AI is reading and writing the web. What does this mean for digital trust?

When a transformational invention comes along, people look for new words to describe it. Sometimes a brand name does the job. The most obvious example in recent years is, of course, ‘to Google’.

Is Pangram about to become a verb too? 

For those of you unfamiliar with the term, Pangram is an AI writing detection programme. You simply paste your text into the Pangram website to see whether all, some or none of it has been written by a machine.

AI detection is increasingly popular among web users. So can it be long before people start saying: “hang on a minute, I’ll just Pangram that?”

Time will tell. But what seems certain is that AI detection will continue to flourish. Why? Because AI content is swarming the web. 

A 2026 study from Pew Research gives a sense of scale.  It analysed half of a million English-language webpages and found significant signs of AI authorship in 10 percent of all pages sampled.  But among pages published after ChatGPT’s release, the figure rose to 35 percent.

AI is changing what we mean by “the web”

The use of AI writing is growing fastest in the social and business domains. This reflects the changing nature of information on the internet. Over the last 20 years, the web has evolved from something we primarily read into something we collectively produce. The online domain is now filled with social posts, product descriptions, reviews, videos and more. 

Perhaps unsurprisingly, non-professional writers are especially open to AI assistance. Indeed, Pangram’s own research found that, of all posts flagged as AI-generated, two-thirds came from LinkedIn. It also revealed that nearly half of X/Twitter articles contain AI writing.

Does this matter? Some observers are relaxed. They argue that AI is to writing what the calculator is to maths. It can suggest angles, provide structure and add polish. 

Lack of nuance and hallucinations

This is a valid argument. But even the optimists recognise the current limitations of AI writing. Large languages model (LLM) writing can lack nuance, invent ficticious citations and occasionally hallucinate. Earlier this year, TikTok scaled back its AI content generation when its software described dancer Charli D'Amelio as a "collection of various blueberries with different toppings.”

Even if the models improve, people still fear that AI writing will diminish human abilities. Linguist Naomi S. Baron explores these ideas in her book: Who Wrote This? How AI and the Lure of Efficiency Threaten Human Writing. She gives the example of AI translation. It’s useful now, but what if it stops people learning languages? Baron also points out the risk of using GenAI to polish a human-authored article. Again, this can be helpful. But only if the writer absorbs the improvements. “I’m enthusiastic when AI becomes a pedagogical tool but less so when it simply mops up after the writer, with no lessons learned,” she told Scientific American.

The coming trust crisis

Pessimists believe the ‘outsourcing’ of writing to AI is well under way. There is a growing concern that this could erode trust – and that the problem could get worse. Why? Because, in a world of machine-generated content, AI models might increasingly end up learning from AI-generated content.

Consider the changes introduced by Google. In 2026, it launched the biggest upgrade to its Search box in 25 years. The new version offers AI-powered suggestions that go way beyond autocomplete. It also lets users create and manage AI agents that can summarise articles, compare products and recommend what to read or buy. 

These useful changes look set to accelerate the spread of AI-generated content. But could this lead to a future in which machine-generated content feeds machine-generated answers?

Researchers believe it could. And they think this could cause a worst-case scenario called “model collapse.” To test their thesis, researchers trained an AI model on AI-generated material about the 14th century English church steeples. By the ninth generation, the output had drifted into lists about the colour of jack-tailed rabbits.

Regulation and the road to transparency

So, what can be done? The most obvious short-term fix is to give users the ability to spot AI-generated content. This is already happening. 

In August 2026, Anthropic announced that all Claude models now embed an invisible watermark into their text output. The text still reads normally to humans, but those with a software key can detect patterns that show it came from Claude. Meanwhile Google developed a watermark system called SynthID, but chose to publish it as an open source tool for developers working on their own models. OpenAI built a watermarking tech for Chat GPT in 2024, but has so far chosen not to release it.

All the AI companies are considering their strategies – because of regulation. In order to trade in Europe, LLMs have to comply with the EU's AI Act’s Article 50. It mandates that generative AI providers must “add machine-readable marks to enable the detection of AI-generated or manipulated content.” 

Article 50 does not specify a technique. It offers watermarking as one option, alongside metadata, fingerprinting and others. Regrettably, none of these methods are foolproof. AI detection carries the risk of false positives, which can ruin an honest writer’s reputation. Then there’s the grey area between an entirely “AI-generated” article and an “AI-assisted” one. Should the latter be tagged or not?

Besides, there are now AI “humaniser” services to factor in too. Users can paste AI text into these programmes to remove the markers of AI authorship and replace them with more human phrasing. The arms race has already begun.

AI-generated content is not going away. Nor should it. For millions of web users, it makes information faster to create and easier to access. So the challenge is not to keep AI out of the information ecosystem. It is to build transparency and provenance around it. Because as AI writes the web and reads it for us, digital trust will depend on something more fundamental than detecting AI: knowing the origin, context and integrity of the information we consume.

Can you trust what you’re reading?  Eight questions to ask

• Can I identify the author?
• Does the author have relevant expertise?
• Does the article include personal anecdotes or experiences?
• Does it provide sources for its claims?
• Does it contain original information and insights?
• Is there a clear narrative flow?
• Is it just filling an SEO need?
• Would I trust the information without knowing whether AI was involved?