More than half of all new articles published online are now generated by artificial intelligence (AI), according to independent research by search analytics firm Ahrefs, which found that 74.2% of newly created web pages in April 2025 contained AI-generated content. A separate study by search engine optimisation (SEO) firm Graphite, which analysed 65,000 English-language web articles, put the share of AI-written articles at approximately 52%. Against that backdrop, a new comparative study has tested which AI platforms are hardest for detection software to identify.
Experts at Open Resource Applications, a company developing free AI tools, gave 12 AI writing platforms identical instructions to produce articles of between 1,000 and 1,500 words written to sound as human as possible. Each piece was then evaluated by three detection tools: Grammarly, QuillBot, and GPTZero.
Google’s Gemini emerged as the hardest AI platform to detect, with an average detection rate of 39% across the three tools. Grammarly flagged just 17% of Gemini’s output, while QuillBot detected none of it. Anthropic’s Claude followed at 41%, with Grammarly finding no AI signals at all in its text. Grok, the AI assistant developed by xAI, ranked third at 46.33%, despite Grammarly identifying 39% of its content.
A spokesperson for Open Resource Applications attributed Gemini’s performance to the diversity of writing styles now possible across newer platforms. “Tools like GPTZero flag predictability and overall structure, so a model that actually reasons through ideas rather than recycling familiar phrases is going to be a lot harder to catch,” the spokesperson said.
ChatGPT, the most widely used AI platform globally with between 800 million and 900 million monthly users, ranked ninth out of twelve tools in the study. Grammarly detected 50% of its output, while QuillBot and GPTZero flagged between 90% and 100% of its text as machine-generated. The spokesperson said ChatGPT’s early market dominance works against it: “Everyone knows what it sounds like. Many models that came after it sounded like ChatGPT first, before they became more unique.”
Among the three detection tools tested, GPTZero proved the most difficult to deceive, catching an average of 98.8% of AI-generated content across all platforms. Grammarly showed the most limited ability to identify AI texts, correctly flagging an average of just 43.5% of generated content. Only Claude and Meta AI managed to confuse GPTZero with any consistency.
The study also found that none of the three detection tools falsely identified properly written human text as AI, adding confidence to the reliability of their positive detections when they do flag content.
Independent research suggests the growth of AI-generated content online has plateaued in recent months, with analysts hypothesising that publishers have found AI articles tend not to perform as well in search results. The detection arms race between generative AI platforms and the tools designed to identify them, however, continues to accelerate, with the gap between the best and worst-performing AI systems growing wider with each model generation.
