What Is an AI-Generated Submission? Definition, Detection, and Data

TL;DR

AI-generated submission is the AMI's D2 dimension — student work produced by AI systems and submitted as the student's own. The category did not exist before ChatGPT's late-2022 launch. Top D2 scores: China (68 — slightly below top due to D6 dominance), with several countries at 100 including Egypt, Iran, Saudi Arabia, Italy, France, Spain, Greece, Colombia, Argentina, Poland. Detection remains limited (Scarfe et al. 2024).

What counts as AI-generated submission

An AI-generated submission is academic work produced by an AI system and submitted as the student's original work without authorisation or proper attribution. Common forms:

• Direct submission — copying ChatGPT/Claude/Gemini output into a document and submitting
• Lightly edited submission — making cosmetic changes to AI output and submitting
• AI-assisted drafting beyond institutional policy — using AI for substantive content generation when institutional policy permits only limited use
• AI use in proctored examinations — using AI tools during exams that prohibit them

The misconduct is in the *misrepresentation of authorship*, not necessarily in AI use itself.

How institutional policies vary

Universities have responded to AI tools with widely varying policies:

Permitted with disclosure

Most common model. Students can use AI for brainstorming, grammar checking, and similar support tasks with disclosure of the specific use. Original drafting and analytical work must be the student's own.

Prohibited entirely

Some courses and institutions ban AI use entirely. Common in fundamental skills courses (writing, mathematics) where the assessment specifically targets capabilities AI can perform.

Permitted with attribution

Some courses permit substantial AI use provided the student properly attributes the AI contribution. Less common; raises questions about what is being assessed.

Permitted without restriction

Rare; some institutions and instructors permit AI use without restriction, treating it as a writing tool. Often combined with assessment design that requires demonstrable understanding beyond text generation.

What the AMI data shows

D2 scores on a 0–100 scale across the 39-country set. Top D2 scores (multiple countries at 100):

• Colombia, Argentina, Greece, Egypt, Iran, Saudi Arabia, Italy, France, Spain, Poland — all D2=100

Lowest D2 scores: Sweden (31), New Zealand (31), Japan (31), Norway (31). The lowest D2 scores cluster in Nordic and Pacific countries with smaller language markets.

Top scores reflect demand-side search-volume signals. Many of the maxed-D2 countries have Spanish, Arabic, Italian, French, or Polish language markets where AI submission tool search volume is at the top of the European or regional distribution.

Detection methods

Automated detection

Turnitin, GPTZero, Originality.ai, Copyleaks, and others provide AI-generated content detection. Capabilities have improved since 2023 but reliability remains limited:

• False positives are common — human-written text can register as AI
• False negatives are common — AI text can pass detection, particularly if lightly edited
• Reliability degrades for shorter text, paraphrased text, and non-English text

Human review

• Stylistic inconsistency between the submission and the student's known writing
• Conceptual errors typical of LLMs (incorrect citations, factual hallucinations)
• Response to follow-up questions — students who didn't write the work struggle to discuss it substantively

Scarfe et al. (2024)

The most cited study on AI submission detection. The University of Reading study submitted AI-generated work through normal coursework channels; 94% went undetected. The study established that current human + automated detection misses the majority of AI submissions.

The AMI's D2 methodology

D2 is primarily measured through Google Trends signals for AI submission tool keywords. The signal captures demand for tools rather than confirmed incidence. The methodology applies a detection-ratio correction (Scarfe 94% undetection) to estimate true incidence from detected case rates where institutional data is available.

The reliance on demand signals is a methodology limitation. The Norway case illustrates the issue — high D2 search volume in Norway partly reflects academic and policy discussion rather than student demand. Future AMI versions will improve language-disambiguated signal interpretation.

Why AI submissions matter for the AMI

AI submissions are the newest misconduct category. Before late 2022, the category did not meaningfully exist. The rapid adoption of ChatGPT (over 100 million users within two months of launch) created a structural shift in academic misconduct that the AMI captures.

The category interacts with other dimensions. AI tools can be used for:

• D1 contract cheating substitution (cheap AI replacing paid human writers)
• D4 plagiarism (AI text may incorporate uncredited training-data content)
• D6 data fabrication (AI tools can generate plausible-looking fake research data)

Sources

• Scarfe, P., et al. (2024). "A real-world test of artificial intelligence infiltration of a university examinations system: A 'Turing Test' case study"
• Guardian FOI reporting on UK AI misconduct cases
• Turnitin AI detection deployment data
• AMI v1.5 methodology document

Full methodology | Download dataset

Related

Read the full methodology