Why can't people just tell the difference themselves?
The honest answer is that most people cannot reliably tell AI-generated content from human-written content, even when they think they can. In a series of six experiments involving 4,600 participants, researchers found that people were unable to detect self-presentations generated by state-of-the-art AI language models in professional, hospitality, and dating contexts [2]. The study showed that humans rely on intuitive but flawed heuristics—for example, assuming that using first-person pronouns like 'I' or 'we,' contractions like 'don't,' or talking about family topics signals a human author. These cues are easy for AI to mimic, and the researchers demonstrated that AI systems can be deliberately tuned to produce text that people perceive as 'more human than human' [2]. This means relying on users' own judgment is not a viable strategy for platforms.
Do platforms need brand-new policies for AI content?
Counterintuitively, the most effective step platforms can take is not to create separate rules for AI-generated content, but to rigorously enforce the rules they already have. A 2024 analysis argues that the harm caused by AI-generated content—misinformation, propaganda, non-consensual deepfakes—is no different in kind from the harm caused by ordinary harmful content [5]. Generative AI massively increases the scale and speed of the problem, but it does not create a fundamentally new type of threat. Therefore, platforms should 'double down on improving and enforcing their existing rules, regardless of whether the content they were dealing with was produced by humans or machines' [5]. This approach avoids the trap of trying to police the technology itself and instead focuses on the actual harm, which is a more manageable and legally sound strategy.
What practical tools can platforms use right now?
Two concrete, evidence-backed tools are available: human-in-the-loop quality checks and technical markers like watermarks or 'AI accents.' A 2024 study introduced the HEAT heuristic (Human experience, Expertise, Accuracy, and Trust) as a rating mechanism for evaluating AI-generated content [3]. When tested with beginner technical communication students, HEAT showed good reliability, with an intraclass correlation coefficient of .743 in a pilot and .825 in subsequent scenarios—meaning different evaluators using the tool gave consistent scores [3]. This kind of structured human review can catch errors and improve content quality, especially for complex tasks like step-by-step instructions where AI often performs poorly [3]. On the technical side, researchers have proposed 'AI accents'—subtle, detectable markers embedded in AI output—to reduce the deceptive potential of synthetic language [2]. Similarly, a 2023 survey on ChatGPT and AI-generated content reviewed state-of-the-art watermarking approaches that can tag AI-produced text or images, making them traceable without breaking the user experience [4]. These tools work best in combination: technical markers for detection, and human reviewers for quality and ethical judgment.
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2023 to 2025, 3 from 2024 or later, 2 in Q1 journals, collectively cited 499 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 44 papers retrieved from a database of over 500 million.
Sources used in this answer
Evil Cannot Create: J.R.R. Tolkien’s Philosophy and the Misuse of AI-Generated Content.
Drawing on Tolkien's philosophy, this study argues that AI cannot create genuine meaning—only corrupt existing human creativity—and concludes that AI should augment, not replace, human creative work, requiring ethical awareness and policies to protect authorship.
Human heuristics for AI-generated language are flawed
In six experiments with 4,600 participants, humans were unable to detect AI-generated self-presentations and relied on flawed heuristics (e.g., first-person pronouns, contractions, family topics) that AI can easily mimic, allowing AI to produce text perceived as 'more human than human'; the study proposes 'AI accents' as a solution.
Incorporating Human Judgment in AI-Assisted Content Development: The HEAT Heuristic
In an exploratory case study with beginner technical communication students, the HEAT heuristic (Human experience, Expertise, Accuracy, Trust) showed good reliability for evaluating AI-generated content (intraclass correlation coefficient of .743 pilot, .825 scenarios), and expert human input improved prompting for better AI output.
A Survey on ChatGPT: AI–Generated Contents, Challenges, and Solutions
This survey of ChatGPT and AI-generated content reviews security, privacy, ethical, and legal challenges, and discusses state-of-the-art watermarking approaches for making AI-generated content regulatable and traceable.
Moderating Synthetic Content: the Challenge of Generative AI
This analysis argues that the threat from AI-generated content is not different in kind from ordinary harmful content, so platforms should enforce existing rules rather than create new sui generis policies for synthetic content.