What does Britain really know about AI? What the 2026 evidence shows

Six in ten adults in the UK used generative AI in the last three months. Only about one in seven say they could explain what AI is in any detail. I went through the biggest surveys of 2025 and 2026, including a government study of more than 30,000 people, to find out what Britain really knows about AI, where the gaps are, and why they matter.

• 59% of UK adults used generative AI in the three months before they were asked¹
• 15% say they could explain what AI is in detail, and 7% of over-55s²
• 8.7m adults say they know hardly anything about AI, or have never heard of it (my estimate from government survey data)¹
• 10% feel very confident they can tell AI-made content from content made by a person³

1. Use has raced ahead of understanding

The Department for Science, Innovation and Technology (DSIT) asked 30,663 adults across the UK, between late 2025 and early 2026, how much they know about AI. It's one of the largest surveys of its kind, and it's weighted to represent the whole adult population.¹

Put the bottom three answers together and 42% of adults, roughly 23 million people, say they know a little about AI or less. Yet 59% had used a generative AI tool, such as ChatGPT, Copilot or Gemini or an AI image tool, in the three months before they answered.¹ Ofcom's survey, which asks the question differently, puts current use of AI tools at 54% of adults and 79% of 16 to 24-year-olds.³

A tougher question gives a starker answer. In DSIT's tracking survey, only 15% of adults said they could explain what AI is in detail. Among over-55s it was 7%, and a third of over-55s (34%) had heard of AI but couldn't explain it at all.² Seven in ten adults (71%) said they have little or no knowledge of how AI is trained on data.² Asked for one word to describe how they feel about AI, the most common answer was "scary".²

There's nothing unusual about using something you can't explain. Most of us couldn't explain how a fridge works. But AI is different in two ways that matter: it can be confidently wrong, and the same technology can be used to deceive you. A little understanding is what protects you from both.

The gap shows among regular users too. Of the generative AI users who answered online, a quarter (25%) said they were still learning the basics and one in eight (13%) had only tried it a few times.¹ At work, KPMG and the University of Melbourne found that 58% of UK workers had relied on AI output without checking whether it was accurate, and only 27% of UK adults had had any AI training at all, which puts the UK in the bottom third of the 47 countries studied.⁴

2. Most of us meet AI without choosing to

Asking whether someone "uses AI" misses most of the AI people actually meet. Search for almost anything on Google and an AI-written summary often appears above the links. Three-quarters of online adults (75%) say they read these summaries at least some of the time, and 42% often or always. More than half of people who never use an AI chatbot (54%) still read them.³ Voice assistants such as Siri and Alexa are AI too, and 40% of adults used one in the past week.¹

Yet when the Office for National Statistics asked people in June 2026 whether they had used AI in the past year, a third (33%) said they had not knowingly done so, including two-thirds (66%) of over-70s.⁵ These are different surveys, so the figures can't be compared exactly, but the message is clear: many people are using AI without counting it as AI. Nearly half of adults (47%) say they've never seen a label or icon telling them something online was made or edited with AI, even though 85% think platforms should tell them.⁶

So AI know-how isn't only for people who choose to open ChatGPT. It's for everyone who searches the web, scrolls a social media feed or answers the phone.

3. Confidence isn't the same as ability

Ofcom asked people how confident they feel telling whether something online was made by a person or by AI. 44% said they feel confident, 24% said they don't, and only 10% feel very confident. Ofcom itself warns that confidence "does not necessarily translate into ability".³ The research on what people can actually spot backs that up:

• Faces. In a 2023 study, AI-generated white faces were judged to be human 66% of the time, more often than real human faces (51%). The people who made the most mistakes were the most confident.⁷
• Voices. Researchers at Queen Mary University of London made voice clones from a few minutes of recordings. Listeners judged the clones human 58% of the time, about as often as they judged real voices human.⁸
• Conversation. In a 2025 Turing test study in the US, people chatted for five minutes with a human and an AI at the same time. Told to play a human-like character, GPT-4.5 was picked as the human 73% of the time, more often than the real person.⁹
• Video. When Which? tested 514 UK adults in March 2026, 70% failed to correctly sort all the real and fake videos they were shown.¹⁰

There is good news: practice helps. In a study led by the University of Reading, about five minutes of training raised ordinary people's success at spotting AI faces from below chance to around 51%.¹¹ That's a big improvement, but it still means guessing wrong half the time. Spotting fakes by eye will only get harder.

The surprising part: overconfidence is a younger person's problem

The stereotype is that older people are the easy targets. Ofcom's realistic tests point the other way. Shown a fake parcel-delivery email, 86% of 45 to 74-year-olds chose only safe responses, compared with 76% of 16 to 34-year-olds.³ Shown a Google search, 55 to 64-year-olds were the best at recognising the top results as paid adverts (58% got it right), while nearly half of 25 to 34-year-olds (47%) were confident but wrong.³

Older adults aren't immune. They're more likely to say they're unsure, and when they are caught the losses can be large. But their caution is an asset. The worry is that the clues careful people rely on are disappearing. In Ofcom's 2026 survey on suspicious texts, 69% of adults said they judge whether a message is genuine partly by "the spelling, grammar or tone", rising to 77% of 55 to 64-year-olds.¹² That's exactly the clue that AI-written scams no longer give away. I've written about what that means for scam advice in my own research paper.

4. The age gap is real, but it isn't about interest

Use falls steeply with age, from around eight in ten adults under 45 to three in ten 65 to 74-year-olds and fewer than one in five 75 to 84-year-olds.¹ Knowledge follows the same curve. Among 75 to 84-year-olds, 42% say they know hardly anything about AI or have never heard of it; among those aged 85 and over it's 68%. By my estimate that's close to 5 million people aged 65 and over.¹

But ask the people who don't use AI why not, and lack of interest isn't the main answer. Among non-users aged 75 to 84, the most common reason was "I don't know enough about it" (47%), well ahead of "I don't find it useful or interesting" (27%).¹

That matters, because knowledge is the easiest barrier to fix. Older people's own optimism about AI is low, though. In the ONS survey, only 16% of over-70s agreed that "AI will benefit me", against 47% of 16 to 29s.⁵ One surprise: over-70s were the age group most willing to trust AI to help with medical diagnosis or treatment decisions (29%, against 13% of 16 to 29s).⁵

How do older people learn? Mostly from the people around them. In the last year, 29% of 75 to 84-year-olds picked up new digital skills from family and friends, compared with 1% through training at work.¹³ And among 65 to 74-year-olds who do use generative AI, nearly a third (31%) say their experience is very limited and they've only tried it a few times.¹

5. Four common misunderstandings

"It looks the answer up"

A chatbot writes its answer by predicting which words are likely to come next, based on patterns in what it was trained on. Some chatbots also search the web, but even then the answer is written by the AI and can still mix things up. When 22 public broadcasters, including the BBC, tested AI assistants on more than 3,000 news questions in 2025, 45% of the answers had at least one significant problem, such as wrong facts or sources that didn't support the claim.¹⁴

"If it sounds sure of itself, it's right"

AI writes just as fluently when it's wrong as when it's right. People sense the risk: 74% of adults name inaccurate information as a risk of AI, the most common concern of all.¹ Acting on it is harder. In research for the BBC, 64% instinctively agreed it's important to watch out for errors in AI summaries, but only 38% said they would actually question one.¹⁵ One older participant summed up the assumption: "you'd expect it to always know the answer".¹⁵

"There's someone in there"

Chatbots are designed to sound friendly and human. In a 2024 study of 300 people in the US, two-thirds (67%) thought there was at least some chance that ChatGPT is conscious, and the more people used it, the more likely they were to think so.¹⁶ AI tools don't have feelings or intentions, but they are very good at seeming to, which is part of why AI-powered romance scams work.

"I'd spot a fake"

As section 3 shows, most of us can't reliably, and the people who are surest are often the most wrong. The safer habit is to check the source rather than trust your eyes and ears. If a message, call or video asks for money or personal details, contact the person or company another way, on a number you already have.

6. What worries people, and whether they're right

Concern is close to universal: only 2.7% of adults said they had none of these worries, and 88% are concerned about harmful AI-generated content spreading online (57% very concerned).¹ Pessimism is growing too. The share of adults who think AI brings more risks than benefits rose from 25% in August 2024 to 38% in June 2026.⁵ And the UK is unusually sceptical: in the Reuters Institute's 2026 study of 48 countries, Britain had the lowest trust in AI chatbots' answers about news (6%).¹⁷

Are people right to worry? Mostly, yes. Inaccuracy, the top concern, is the problem experts worry about most, and deepfakes are already used in scams, as I describe in my guide to how criminals are using AI. The government's own scientists add some useful perspective: in a 2025 review they found no clear evidence yet that deepfakes are widespread in the UK.¹⁸ Healthy scepticism is good. What helps most is turning that worry into a few practical habits.

7. What good AI know-how looks like

You don't need to know how AI works inside to use it safely. These six points cover most of it:

1. It predicts; it doesn't know. AI is brilliant at drafting, summarising and explaining, but it can be confidently wrong.
2. Check anything that matters, especially health, money, legal and news questions, against a trusted source such as the NHS website, GOV.UK or your bank.
3. Keep private details out. Don't type bank details, passwords or medical records into a chatbot.
4. Seeing and hearing is no longer believing. Voices, faces and messages can be faked. If someone asks for money, check by calling back on a number you already have, and agree a family safe word.
5. You meet AI even if you never open ChatGPT, in search summaries, social media feeds, phone keyboards and photo apps.
6. It's genuinely useful. The most common use among people who use generative AI is writing or editing documents, texts and emails (58%).¹ The best way to understand AI is to try it on something real, with someone alongside.

If you'd like to do that on your own phone, tablet or computer, that's what my AI lessons at home are for. And if you run a club or group, I give a free talk on AI and the new scams. Ring 07500 511462.

How I did this

I read each report and, where they were published, the underlying data tables. Wherever possible I used official and academic sources. Where a figure comes from a company that sells related products, such as deepfake detection, I've said so.

Figures marked "my estimate" turn survey percentages into numbers of people using the survey's own weighting. DSIT says each weighted response represents about 1,846 adults, so its sample stands for about 56.6 million UK adults. Different surveys ask different questions in different ways, so their figures aren't directly comparable; I've said which survey each number comes from. Some questions in the DSIT survey were asked only of people who answered online, and I've noted where that applies.

Every figure was checked against its original source. If you spot an error, please email enquiries@jameshayesconsulting.com and I'll correct it.

Sources

1. Department for Science, Innovation and Technology, DSIT Public Engagement Survey 2025/2026: artificial intelligence data tables. 30,663 UK adults, online and paper. Tables E1, E3, E5, E11, E15, E17, E21 and E27.
2. DSIT, Public attitudes to data and AI: Tracker survey (Wave 4), December 2024. 4,947 UK adults online, July to August 2024. Still the latest wave.
3. Ofcom, Adults' Media Use and Attitudes Report 2026, 2 April 2026. Fieldwork September to November 2025.
4. KPMG and University of Melbourne, Trust, attitudes and use of AI: a global study 2025, country insights (UK pages). UK sample 1,029, fieldwork November 2024 to January 2025. Part-funded by KPMG.
5. Office for National Statistics, Public opinions and social trends, Great Britain: June 2026, 17 July 2026. 3,510 adults.
6. Ofcom, Online Nation 2025, December 2025, citing Ofcom's AI labels survey (December 2024).
7. Miller et al., "AI Hyperrealism: Why AI Faces Are Perceived as More Real Than Human Ones", Psychological Science, 2023.
8. Lavan, Irvine, Rosi and McGettigan, "Voice clones sound realistic but not (yet) hyperrealistic", PLOS One, September 2025.
9. Jones and Bergen, "Large language models pass a standard three-party Turing test", PNAS, 2026. US participants.
10. Which?, "Ultra-realistic deepfakes herald the next generation of scams", May 2026. Deltapoll, 514 UK adults, March 2026.
11. Gray et al., "Training human super-recognizers' detection and discrimination of AI-generated faces", Royal Society Open Science, November 2025. Percentages from the University of Reading's press release.
12. Ofcom, Experiences of suspicious calls, texts and app messages 2026, data tables, March 2026. Online survey by Yonder, 4 to 5 February 2026; table 100.
13. DSIT, DSIT Public Engagement Survey 2025/2026: digital inclusion and skills data tables, table A17.
14. European Broadcasting Union and BBC, News Integrity in AI Assistants, October 2025. Tests of 2025 versions of ChatGPT, Copilot, Gemini and Perplexity.
15. BBC and Ipsos, Audience Use and Perceptions of AI Assistants for News, October 2025. 2,000 UK adults aged 16 to 75.
16. Colombatto and Fleming, "Folk psychological attributions of consciousness to large language models", Neuroscience of Consciousness, 2024. 300 US residents.
17. Reuters Institute, Digital News Report 2026, June 2026. YouGov, about 2,000 adults per market.
18. Government Office for Science, Deepfakes and media literacy, May 2025.