systems

The Guardian investigation effectively demonstrated FOI as a viable data source for AI misconduct measurement.

Why the UK could produce this data

The UK FOI Act applies to universities (most are public-funded). Combined with:

• Active investigative journalism (the Guardian's higher education team)
• Substantial UK university sector size enabling meaningful statistical analysis
• Standardised institutional misconduct reporting categories

The UK is structurally well-positioned to produce this kind of national dataset. Other countries with weaker FOI laws, less centralised university systems, or less active education journalism cannot easily replicate.

What the data does not show

Undetected misconduct

The 5.1/1000 rate is *confirmed-case* — students formally identified and processed for misconduct. The Scarfe correction (94% undetected) suggests the actual rate is substantially higher.

Severity gradient

Confirmed cases include a wide range of misconduct severity — from minor undisclosed AI assistance through wholesale AI-generated work submitted as original. The Guardian data does not differentiate.

Course type variation

Different subject areas likely have very different AI misconduct rates. Essay-based humanities and social science courses are likely more affected than mathematics or laboratory-based STEM courses. The Guardian data aggregates across course types.

Trend data

The Guardian data covers 2023–24. The 2024–25 data (still being collected) will be more informative for trend analysis.

Broader significance

The Guardian investigation established that:

• FOI is a viable data source for AI misconduct measurement
• Institutional variation in reporting reflects detection capability
• Confirmed-case rates substantially understate true incidence (Scarfe correction)
• AI misconduct is a measurable population-level phenomenon, not an anecdotal one

The reporting accelerated UK policy discussion of AI in education and contributed to detection tool investment decisions at institutions across the sector.

Sources

• The Guardian (June 2025), FOI investigation [verify specific article reference]
• Times Higher Education Russell Group FOI reporting [verify]
• Scarfe, P., et al. (2024), University of Reading detection study
• AMI v1.5 D2 dimension methodology

Full methodology | Download dataset