AI Submission Misconduct Figures 2026: Global Data

TL;DR

AI submission (D2) data from AMI v1.5. Demand signal maxed in 11 countries. Guardian FOI confirms 5.1/1000 UK students caught in 2023–24. Scarfe et al. (2024) detection study: 94% AI miss rate. Detection capability is the principal current limitation.

Demand signal rankings — D2 score

Top — D2 = 100 (11 countries tied)

• Colombia
• Argentina
• Greece
• Egypt
• Iran
• Saudi Arabia
• Italy
• France
• Spain
• Poland
• (one other)

These countries all show the top of the per-country distribution for AI submission tool search volume. The signal includes searches in Spanish (driving the Latin American and Spanish cluster), Italian, French, Arabic, Persian, Polish, and Greek.

Middle band — D2 in 50s–60s

• China (68)
• Pakistan (66)
• India (62)
• Indonesia (62)
• Malaysia (62)
• South Korea (56)
• Brazil (56)
• Mexico (56)
• Thailand (56)
• Philippines (56)
• Vietnam (50)
• Turkey (50)

Lower band — D2 in 30s–40s

• US (44)
• Canada (44)
• UK (44)
• Australia (44)
• South Africa (44)
• Singapore (44)
• Germany (44)
• Japan (31)
• Norway (31)
• Sweden (31)
• Netherlands (31)
• Ireland (31)
• New Zealand (31)

The Guardian FOI investigation

In June 2025, The Guardian published a Freedom of Information investigation showing UK university confirmed cases of AI misconduct:

• Total cases: nearly 7,000 across UK universities in 2023–24
• Rate: 5.1 per 1,000 students
• Coverage: most UK universities responded to the FOI request
• Categories: formal misconduct findings for AI use including ChatGPT, Bard, and other tools

The Guardian data is the largest single confirmed-case dataset for AI submissions globally. Times Higher Education has published similar FOI data for Russell Group universities specifically [verify].

The Scarfe et al. (2024) study

The University of Reading study tested AI submission detection capability:

• Method: AI-generated submissions submitted through normal coursework channels at the university
• Coverage: psychology undergraduate assessments
• Detection rate: 6% (i.e. 94% of AI submissions went undetected)
• Detection methods tested: combined human marker review and Turnitin AI detection

The study established that the current detection regime catches only a small minority of AI-generated submissions. The 94% miss rate has been widely cited as the empirical benchmark for AI detection capability.

Implications

Applying the Scarfe correction to the Guardian FOI data:

• 5.1 per 1,000 confirmed cases × (1 / 0.06) detection factor = 85 per 1,000 true rate estimate
• This implies an 8.5% true incidence rate — substantially above the confirmed rate

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

US data — Stanford and other studies

Stanford has run confidential undergraduate surveys post-ChatGPT, reporting:

• 9–16% of undergraduates use AI for assignments [verify specific Stanford study citations]
• Rates vary by discipline and course type

Other US studies have produced comparable estimates. The US D2 score of 44 reflects this moderate range — well below the maxed-D2 countries but consistent with substantial actual use.

Time series

The category did not meaningfully exist before late 2022:

• Pre-November 2022: ChatGPT had not launched; AI tools were available but not at scale
• December 2022 – mid 2023: rapid adoption of ChatGPT by students; many universities reactive rather than proactive
• 2023 – present: institutional policies developing; detection tools rolling out; FOI data emerging

The two-and-a-half-year history of the category means time-series data is limited. The 2024 and 2025 FOI data will be more informative than the 2023 data given institutional policy maturation.

Country detection asymmetry

The detection gap between countries is substantial:

• High detection deployment: UK, Australia, US, Canada have widespread Turnitin AI detection
• Moderate deployment: most European countries
• Limited deployment: many Q3 and Q4 countries have minimal AI detection beyond basic plagiarism tools

The detection-incidence confound applies strongly to D2. Countries with stronger detection report more cases — not necessarily because they have more actual incidence but because they catch more of what occurs.

Substitution dynamics

A major open question: are students substituting AI for contract cheating, or using both?

Evidence for substitution:

• Some essay mill brand name search declines in 2023–2024
• Free AI alternatives reduce price advantage of contract cheating
• AI is faster than commissioning a human writer

Evidence against substitution:

• AI detection is improving (slowly); contract cheating is harder to detect
• Some assessments (longer, more complex) still favour human writers
• Contract cheating revenue has not collapsed per industry reporting [verify]

The likely answer: substitution at the margins but coexistence overall. Future AMI versions will track the dynamic.

Sources

• The Guardian (June 2025), FOI investigation [verify specific article reference]
• Scarfe, P., et al. (2024), "A real-world test of artificial intelligence infiltration of a university examinations system"
• Stanford and other US-based confidential survey data [verify specifics]
• AMI v1.5 dataset and methodology

Full methodology | Download dataset

Related

Explore the full dataset