AI-Labelled Content – Do Your Visitors Even Care?

AI Labels Are Becoming Mandatory – But How Does That Affect Readers?

In August 2026 some pretty dry-sounding legislation came into force. It’s boring AF to read, but vital if you’re in digital marketing. Let’s get through that, then discuss if your readers even care.

EU

California

China

Well, like many things, they were way ahead of us:

Studies Into AI-Labelling

OK so we know the law (for now – let’s be honest, this is the tip of the iceberg). What I’ve been interested in is whether this labelling of AI content makes any difference in behaviour and thoughts of website users.

Well, back in 2024, Sacha Altay and Fabrizio Gilardi did a study of around 5,000 people in the UK and USA. They showed participants news headlines, some of which were labelled as AI-generated.

They asked the participants to rate how accurate they thought the article was, and how likely they’d be to share it.

You won’t be surprised (neither was I) that the headlines labelled as AI-generated scored worse for both.

So is that it? People just get put off by AI-labelled content?

Nope (otherwise I wouldn’t write this article).

People scored AI-labelled content badly even if it was actually human-written content.

So it was the disclosure – not whether or not the content was actually produced by AI – that counted.

That Study Wasn’t A One-Off

If we go back further to 2022 and a study from Chiara Longoni et al, they used 4,000 participants across two experiments, showing them news article headlines (all of which were genuine news headlines tagged as either AI or human written.

Here’s what they found:

• AI tag reduced perceived accuracy by 7.6 percentage points (experiment one) and 14.5pp (experiment two)
• Trust in the reporter dropped sharply: 2.57 vs 3.30 on a 5-point scale for AI vs human

But remember this was back in 2022 – pre-ChatGPT and therefore at a time where most of us weren’t even used to seeing ‘AI’ everywhere like we do now.

Another But: Both Of These Used Opt-In Panels

Why this is important: If you recruit people who volunteer for online surveys, you’re not necessarily getting a picture of the general public.

So in 2024, Chuyao Wang, Patrick Sturgis and Daniel de Kadt at the LSE ran it again, this time on 3,861 UK adults from a proper nationally representative probability sample.

They showed people a news article about Universal Basic Income, made to look like a BBC News page. Half saw a disclaimer at the bottom saying the report was generated by ChatGPT. Everything else was identical.

The penalty showed up again. Labelled article, lower perceived accuracy.

So far, so predictable. But this study asked the question nobody else had: does it actually go anywhere?

They measured three things beyond accuracy:

• Interest in the topic (dropped slightly)
• Support for the policy (no effect at all)
• General concern about misinformation (no effect at all)

People rated the labelled article as less accurate. Then went on to support UBI at exactly the same rate as everyone else.

The label changed what they said about the article. It didn’t change what they thought about the subject.

And it didn’t make them more suspicious of online content generally, which is notable, because “this is misinformation” labels do exactly that.

Half the participants were shown a short explainer about generative AI before they read anything — just a neutral paragraph on what it is and how news outlets use it.

The researchers expected this to make the label sting more. It did the opposite. Priming people about AI reduced the damage the label did.

Their explanation: the penalty is partly a novelty shock. Over half their sample had heard little or nothing about generative AI. Familiarise people with it first, and the label loses some of its bite.

Which raises an obvious question: if the penalty shrinks as people get used to AI, how long does the penalty last?

What People Actually Hear When You Say ‘AI-Generated’

Going back to Altay & Gilardi, they dug into why people marked the labelled content down. It came down to this: people took ‘AI-generated’ to mean entirely AI-generated. No human anywhere near it.

And look at how Longoni’s study set things up. Before participants rated a single headline, they were told AI reporters were algorithmic processes with “limited to no human intervention.” So the assumption wasn’t something people turned up with — the researchers handed it to them.

Which fits with the novelty thing from the LSE study. If you don’t know how AI content actually gets made, you fill in the blanks. And people fill them in badly.

So when you slap ‘AI-generated’ on something, what a lot of readers hear is ‘nobody was responsible for this’.

Now remember that EU carve-out from earlier: AI text is exempt if a human reviews it and takes legal responsibility. Lawyers drafting legislation and psychologists running experiments, working completely separately, landed on the same thing: it’s about whether a human is accountable, not which tool did the typing.

So I think the answer isn’t to hide the label, it’s to put the human back into it.

How You Do It vs Whether You Do It

Research shows that publishers can mitigate the loss of trust by not just using a bare ‘this was AI’ label and specifying that there was decent human oversight. So – what AI did and what the human did.

Specifically, generating content v editing. Perception of content generated by AI is worse than content edited by AI.

I’d also call out a specific point from the LSE study – which is that they used ‘generated by ChatGPT’ rather than ‘generated by AI’. The potential issue with that is it tested people’s perceptions specifically about that platform (which, let’s be honest, is often influenced by how Sam Altman is perceived).

Paige Maguire told me: “Every AI disclosure I’ve seen was drafted by legal and shipped by engineering. Nobody designed it. That’s why they read like allergy warnings. A disclosure is interface copy, it has placement, hierarchy, and tone, and all three change how it lands. If you treat labelling as a compliance task, you will get a compliance outcome, which is the worst-performing version of a thing you’re legally required to do anyway.“

If You Don’t Need To Label, Should You?

So a situation where a publisher says “this was written with AI, reviewed and edited by [X person”, this would reduce AI’s negative impact on visitor perception. But it would also mean the label isn’t even needed, as that kind of content is exempt from the EU legislation.

Which begs the question, should publishers label content if they legally don’t need to?

I don’t think there’s a right or wrong answer to this. My personal opinion would be no. My rationale:

• Visitors will see it as negative
• How do we determine a threshold for whether a label is warranted (e.g. a human writes a piece then uses AI to proofread it)

On the flip-side, you could argue that not labelling content as AI-involved implicitly says ‘this was all done by a human’.

And there is a risk that AI detection tools (and personally I think many of these are BS) flag yours as AI generated. So, possibly there is an argument for proactive labelling even if there’s no legal requirement.

Jake Wardle from EV Cable Hub told me:

What People Say Isn’t Always How They Behave

A caveat with the studies we looked at is – what people say isn’t always how they actually behave (especially if they feel they should be saying a certain thing). AI can be quite an emotive and controversial topic and I’d bet if you asked someone at random “would you trust AI labelled content less” they’d instinctively say no.

And here’s where as publishers we have a challenge – how can we measure impact of AI labelling on trust and perceived credibility? I’m not convinced we can.

The first challenge is – yes, we can split test labelling and measure metrics like scroll depth, time on page, engagement etc. But none of those really tell us if a visitor felt trust in the article. Someone could read a whole article top to bottom and think “that was great” or “that looked like AI slop” and we wouldn’t know.

My colleague Paige says:

The second one is with legislation. If we’re talking about content that falls under legal requirements, we can’t split test removing labelling whilst being compliant.

My gut feeling on it all:

• It depends (sorry!) on the topic, situation and niche
  • Someone reading a summary of a stock movement is unlikely to care it’s AI-written (most are)
  • Someone reading an in-depth guide to symptoms of a serious disease would care
• If the content truly satisfies the user intent, is good quality etc. then I think most readers aren’t going to be bothered in most cases

As Chris Selland says, and I think this sums it up very well:

What Happens Next

My theory, completely unsupported by any studies is that the effect of AI-labelling will reduce quite rapidly over the next couple of years due to the snowball effect of:

• Legislation means more labelling = people become more used to it
• AI tech gets more ingrained in our lives = people become more used to it
• AI content becomes more of a norm = people become more used to it

I say ‘reduce’ rather than ‘stop’ because I can’t see a world where anyone wouldn’t trust AI as much as humans.

As Davide Bertolino from Press And Pillow told me:

That sums it up nicely. I think we will move away from so much ‘AI slop’ being published as it’s so often ridiculed. I also like the point of who is prepared to put their name behind it; and again I’d assume most people wouldn’t want to ‘claim’ AI slop.

Similarly, Callum Gracie from Otto Media told me:

That Was All About Humans Though..

All the studies were about how humans viewed AI-labelled content. What I’m really curious about though, especially for digital marketers dealing with the rise of agents – do AI systems view AI-content as less trustworthy?

I don’t know.

Things to consider here:

• If an AI search platform crawls content it thinks was AI-generated, isn’t labelled as AI but should be by law, would it (or should it) see this as a trust issue?
• So, almost an extension of EEAT – the Trustworthiness could then include ‘did this publisher label its AI content’?
• But – systems that decide whether something was AI-written aren’t reliable, and what if there was a false positive?

I lean towards thinking it’s not going to be an issue, but given how quickly things move, who knows?

Google Doesn’t Mind AI Content

I would say this though – Google doesn’t give a shit if your content is AI-written. It does care about ‘scaled content abuse’ and content depth. But this myth that Google penalises your site because you use AI content is nonsense.

OK Andy, So What Should We Actually Do?

With the caveat that I’m not a lawyer and this isn’t legal advice:

• Assess if you are legally obliged to be labelling
• If you are, label it
• If you label, do it in a way that clearly explains what was AI and what was human and design it in a nice way
• Don’t panic whether it’s going to affect trust but don’t ignore it either

To address this head-on as publishers, we also need to leave our own biases at the door. By that I mean that if we take a negative approach to AI-labelling, it’s going to show. If we end up with AI-labelling that’s in red text with ❗emojis and the whole think looking like a warning, it’s going to fail.

If we take the approach of ‘here’s our opportunity to be candid and positive’ then that will come across to the visitors.

Bottom line: We need to consider legislation, integrity, visitor perception and future changes. So, no pressure then 😉

Appendix: Examples Of AI Content Labelling

Yahoo

CNET

BuzzFeed

Washington Times

Citations/Links

• People are skeptical of headlines labeled as AI-generated, even if true or human-made, because they assume full AI automation (Altay & Gilardi)
• AI labeling reduces the perceived accuracy of online
  content but has limited broader effects (Wang et al)
• Effect of disclosing AI-generated content on prosocial advertising evaluation (Baek et al)
• I Don’t Care Who Wrote This (Chris Selland)
• More than a third of the internet is now being written with AI, study suggests (Yahoo Finance)
• Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems (EU Artificial Intelligence Act)
• California’s AI Transparency Act (SB 942) Is Now in Effect — Not January 1 (Casrai.org)