Date
Publisher
arXiv
The rapid advancement of Large Language Models (LLMs) presents a significant challenge to academic integrity within computing education. As educators seek reliable detection methods, this paper evaluates the capacity of three prominent LLMs (GPT-4, Claude, and Gemini) to identify AI-generated text in computing-specific contexts. We test their performance under both standard and 'deceptive' prompt conditions, where the models were instructed to evade detection. Our findings reveal a significant instability: while default AI-generated text was easily identified, all models struggled to correctly classify human-written work (with error rates up to 32%). Furthermore, the models were highly susceptible to deceptive prompts, with Gemini's output completely fooling GPT-4. Given that simple prompt alterations significantly degrade detection efficacy, our results demonstrate that these LLMs are currently too unreliable for making high-stakes academic misconduct judgments.
Study at a glance
- Question: Can prominent AI models reliably detect human-written text and their own or other models' writing in computing education?
- Setting and data: Undergraduate computer science tasks, May to early June 2025, evaluating 25 human texts and 150 AI-generated texts across five prompt categories.
- Method: Model evaluation using descriptive analysis of classification accuracy and reported confidence scores under standard and deceptive conditions.
Why it matters
Computing educators lack reliable ways to tell whether student work is original or AI-generated. Without accurate detection, institutions risk incorrectly penalizing innocent students or failing to address academic dishonesty.
Key findings
- Models struggled to identify verified human-authored work, with GPT-4 misclassifying 88% of human text as AI-generated, while Claude and Gemini had error rates of 28% and 32%.
- Adversarial prompts instructing models to write informally drastically reduced detection accuracy, with GPT-4 correctly identifying 0% and Gemini identifying only 4% of deceptive Gemini texts.
- Models' self-reported confidence scores bore no meaningful correlation to their actual classification accuracy, frequently reporting over 90% confidence on incorrect judgments.
Summary drafted with AI from the full paper and may contain errors. See the full paper for details.
What is the application?
Who is the user?
Who age?
Why use AI?
Study design
