As more and more of the Internet becomes artificial intelligence (AI)-generated following the technology’s boom after ChatGPT’s debut in late 2022, concerns about content created using the technology have run rampant – be it due to their potential to spread misinformation, or even the idea that human beings are being displaced from online discussions entirely because of AI-powered bots.
Well, according to a study from a trio of researchers from Imperial College London, the Internet Archive, and Stanford University, titled “The Impact of AI-Generated Text on the Internet”, there’s a lot more to the issue than meets the eye.
Findings from the study include a significant figure: around 35% of newly published websites uploaded between August 2022 (around the time ChatGPT hit the scene) and May 2025 to the Internet were either AI-generated or AI-assisted.
This may not come as a huge surprise to some, with the growing amount of content being labelled “AI-generated”, and even some AI-powered “influencers” already making waves on platforms like Instagram and TikTok.
Others might be quick to shout “Dead Internet Theory”, referring to the idea that most online activity now consists of bots, generating synthetic content or falsehoods, which are then used to tell human onlookers how and what they should be thinking about.
However, the study found that this isn’t the case. While there is a lot more AI-generated content online now, human fact-checkers involved in the study found that the information itself was not wrong.
In fact, it also found that several other common assumptions were not necessarily true, including the belief that AI-written content rarely cites sources of its information, becomes longer while saying less, or causes writing styles across the Internet to become less distinct.
These findings were unexpected, according to one of the paper’s authors, Jonas Dolezal, in an interview with online publication 404 Media.
“The most surprising result was that our Truth Decay hypothesis wasn’t confirmed. It’s worth noting that we were specifically looking for an increase in verifiably untrue statements, which we didn’t find.
“But it could still be the case that AI is quietly increasing the volume of unverifiable claims, ones that can’t be checked against existing fact-checking tools and infrastructure. Or it may simply be that the Internet wasn’t a particularly truth-adhering place to begin with,” he says.
Instead, the study found other shifts. It saw the range of unique ideas and diverse viewpoints shrink with the use of AI, while the writing became more sanitised and artificially cheerful.
The question is, what does this actually mean for users on the Internet?
More than meets the eye
AI Society president Dr Azree Nazri says while the technology could be seen as a way to increase productivity for text-based online content creation, it may also have a knock-on effect on the variety and diversity of what users see on the Internet.
He believes this may make the Internet feel repetitive, as a search on a topic may return many different AI-generated or AI-assisted articles drawing from the same datasets, reports and sources, without introducing new ideas.
“The risk is that while content volume increases massively, diversity of perspectives, original field research, and unique human analysis begin to shrink,” Azree says.
Asia Pacific University of Technology and Innovation (APU) vice-chancellor Professor Dr Ho Chin Kuan shares similar thoughts, particularly on what will happen once the majority of the Internet consists of AI-generated content.
“Firstly, AI models will increasingly train on AI-written content, which will narrow the range of ideas over time. Secondly, the Web stops being a reliable mirror of human thought and becomes more of a synthetic layer built on top of it. The Internet won’t break – but its role as a record of what people actually think and say will weaken,” he explains.
Azree, who is also a senior lecturer of AI at University Putra Malaysia, believes that this is the largest long-term risk, given how AI models are trained on existing data and naturally tend to reproduce patterns, assumptions, and viewpoints that already dominate online sources.
In cases like these, while the information itself is not incorrect, it can then be used to amplify certain topics with the massive amount of content produced.
This rings true for Professor Ho, who points out that the ability to flood content on certain issues to set the agenda is often overlooked.
“Agenda-setting has always been about what gets covered, not just how. If AI tools make it easier to publish on certain topics, drive traffic, or align with whoever is deploying the tools, those topics gain visibility while others quietly fade. Volume then becomes a form of editorial power,” he says.
He adds that the information being spread does not necessarily have to be false or inaccurate.
“Numbers can be made misleading as well. One can be technically accurate and still manipulative, through framing, selective emphasis, or sheer volume.
“The bigger risk isn’t false claims, but coordinated campaigns that flood comment sections, reviews, and social feeds designed to shift perception.
“Spam and astroturfing (a deceptive practice where an organisation or group creates the false appearance of grassroots or public opinion) scale far more cheaply than fact-checking does,” he says.
Dr Afizan Azman, an Associate Professor at Taylor’s University School of Computer Science, believes that beyond raising the perceived importance of certain issues, this could also be used to influence online algorithms.
“Algorithmic reinforcement, where many social media platforms use recommendation engines to drive user engagement, can lead to trending topics, repeated narratives and echo chambers,” he says.
“This raises concerns about disproportionate visibility, where certain topics dominate attention simply because they are easier or cheaper to produce at scale,” he adds.
Polite problems
Experts believe that the nature of AI-generated text, which tends to be more sanitised and generally positive in tone, raises a different set of concerns, particularly when it comes to making certain issues less urgent.
“In many instances, AI models are trained to be balanced and non-alarming (could be part of their guardrails), which is useful in many contexts,” Professor Ho says. “But this doesn’t augur well when the subject genuinely warrants concern.
“For example, a pandemic, a war, a climate crisis, or a corporate scandal described in a pleasant, even-handed way can feel less like a problem demanding attention. Over time, that tonal smoothing can reduce public urgency,” he explains.
This tendency comes as part of the design of large language models (LLMs), according to Afizan.
“AI-generated content is typically designed to be polite, balanced, and helpful. It often avoids extreme or emotionally intense language and tends to present information in a calm and neutral tone.
“However, people often judge importance based on emotional cues. As a result, content that appears calm and neutral may feel less urgent or less morally pressing.
“A key concern is if AI-generated content consistently sounds optimistic or neutral, important societal issues such as public health or climate change may appear less alarming than they actually are,” he says.
From Azree’s perspective, this could even extend to building a skewed understanding of reality. In cases where new opportunities or developments are being discussed, the benefits may receive more focus, while the risks and complexities involved may be undersold.
This can happen all while the information being presented remains factually accurate.
“For example, if there are real risks in an industry – such as drone accidents, privacy concerns, or airspace conflicts – AI-generated content may still overwhelmingly emphasise growth, innovation, investment opportunities, and future economic benefits.
“That creates a highly optimistic narrative, but it can reduce attention on safety, regulation, or operational risks.
“Over time, decision-makers may start underestimating the seriousness of those issues because the dominant online discussion appears overwhelmingly positive,” he says, adding that this could happen in other sectors as well.
This is also an aspect the study highlights, touching on how cheerful and homogenised text may end up influencing public thinking and pacifying those with more dissenting views.
Another concern is how online platforms have mechanisms aimed at moderating harm, such as hate speech or factual inaccuracies, but are not equipped to address subtler issues like a decline in unique ideas or diverse viewpoints.
Mitigating measures
The way that Afizan sees it, the biggest thing to worry about isn’t that humans will be replaced by AI online, but rather that more of the Internet will instead be shaped by AI rather than the human experience.
“As a result, online discourse may become more uniform, less experimental, and less culturally distinctive. This may also reduce the visibility of minority or niche perspectives, even when they are important.
“Human writing often includes imperfections, humour, contradictions, strong emotions, and personal or cultural nuance. In contrast, AI-generated content tends to be more structured, balanced, and optimised.
“If AI-generated content becomes dominant, it may reduce critical debate, weaken intellectual diversity, and make it more difficult to identify trustworthy information. This could lead to confusion, cynicism, and reduced trust in institutions and media, while also making attention manipulation easier,” he says.
This is something that Professor Ho believes is already happening, as widely used AI models are making language and viewpoints more uniform. At the end of the day, unconventional perspectives and minority viewpoints are becoming harder to find.
“The ‘Dead Internet’ idea is an overstatement of things, but the underlying intuition makes sense: more of what we scroll past is generated rather than written, and more interactions are with systems rather than people.
“That changes public discussion in subtle ways – debates feel dull, consensus appears manufactured, and it gets harder to tell whether an idea is genuinely popular or just heavily produced and propagated,” he says.
Both Afizan and Professor Ho call for clearer labelling on AI-generated content online, but recognise the difficulty in enforcement.
Afizan says implementing reliable labelling at Internet scale is both technically and socially challenging, as AI use often exists on a spectrum rather than being fully human or fully machine-generated.
“Detection tools are not always reliable, and the sheer volume of online content makes enforcement difficult,” he says.
Professor Ho concurs, adding that while people deserve to know whether content is written by a person or generated by AI, trying to get content creators to label every single piece may not be practical.
“A more workable approach is disclosure at the platform level. This requires publishers and major platforms to declare their use of AI in content production,” he says.
“The Impact of AI-Generated Text on the Internet” study itself suggests that, as the technology develops, detection methods for AI and text watermarking (the embedding of imperceptible markers in AI-generated text) may become inadequate.
Instead, the study suggests using digital verification systems such as C2PA (Coalition for Content Provenance and Authenticity), which embed a tamper-resistant cryptographic credential in a file, such as a photo, to verify that it is human-created.
This credential would include details such as who created the original photo or document, how it was edited, what tools were used, and whether AI was involved in its creation, with each change or save recorded to create a traceable chain of edits.
The researchers also suggest changes to search engines and recommendation algorithms to reward content containing diverse human-made material instead of raw volume or engagement.
Meanwhile, Azree says that the technology itself is not going to kill the Internet, though humans will need to live alongside it and be aware of both the benefits and the risks involved.
“People will still post, create, argue, investigate, and share experiences – but there is a real concern that the balance could shift toward a world where synthetic content overwhelms original human thought, making the Internet feel larger, faster, and more active, yet sometimes less meaningful and less diverse intellectually,” he says.
