Artificial intelligence is changing the way people seek information and communicate, but recent research and industry reports suggest that the technology still has significant hurdles to overcome in truly mimicking human conversation—and its rise is already reshaping the digital landscape for major online publishers. On October 20, 2025, two developments highlighted both the promise and the pitfalls of AI: a new study revealed the persistent awkwardness of AI-generated dialogue, while Wikipedia reported a notable drop in human traffic as users increasingly turn to AI-powered answers.
Let’s start with the question that’s been on many tech-watchers’ minds: can large language models like ChatGPT-4, Claude Sonnet 3.5, Vicuna, and Wayfarer really pass for human in conversation? According to a study led by Associate Professor Lucas Bietti from the Norwegian University of Science and Technology (NTNU), the answer remains a resounding "not quite yet." The research, published on October 20, compared transcripts of actual human phone conversations with those generated by leading AI systems, and then asked people to distinguish between the two. The result? Most participants could still spot the bot.
So, what’s giving these AI models away? The NTNU study found several telltale signs. First, there’s what the researchers call “exaggerated alignment.” When humans talk, we naturally mirror each other’s speech patterns a little bit—think of it as conversational chemistry. But AI, it turns out, takes this to an almost comical extreme. "Large language models are a bit too eager to imitate, and this exaggerated imitation is something that humans can pick up on," explained Bietti, as reported by Neuroscience News.
Then there’s the issue of “filler words”—those little conversational glue words like “so,” “well,” “like,” and “anyway.” In real life, these words help us sound natural, signal interest, or structure our thoughts. But AI models often misuse or overuse them, disrupting the flow and making the conversation sound just a bit off. "The large language models use these small words differently, and often incorrectly," Bietti noted. It’s a subtle thing, but it’s enough to tip off most listeners that they’re chatting with a machine.
Another area where AI stumbles is in the art of starting and ending a conversation. Humans rarely jump straight into business; we ease in with a “hey” or “how are you doing?” and wrap up with a friendly “talk to you later” or “see you soon.” AI, however, often fumbles these transitions, either skipping them entirely or using them in ways that feel forced or unnatural. As Bietti put it, “This introduction, and the shift to a new phase of the conversation, are also difficult for large language models to imitate.” The same awkwardness plagues their attempts to close a conversation gracefully.
Despite these shortcomings, the researchers acknowledge that AI is evolving rapidly. While today’s models can’t consistently fool humans, tomorrow’s might come closer. Still, Bietti and his colleagues caution that some differences—like the subtle timing, empathy, and social intuition that define genuine human exchanges—may always set us apart from our digital counterparts. "Improvements in large language models will most likely manage to narrow the gap between human conversations and artificial ones, but key differences will probably remain," Bietti concluded.
While AI’s conversational abilities are still a work in progress, its impact on the broader information ecosystem is already being felt—and not always in ways that please the old guard. On the same day as the NTNU study’s release, the Wikimedia Foundation, which operates Wikipedia, reported an 8 percent year-on-year decline in human visits to its site. The culprit? People are increasingly turning to AI tools like ChatGPT and Google’s AI Overviews to get their questions answered directly, often without ever clicking through to Wikipedia itself.
This drop in traffic was only discovered after Wikipedia updated its bot detection systems, revealing that the shift was even more pronounced than previously thought. Marshall Miller, writing on behalf of the Wikimedia Foundation, explained, "After making this revision, we are seeing declines in human pageviews on Wikipedia over the past few months, amounting to a decrease of roughly 8 percent as compared to the same months in 2024." The Foundation believes this reflects the growing influence of generative AI and social media on how people seek information, with search engines increasingly serving up direct answers—often based on Wikipedia content—rather than simply linking to the site.
Wikipedia is hardly alone in feeling the squeeze. Other publishers have also reported significant drops in traffic as AI-generated summaries and answers siphon off would-be visitors. DMG Media, which owns MailOnline and Metro, told the UK’s Competition and Markets Authority that click-through rates from Google’s AI Overviews had plummeted by a staggering 89 percent. Meanwhile, Penske Media Corporation, parent of Rolling Stone, recently filed a lawsuit against Google, alleging that the tech giant’s AI-generated article summaries are cutting into its online readership.
Google, for its part, has pushed back against these complaints. In an August 2025 blog post, Liz Reid, head of Google Search, insisted that the company remains committed to supporting publishers. "Our data shows people are happier with the experience and are searching more than ever as they discover what Search can do," Reid wrote. She argued that while AI responses might provide a quick overview, users still click through to learn more—and that when they do, those clicks are "more valuable." Reid added, "While overall traffic to sites is relatively stable, the web is vast, and user trends are shifting traffic to different sites, resulting in decreased traffic to some sites and increased traffic to others. We continue to send billions of clicks to websites every day and believe that Search’s value exchange with the web remains strong."
This tug-of-war between AI-generated convenience and the traditional web isn’t likely to be resolved anytime soon. On one hand, AI tools are making it easier than ever for users to get quick answers, but on the other, they’re disrupting the traffic—and, by extension, the business models—of the very sites that supply much of that information. It’s a classic case of innovation colliding with the status quo, and both sides are scrambling to adapt.
Back in the world of AI conversation, the NTNU study’s findings are a reminder that, for all their sophistication, language models still have a way to go before they can truly pass for human. As Bietti and his colleagues suggest, the social subtleties that make human dialogue so rich and nuanced may never be fully captured by code. Meanwhile, as AI continues to reshape how we search, learn, and communicate, the ripple effects are being felt across the internet—from Wikipedia’s boardrooms to the daily choices of millions of users. The digital conversation, it seems, is only just beginning.
