AI in practice: Hearst UK

Q: What is the current state of play with regards to AI usage?

A: The advent of AI – and generative AI in particular – has caused the publishing industry to enter a period of rapid change, competing enthusiasms and no small amount of uncertainty. While some companies have launched themselves into evangelical adoption, others have been more circumspect, their curiosity tempered by legal blockers, ethical uncertainties and expense.

The fear of being left behind is often countered by a fear of losing our grasp on the very things that give journalism and publishing its heart and soul. Consequently, our collective adventures in artificial intelligence can resemble people moving at hazardous speed but going nowhere fast.

This is understandable. But as we assess our own position on AI, we risk failing to concentrate on what our audiences think, feel and want from our content and services. The debate about how AI should be used in journalism, image generation and production software is important. But it should sit alongside another, arguably more important, question. To wit: what is the public’s appetite – or tolerance – for content to be AI-influenced? (My admittedly small and anecdotal pool of research would suggest it is low.) This is one of the biggest questions publishers should be asking.

Hearst UK is at an important transition point. We are categorically pro-AI but are still working out how to move from isolated experiments to repeatable, properly governed practice.

Our people have access to an enterprise version of ChatGPT and are trained on how to use it properly and ethically, as well as building their own custom GPTs and bots. We use AI in multiple departments, ranging from auto-tagging in our media systems, to text-based video editing, to the prompted creation of marketing assets.

Possibly the most important thing we’ve done is to develop an editorial AI policy. It recognises that AI can assist journalists but must not take their place. It sets out principles around human-generated content, trust, transparency, accuracy, oversight and innovation, alongside guidance on appropriate and impermissible uses.

The policy is deliberately principles-based, because the technology is changing too quickly for any list of permitted uses to remain sufficient for long. Its central message is simple: AI can be a great support in creating content, but it doesn’t replace the journalist’s responsibility for it. It must not fabricate reporting, substitute for genuine expertise or lived experience, or present machine-generated material as original journalism.

Q: What are the best opportunities for future development?

A: The greatest opportunity is to redesign workflows rather than simply add AI to existing ones. Giving people new tools while leaving old processes intact is a false economy: it simply adds another layer of activity without changing how work gets done.

AI needs to be embodied in the operating model of publishing, in the way we commission, create, edit, package, distribute and learn from content. That’s ultimately what will allow us to improve speed, scale and efficiency while giving the human element more room to do the work that creates real distinction: originality, creativity, ingenuity, judgement and taste.

The most interesting opportunity may be to make our abundance of content more useful to individuals. Our archive contains decades of high-quality journalism, but AI could help us surface, connect and repackage the material most relevant to each customer, powering more intelligent search, recommendations, recirculation and membership services. Personalisation doesn’t mean putting a chatbot on every homepage; it’s about helping audiences find greater value in what we create.

Q: What are your three key learnings?

1. Start with the problem, not the tool: The arrival of a new model can create a powerful sense of urgency, but the best use cases usually begin with a frustrating workflow or unmet audience need. The question should be: what are we trying to improve and why would anyone care?
2. Human oversight is not a handbrake on innovation: With quality journalism, trust is the product. Human accountability can’t just be added at the end as a compliance check; it must be designed into the process from the beginning. AI might be able to knock out something quickly, but speed isn’t the same as value and fluency isn’t the same as truth.
3. Training must lead to implementation: It is easy to give people a new tool, run an inspiring session and declare progress. Much harder is to identify a real use case, change the operating model around it, measure the result and decide whether it should become standard practice. The shift we need is from awareness to adoption and from experimentation to responsible scale.

Full disclosure: I asked ChatGPT to review an early draft of this piece. It offered some useful suggestions, several bloodless ones and, somewhat confoundingly, assured me that it was within the designated word count when it was anything but. That is why we must resist anthropomorphising AI. When it makes mistakes, it’s not a colleague having an off day; it is a technology malfunctioning. We shouldn’t excuse a confident error simply because it’s delivered in an obsequious manner. To paraphrase the late, great Terry Jones, AI is not the Messiah, it’s a very naughty bot.