Machine Generated Text: A Comprehensive Survey of Threat Models and Detection Methods

AI-generated text is becoming indistinguishable from human writing, making it increasingly difficult to identify misuse through manual inspection alone. The survey emphasizes that powerful, freely available open-source models and user-friendly tools like ChatGPT have democratized access to generative AI, expanding both legitimate uses and potential for abuse.

Multiple serious threat models exist beyond simple deception, including academic plagiarism, disinformation campaigns, spam, impersonation, and automated harassment. The paper stresses that addressing responsible AI requires understanding these diverse abuse scenarios, not just detecting whether text is machine-generated.

Current detection methods face significant technical challenges and have no perfect solution. The survey indicates there are "numerous open problems" in reliably identifying machine-generated text, suggesting that detection alone cannot be the sole safeguard against misuse.

Detection systems themselves must be trustworthy and fair. The paper emphasizes that solutions need to demonstrate "fairness, robustness, and accountability"—meaning detection tools should work reliably across different contexts, resist manipulation, and be transparent about their limitations and potential biases.

A comprehensive, multi-layered approach is necessary for responsible AI deployment. Rather than relying on detection as a single fix, the survey indicates that addressing machine-generated text abuse requires understanding threat models within their "cybersecurity and social context" and developing guidance for future safeguards.