NYT's Malachy Browne on AI's arms race in visual reporting

A journalist who cut his teeth hand-verifying phone videos during the Arab Spring is now running AI models over satellite imagery of Gaza looking for bomb craters. That progression, more than any single tool, is the thing worth sitting with in the Reuters Institute's interview with Malachy Browne, senior producer and enterprise director of the New York Times' visual investigations team.

Browne describes using an AI model to scan satellite imagery of areas designated as safe in Gaza, which flagged more than 500 craters consistent with 2,000-pound bombs. The same team used satellite imagery and computer vision to follow a Chinese surveillance balloon from shortly after its launch in China across the Pacific and into North America. His framing is that AI is 'a fantastic tool that we're only beginning to understand,' useful when it lets reporters look at evidence at a scale that would otherwise require desks of analysts.

The part that should be uncomfortable for anyone downstream of visual journalism, whether that is courts, investigators, platforms, or ordinary readers, is what he says in the same breath. Visual and audio verification, he says, are going to become 'extremely challenging'; detection tools are 'struggling to keep pace with generation tools,' and it is becoming 'an arms race.' Browne also calls the rise of AI-generated content 'probably the biggest challenge within our space.' If a well-resourced newsroom's most experienced verifier is saying that, the honest read is that the credibility of visual evidence itself is under pressure, not just any individual clip.

Browne's answer, as best I can tell from the interview, is method rather than magic. The team, which he joined in 2016, tries to 'weave the story together with the evidence, the questions around it, the problems that evidence can solve, and the proof that supports our conclusions.' The bet is that audiences will increasingly recognise the value of that kind of showing-your-work reporting, and that even when a story's immediate impact is limited, it holds up as a record for accountability mechanisms, prosecutors and historians later.

What the reporting doesn't give you is the boring specifics an operator would want: which AI models the Times is actually running for crater identification and object tracking, how the team plans to handle audio (which Browne calls out as the harder problem), or where it sits on provenance standards like signed capture or watermarking. The forward-looking thing worth watching is whether newsrooms with less engineering muscle can copy this workflow, because if scaled visual verification quietly consolidates into a handful of outlets, the loss is not just journalistic.