AI is writing more like humans - but do we actually like it?

As artificial intelligence (AI) advances, large language models (LLMs) such as GPT-4, Claude, and Llama have increasingly produced text that is nearly indistinguishable from human writing. With AI-generated content proliferating across news, academic writing, and social media, an important question arises: Can humans still distinguish between human and machine-generated text? More crucially, do people prefer human-authored text over AI-generated content?

A recent study titled "Is Human-Like Text Liked by Humans? Multilingual Human Detection and Preference Against AI," conducted by Yuxia Wang et al., investigates these questions across multiple languages and domains. Published in 2025, the research examines the extent to which people can correctly identify AI-generated content and whether their preferences lean toward human-written text. By analyzing human detection accuracy across 16 datasets covering nine languages and nine domains, the study challenges existing assumptions about human perception of AI-generated language.

How well can humans detect AI-generated text?

Prior studies have suggested that distinguishing between human and AI-generated text is highly challenging, with detection accuracy often approaching random guessing. However, Wang et al.'s study presents a contrasting finding: Across their dataset, human annotators achieved an average detection accuracy of 87.6%, significantly higher than previous reports.

The study highlights key linguistic features that differentiate human and machine-generated text. Human writing tends to be more concrete and culturally nuanced, often incorporating specific references such as dates, locations, and idiomatic expressions. In contrast, AI-generated text is more structured and neutral, often exhibiting uniform sentence length and polished grammatical structures but lacking the subtle irregularities present in natural human writing. Moreover, AI-generated text frequently adheres strictly to logical progression, whereas human-authored text may include unexpected topic shifts, humor, or rhetorical devices that make detection easier.

Interestingly, detection accuracy varied across languages and domains. Annotators found it easier to distinguish AI-generated content in news articles, peer reviews, and academic abstracts, where nuanced arguments and expertise play a critical role. However, they struggled more with Wikipedia-style entries, summaries, and social media posts, where AI models excel in producing fluent, factual, and stylistically consistent content.

Does AI-generated text meet human preferences?

One of the study's most striking findings is that humans do not always prefer human-written text. Across multiple languages and domains, human participants demonstrated a significant preference for AI-generated content in certain contexts. This preference was particularly strong in Russian and Arabic summaries and tweets, where AI-generated text was often perceived as clearer, more concise, and better organized than human-written equivalents.

However, in emotionally driven content such as opinion pieces and social interactions, human-written text was generally favored. This suggests that AI-generated content is often optimized for clarity and coherence but may still lack the depth of personal expression and emotional resonance that human authors naturally produce.

The study also finds that prompting AI models with more explicit instructions to mimic human writing styles can partially bridge the gap between machine and human-generated text. In over 50% of cases, strategic prompting led to more human-like AI outputs, making detection even more difficult. This finding raises ethical concerns, as improved AI-generated text could further complicate the ability to distinguish between real and synthetic content in critical areas such as journalism and academic writing.

Implications for AI development and content authenticity

The findings of this study have significant implications for AI developers, content creators, and policymakers. As AI-generated text becomes increasingly human-like, there is a growing need for better detection mechanisms to safeguard against misinformation, AI-assisted plagiarism, and unethical content manipulation.

At the same time, the study's insights into human preference patterns suggest that AI developers should focus on enhancing cultural and contextual adaptability in LLMs rather than merely improving grammatical accuracy. Given that human preferences vary by language and content type, future AI models may need to incorporate more sophisticated style adaptation techniques to align with user expectations.

From a regulatory perspective, these findings also emphasize the importance of transparency in AI-generated content. As AI continues to integrate into education, media, and professional fields, ensuring that users are aware of whether they are interacting with AI-generated or human-written content will be crucial in maintaining trust and accountability in digital communications.

In conclusion, Wang et al.'s research presents a nuanced perspective on the evolving relationship between human and AI-generated text. While humans can still detect AI-generated content with relative accuracy, their preferences for AI versus human writing vary based on context. As AI technology continues to evolve, understanding these shifting dynamics will be essential for shaping the future of AI-human interaction in written communication.