Reuters is experimenting with utilizing an AI agent to hurry up its video manufacturing course of. This week, Reuters employed its first AI video producer to, partly, assist oversee the initiative.
Usually, Reuters’ newsroom has used AI know-how to speed up its text-based workflows. However Rob Lang, newsroom AI editor at Reuters, mentioned he noticed a possibility to make use of agentic AI to behave as a form of super-charged video modifying system. Proper now, Reuters is utilizing AI instruments to create and course of video metadata to chop totally different edits of video protection.
“We’ve requested the LLM to choose the very best bits and create what we name a wrap edit. And it’s really doing fairly a great job. In order that’s with out even having that multimodal. So if we are able to really get the AI as an agentic AI system to grasp that, to have the ability to take a look at issues and compile issues, we’d have the ability to get the AI to construct the edits for us, which might be fairly extraordinary,” Lang mentioned on stage on the Digiday Publishing Summit Europe on Oct. 28.
To be clear, this AI agent can be making the tough reduce. There would nonetheless be a human concerned to make choices on the complete video edit, Lang mentioned, equivalent to Reuters’ new AI video producer, Enrique Flores Roldan.
However the concept is it could assist with the extra fundamental and time-consuming video edits. Ideally, it could have the ability to reduce collectively clips that think about continuity errors, equivalent to a topic taking their glasses on and off, or the solar getting into between the clouds impacting the best way a topic is lit, Lang mentioned.
“It’s… writing the directions in order that they’re understood and may be utilized [by an AI agent],” he mentioned.
About 60% of Reuters’ newsroom is utilizing AI, based on Lang. Probably the most “proficient” AI customers – which make up about 50 to 100 individuals, out of Reuters’ 2,500 journalists – are utilizing AI to “vibe code” and assist with investigative journalism.
“We’re at 60% and it’s rising each month. It’s most likely 5% each month in the meanwhile. We’ve been advised to get to 100% by the top of the yr. I believe we’ll get near 80, contact wooden,” Lang mentioned.
The largest hurdle to getting there’s not each newsroom staffer seems like they should use AI know-how, he famous.
“I used to be speaking to a photographer the opposite day, [who asked] how can I take advantage of AI? And I mentioned, why don’t you simply use it that will help you discover a parking area whenever you flip as much as an odd metropolis? Simply easy issues like that. There are methods to make use of AI that aren’t essentially journalistic,” Lang mentioned.
So what’s Reuters’ newsroom not utilizing AI for? Reality-checking, Lang mentioned. AI methods have limitations, particularly in relation to separating out “crap” internet content material it has ingested into its coaching knowledge from dependable info, Lang mentioned.
Due to that, Reuters is constructing a RAG retrieval database system – techy verbiage to explain an archive of Reuters’ content material that an LLM system can use to base its generated responses, and get extra dependable outputs that manner.
“We will use that as a stable grounding that we can provide the AI to do some reality checking,” Lang mentioned.
Nonetheless, it’s nonetheless early days, and Lang’s crew is experimenting with what the true use circumstances can be for this. To start out, Reuters has created a RAG database with its type information for journalists to ask questions, equivalent to find out how to spell Ukrainian President Volodymyr Zelensky’s title (Reuters spells it “Zelenskiy”).
“A number of what I’m doing is definitely saying, ‘Properly, how is that this going to work sooner or later?’” Lang mentioned.
Listed below are another AI instruments Reuters has constructed to hurry up the newsroom manufacturing course of, based on Lang:
- A “sandbox” the place a newsroom staffer can enter a immediate, and the system will create a sequence of prompts to assist with manufacturing processes.
- A number of CMS instruments, equivalent to a headline builder and bullet level features.
- One other CMS instrument that may create the primary draft of a narrative. It could possibly course of a press launch and an interview transcript and put the 2 collectively to write down a primary draft, for instance.
- An AI summarization instrument referred to as Reality Genie that processes 1000’s of press releases per week and makes options to the newsroom on what to cowl.
- A instrument referred to as Federal Bot created by a Reuters journalist that takes the entire revealed knowledge and press releases from the U.S. authorities (3 times a day) and appears for brand spanking new info, flagging something that journalists ought to take a more in-depth take a look at by emailing these closest to the related beat.
