Eighty per cent of communications professionals use generative AI. Eighty-five per cent have never been trained to use it. The distance between those two numbers is now a governance problem.
Generative AI has moved into professional life faster than any workplace technology in memory, and it has arrived from every direction at once: how we write, how we search, how we get found. In research I co-led with the University of Sussex, 80% of content writers told us they use generative text AI tools. Only 11% said they never used them. Set those figures against two others from the same study: 85% of respondents had never received any training in using these tools, and 71% said their organisation had no guidelines on acceptable use, or none they were aware of. That distance between use and understanding is the AI literacy gap. It is quieter than the debates about job losses or brand discoverability, but for knowledge industries it may be the more consequential problem: unmanaged adoption is already shaping the quality, trustworthiness and accountability of everyday work. Our sample was communications professionals, but the pattern will be recognisable to anyone who drafts a client email, briefs a team or represents an organisation externally. Which is to say, most people in most jobs.
This is not a story about idle experimentation, though. In our survey of more than 1,100 practitioners, 68% of respondents said the tools improve their productivity, and most believed AI will ultimately have a positive impact on their work. Optimism aside, the technology is already doing work inside organisations that have, in the main, never formally sanctioned it.
The technology is also reshaping how organisations get found, from the outside in. Google now presents an AI overview above the results, offering an answer before anyone needs to bother scanning the blue links. Increasingly, people skip the search engine altogether and ask the likes of Claude or ChatGPT instead. Research from Search Forward, a Magenta series on the changing search landscape, found two-thirds of B2B decision-makers now use AI to find suppliers, 85% had discovered a supplier through an LLM they would not have found otherwise, and 90% trust AI-generated answers as credible. LLM answers to search enquiries are synthesised, often without attribution, from whatever the model has absorbed about an organisation: its coverage, its reviews, its website copy, its social presence, and increasingly the sheer volume of AI-generated content published in its name. Whoever communicates on behalf of an organisation is, whether they’ve registered it or not, now writing for two audiences: the human reader, and the machine sitting in front of them. There’s more secrecy to search now. Nobody fully understands the algorithms that decide what surfaces and what stays hidden.
CheatGPT?
And there’s secrecy behind-the-scenes, too. Only 11% of communicators talked openly to clients or external networks about their use of generative AI in their work. When we asked why, respondents feared being judged as lazy or less capable. Others described a sense of embarrassment at needing help in the first place. Hence the study’s name – CheatGPT. Secrecy of this kind is a side effect of the way AI is being managed, or rather, not being managed. When the organisations people work for fail to set norms, policies or ethical use guidelines, individuals improvise their own, and improvised rules make inconsistent, potentially irresponsible use the default. It’s also difficult to expect anyone to know what responsible use even looks like when they have not been told. If public relations is, in the PRCA’s definition, the “strategic management discipline that builds trust, enhances reputation and helps leaders interpret complexity…”, then someone is on the receiving end of it. You. And if AI is being used in the shadows, without the understanding that should inform its use, then what could that be doing to our information environment, to information integrity, to trust?
Every fabricated statistic, unlabelled AI-generated image or unchecked claim that goes out under an organisation’s name adds a little more noise to an information environment already struggling with synthetic content and coordinated inauthentic activity. The EU AI Act’s transparency obligations under Article 50 exist because audiences can no longer reliably distinguish AI-generated content from human-made content unless someone tells them. Communicators, of all people, understand what happens to trust once it slips away. It doesn’t come back. If the function most responsible for protecting an organisation’s credibility is also the one experimenting most heavily, and most privately, with tools that can degrade that trust, then AI literacy becomes a trust and integrity ‘must have’.
Theory v. practical
Part of the problem is that ‘AI training’ is usually understood as tool training: which buttons to press, which prompts to write. There is a difference between skills for use and skills for understanding. Employers tend to invest in the former. But employees need the latter. The OECD describes AI literacy as the knowledge, skills and competencies needed to master and critically assess AI, to work with it effectively and ethically. This is particularly relevant for anyone who communicates as part of their role, because the currency of that work is trust. A tool that fabricates a statistic burns credibility that took years to build. Actually, no. Let me rephrase. A person who doesn’t use that tool properly, doesn’t creatively or critically review the output, notice the fabrication – they burn credibility.
A useful way to think about literacy is as a spectrum with three stages. Illiterate is the passenger stage: you can steer the tool and accept its output, but you have no idea what is happening under the bonnet, so when it goes wrong you cannot tell. Literacy is the pilot stage: you still use the automation, but you scan the instruments, understand the limitations and know when to distrust the machine. It includes knowing when a task demands judgement, cultural sensitivity or accountability that should not be delegated to a model. Fluency is the stage beyond that, where professionals understand what goes into designing the engine. Reaching the pilot stage does not require a computer science or engineering degree. It requires understanding a few core facts, chief among them that a large language model predicts plausible language based on patterns in its training data. It calculates probability rather than comprehending truth, which is why a fluent, confident answer can be wrong. Throw in a bit of sycophancy – your LLM of choice telling you how amazing you are, how every idea you have is brilliant – and it’s easy to see how an unaware user could mistake flattery for accuracy.
AI literacy rests on four pillars. The first is understanding capabilities and limitations: knowing what these systems actually do, where they are strong, where they are weak, why they tell you sweet little lies from time to time. Demystifying the technology removes both the hype and the fear that distort decision-making. It may also reduce the fear of missing out, which is clearly driving a good deal of AI adoption. The second is critical thinking. Every output should be treated as an unverified first draft, and no fact, quotation or statistic should ever be published without checking the source. The third is responsible use: guarding against bias in outputs, protecting confidential information and respecting privacy. In our research, 50% of managers said they were concerned about the legalities of generative AI use. That figure should be 100% in my view. The fourth pillar is keeping a human in the loop. The Global Alliance’s Responsible AI Guiding Principles, co-created by professional bodies worldwide, are clear that accountability for AI-assisted work sits with the professional, not the tool.
That principle has already been tested at a tribunal, and it did not go the AI’s way. In 2022, a passenger asked Air Canada’s website chatbot about bereavement fares after a death in the family, and the bot told him he could book at full price and claim the discount back afterwards. That wasn’t the airline’s policy; the correct terms sat on a different page of the same site. But the passenger had no way of knowing which part of Air Canada’s website to trust, and he booked on the chatbot’s advice. When he tried to claim the difference back, Air Canada refused. Its defence: the chatbot was “a separate legal entity”. They were not responsible. The AI was. A tribunal disagreed. It ruled that it was all the company’s website, whoever or whatever was doing the talking, and ordered Air Canada to pay up. The lesson generalises well beyond airlines. Don’t blame the AI when it goes wrong.
Three moves
The encouraging news is that practitioners want to understand these systems. In our study, 66% said workplace training on generative AI would be useful. Three moves make the biggest difference. First, write AI guidelines collaboratively, with input from the people who use the tools daily, as rules imposed from above tend to drive use underground rather than improve it. Second, invest in training that builds understanding. A one-off webinar on prompting is not literacy. Literacy is a habit: formed at the start, then reinforced so it holds. On the AI literacy programme I co-lead with Dr Tanya Kant at the University of Sussex, the shift tends to come at the point where someone stops asking for a better prompt and starts asking why the model gave them the answer it did. Third, establish use and disclosure norms early, internally and externally, so that transparency becomes part of doing business. It will feel less like a confession that way.
Generative AI will keep improving, and the tools of 2026 already make those of 2023 look primitive. What will not change is the gap between understanding how to use the tools, and understanding how the tools work. Unless there is more investment in AI literacy. For any organisation whose currency is trust – which, in the end, is most of them, surely? – closing that gap is a professional obligation, and the organisations that treat it that way will be the ones still trusted when the novelty has worn off.
About the author
Jo Sutherland is managing director of Magenta Associates, a UK communications consultancy. She is studying for an MSt in AI, Ethics and Society at the University of Cambridge and is a contributing author of AI for PR (Kogan Page, 2026). With Dr Tanya Kant at the University of Sussex, she co-led the CheatGPT research into generative AI use in the UK PR and communications profession, and co-leads the AI Literacy for Communicators CPD programme.
