Inside Reuters’ agentic AI video experiment 

“We’ve asked the LLM to pick the best bits and create what we call a wrap edit. And it’s actually doing quite a good job. So that’s without even having that multimodal. So if we can actually get the AI as an agentic AI system to understand that, to be able to look at things and compile things, we might be able to get the AI to build the edits for us, which would be quite extraordinary,” Lang said on stage at the Digiday Publishing Summit Europe on Oct. 28.

To be clear, this AI agent would be making the rough cut. There would still be a human involved to make decisions on the full video edit, Lang said, such as Reuters’ new AI video producer, Enrique Flores Roldan.

But the idea is it would help with the more basic and time-consuming video edits. Ideally, it would be able to cut together clips that take into consideration continuity errors, such as a subject taking their glasses on and off, or the sun going in between the clouds impacting the way a subject is lit, Lang said.

“It’s… writing the instructions so that they’re understood and can be applied [by an AI agent],” he said.

About 60% of Reuters’ newsroom is using AI, according to Lang. The most “proficient” AI users – which make up about 50 to 100 people, out of Reuters’ 2,500 journalists – are using AI to “vibe code” and help with investigative journalism.

“We’re at 60% and it’s rising every month. It’s probably 5% every month at the moment. We’ve been told to get to 100% by the end of the year. I think we’ll get close to 80, touch wood,” Lang said.

The biggest hurdle to getting there is not every newsroom staffer feels like they need to use AI technology, he noted.

“I was talking to a photographer the other day, [who asked] how can I use AI? And I said, why don’t you just use it to help you find a parking space when you turn up to a strange city? Just simple things like that. There are ways to use AI that aren’t necessarily journalistic,” Lang said.

So what’s Reuters’ newsroom not using AI for? Fact-checking, Lang said. AI systems have limitations, especially when it comes to separating out “crap” web content it has ingested into its training data from reliable information, Lang said.

Because of that, Reuters is building a RAG retrieval database system – techy verbiage to describe an archive of Reuters’ content that an LLM system can use to base its generated responses, and get more reliable outputs that way.

“We can use that as a solid grounding that we can give the AI to do some fact checking,” Lang said.

However, it’s still early days, and Lang’s team is experimenting with what the true use cases would be for this. To start, Reuters has created a RAG database with its style guide for journalists to ask questions, such as how to spell Ukrainian President Volodymyr Zelensky’s name (Reuters spells it “Zelenskiy”).

“A lot of what I’m doing is actually saying, ‘Well, how is this going to work in the future?’” Lang said.

Here are some other AI tools Reuters has built to speed up the newsroom production process, according to Lang:

• A “sandbox” where a newsroom staffer can input a prompt, and the system will create a chain of prompts to help with production processes.
• A host of CMS tools, such as a headline builder and bullet point functions.
• Another CMS tool that can create the first draft of a story. It can process a press release and an interview transcript and put the two together to write a first draft, for example.
• An AI summarization tool called Fact Genie that processes thousands of press releases a week and makes suggestions to the newsroom on what to cover.
• A tool called Federal Bot created by a Reuters journalist that takes all of the published data and press releases from the U.S. government (three times a day) and looks for new information, flagging anything that journalists should take a closer look at by emailing those closest to the relevant beat.