“When talking about journalism in the newsroom, there is too much focus on the technology and not enough on what it does,” said Marcela Kunova, the editor of Journalism.co.uk.

I agree.

Professor on Innovation Tor Wallin Andreassen at The Norwegian School of Economics (NHH) recently noted: “When Big Tech is massively leading the AI race, we need to shift focus from developing ground-breaking AI technology to becoming leaders in the use of AI.”

I agree with that as well.

If you go to the hotel for Artificial Intelligence models, Hugging Face, you will find more than half a million different models. And there are way over 100,000 datasets you can use to train these models.

Before large media groups manage to fire up their cloud Nvidia GPUs, some of the larger players have rolled out an even better large language model (LLM) that can swallow longer texts, handle more context, run even faster and lighter, and understand more languages.

Being allergic to commercial vendors of AI worked for a while in 2023 when GPT3 was alone in the space. Now you can connect your newspaper to state-of-the-art language models including Anthropic (even better sentences than GPT4 in many cases), Gemini (faster and more scalable, better to transform code), Microsoft Azure (trusted by your IT manager, easier), or the refreshing non-Californian Mistral.

Our experience is that all of these models (and many more) are surprisingly affordable to put to work in a breaking news editorial environment.

So the game will change. Development efforts will be pushed upward in the newspaper’s hierarchy of needs. You don’t code your model using open source tools; that’s like making your tofu. You can, but it’s not worth it and it will taste the same as the one in the shop or worse.

Bring in your preferred model, fine tune it if you must — though you probably won’t need to — and start experimenting with your editors.

Work downstream to implement AI in the newsroom:

Create tools that help reporters in their daily work: Make your AI create captions, help with SEO and some distribution, classify stories, tag them, and add boring meta-data. Journalists are terrible at meta-data.

Don’t get in the way of the hand-crafted text: Surprisingly few reporters want to auto-generate titles — it’s too close to home. Reporters want to do this manually. Also, do not make a text longer using AI. Use it to extract and make text shorter. It excels in that.

Advanced functions should look easy: If Google manages to navigate the entire Web with one little search box, how many options do you need to provide in your AI tools? Make it simple. We are in a hurry, and the news can’t wait. Hide the advanced settings (we could be better as well here).

Create auto summaries: They work well and increase readability and read time for some of your users.

Be international: Language barriers are vanishing with LLMs. We see several smaller publishers growing fast with new approaches to language. A small niche site in a local language can suddenly be world-class, niche-generated content now and will be removed from the search index. You will kill your traffic and your trust.

Do not auto-generate articles: Just don’t do it. You became a publisher for a reason. Don’t waste your trust. Google is increasing its penalties against auto-generated content. You will be removed from the search indexes and you will not add value to your readers.

Use AI-native production tools: Make sure your most important editorial tools use machine learning and AI to make your company more efficient. Choose a CMS that has AI built in and encourages you develop it further (a glass house here, I know). Choose a paywall provider that adapts to different reader groups, and choose a data layer (stats) that can play with the paywall and the editorial content to make new products fast. Make sure your tech stack can launch new products in days, not weeks.

Don’t hire AI experts: Make everybody an AI expert, both in your newsroom and in your development department. You need experts in practical use, not in fundamental machine learning research.

Use the tools that exist: Don’t build unless you need to. The world is changing faster.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .