Reuters Runs AI Across a Century of Archive Video to Make a Million Clips Searchable

Reuters has quietly solved a problem that's plagued every large media archive: how do you make a century of unstructured video actually searchable? The news agency announced it has applied AI to its entire video archive, nearly one million clips stretching back to 1896, with support from the Google Digital News Innovation (DNI) Fund. The result is time-code accurate speech-to-text transcription across 11 languages, plus automated identification of public figures appearing in the footage.

This isn't a minor tagging update. We're talking about content ranging from the Wright brothers' first flight in 1903 to pivotal World War II footage, all now sitting inside Reuters Connect with searchable transcripts attached. For an archive this old, that's a genuinely hard engineering problem. Old film has degraded audio, inconsistent formats, and decades of varying recording standards. Getting a speech-to-text model to perform reliably across that range isn't trivial.

The Technical Approach

The project leaned on machine learning for two core tasks: transcription and facial recognition, both applied at a scale that would be impossible manually.

• Speech-to-text transcription: Reuters generated transcripts in English, Spanish, German, French, Italian, Portuguese, Russian, Chinese, Japanese, Arabic, and Hindi. Each transcript also gets an automatic English translation, meaning a Russian-language clip from decades ago becomes instantly readable to an English-speaking editor.
• Time-code accuracy: The transcripts aren't just bulk text dumps. They're synced to specific moments in the footage, so a researcher can jump directly to the exact second someone starts speaking rather than scrubbing through an entire reel.
• Public figure identification: The system flags recognizable public figures within clips, letting users navigate straight to the segments featuring them instead of watching full videos hunting for a specific appearance.
• Data labelling at scale: Reuters describes this as part of a broader "archive modernisation project," where the heavy lifting was in analysis and labelling, essentially building a structured dataset out of what was previously just a pile of unindexed footage.

That labelling step is where most of the real engineering effort probably went. Raw video isn't useful to a machine learning pipeline until it's been annotated, timestamped, and organized into something a model can actually parse. Reuters says it leveraged its "growing AI and machine learning capabilities" for this, which suggests an internal pipeline built specifically for processing archival footage rather than an off-the-shelf tool bolted onto old tape.

Sue Brooks, Reuters' Managing Director of Product Development & Agency Strategy, framed the payoff in practical terms: "By applying AI technology to more than 100 years of archive video, we can offer customers around the world a vastly enhanced experience. This innovation means users of Reuters Connect can discover the exact moments that matter in history, from Lenin to Trump, WWI to 9/11, Sarajevo to Gaza and so much more. By unlocking this data, we will deliver unparalleled value to customers for years to come."

That range of examples, Lenin to Trump, WWI to 9/11, tells you something about the scope here. This isn't a curated highlight reel. It's an attempt to make the entire archive queryable, treating a century of footage as a searchable dataset rather than a static vault.

Why This Matters for Archive Work Generally

Media archives are a classic unstructured data problem. You've got millions of hours of footage, much of it undigitized in any meaningful sense, sitting in formats that make search nearly impossible without watching the whole thing. Editors and researchers have historically relied on manual metadata, written descriptions, and institutional memory to find anything specific. That doesn't scale, and it definitely doesn't scale across 130 years of content.

What Reuters built here is essentially a retrieval system layered on top of raw video. Speech-to-text turns spoken audio into indexable text. Facial recognition turns visual content into structured, searchable metadata. Combine both and you've converted video, historically one of the hardest media types to search, into something closer to a text document you can query.

The multilingual angle is worth calling out too. Building transcription pipelines that work reliably across 11 languages, especially on decades-old audio with inconsistent recording quality, is a nontrivial NLP challenge. Automatic translation into English on top of that adds another layer of complexity, since translation quality depends heavily on how clean the original transcription is. Garbage in, garbage out applies just as much to translation pipelines as anything else.

This kind of project also hints at where AI-assisted archive work is heading more broadly. Museums, broadcasters, and government archives are all sitting on similar mountains of undigitized or under-indexed footage. The techniques here, transcription, translation, facial recognition, timestamped indexing, aren't exotic. They're relatively mature ML capabilities. What's notable is the scale of application: one million clips, over a century of history, processed comprehensively rather than piecemeal.

Reuters Connect is positioning this as a product differentiator, giving customers, editors, producers, researchers, faster access to historically significant footage. That's the business case. But the underlying technical achievement, making a century of degraded, multilingual, unstructured video genuinely searchable, is the more interesting part for anyone working on similar large-scale digitization or media processing problems.

Key Takeaways

Reuters processed its entire video archive, nearly one million clips dating to 1896, using AI-driven speech-to-text and facial recognition, backed by the Google DNI Fund. Transcripts now exist in 11 languages with automatic English translation, all time-code synced for precise navigation. The bigger lesson: at sufficient scale, transcription and recognition models can turn even century-old, degraded archival footage into a structured, queryable dataset, a template other large media and institutional archives will likely follow.