A photograph arrives in a newsroom claiming to show a flood, a protest or the aftermath of a strike. Twenty years ago, checking it meant phone calls and patience. Today it must be checked in minutes, against the possibility that it is old, mislabelled, staged or generated entirely by software, while social media carries it around the world. Professional fact-checkers, the specialists news organisations rely on to make those calls, are now testing a new generation of AI verification tools that promise to speed up the detective work without lowering the standard of proof.
The need is not in doubt. Synthetic media has become cheap and convincing, and the volume of questionable material spikes exactly when verification matters most: during wars, disasters and elections. Research on misinformation trends, including coverage by the BBC, shows false images and clips routinely outpacing corrections. A single misleading video can consume hours of a fact-checker’s day, tracing its first appearance, comparing landmarks and weather, contacting witnesses, while dozens more queue behind it. Anything that reliably shortens that process is valuable, which is why trials are running in newsrooms covered by our media coverage and far beyond it.
The Verification Problem Keeps Getting Harder
The tools under test do several distinct jobs. Reverse image search has grown smarter, finding earlier versions of a picture even when it has been cropped, filtered or flipped. Geolocation assistants scan a scene for clues, vegetation, architecture, shadows, signage, and propose where it was taken. Audio analysis flags synthetic voices, and provenance systems check cryptographic credentials that some cameras and editing tools now attach to genuine files. The best of the new software combines these signals into a single workbench, presenting evidence for a human to weigh rather than a verdict to accept, a design choice the testers consistently praise.
The limits are equally clear, and the trials have surfaced them honestly. Detection tools trained on yesterday’s generators struggle with today’s, and the makers of synthetic media adapt quickly. Provenance credentials help only where they exist, and most material circulating online has none. AI assistants also make confident mistakes, inventing plausible-sounding explanations for why an image looks authentic, which makes them dangerous in the hands of a rushed user. Several fact-checking organisations describe the tools as excellent junior researchers and poor senior editors: tireless at gathering, unreliable at concluding.
What the New Tools Can and Cannot Do
What emerges from the testing is a workflow rather than a replacement. Machines triage: they sort the flood of material, surface the items spreading fastest, and assemble first-pass evidence. Humans decide: they make the calls that carry reputational and sometimes legal weight, they contact sources, and they write the careful explanations that show readers how a conclusion was reached. That transparency is itself part of the craft. A fact-check that merely asserts true or false persuades few people; one that demonstrates the checking teaches readers to check the next claim themselves.
There are pressures the tools cannot solve. Fact-checking teams are small, funding is fragile, and the platforms distributing false material move faster than any newsroom. Some testers worry that impressive software will become an excuse to cut staff, leaving fewer humans supervising more machine output at exactly the moment the material grows trickier. Others note the geographical imbalance: verification tools work best on English-language content and well-documented places, while misinformation flourishes in languages and regions where the tools are weakest and local fact-checkers fewest.
Why Human Judgement Stays in Charge
The fairest verdict from the trials so far is cautious optimism. The tools genuinely save time on the mechanical parts of verification, sometimes hours per case, and they occasionally catch manipulations a tired eye would miss. They have not changed what verification fundamentally is: an evidence-based argument made by accountable people. Newsrooms that treat them that way, as instruments inside a human process with published standards, report real gains. Those hoping the software will simply declare the truth are likely to be the next cautionary tale the fact-checkers have to write about.
For readers, the practical lesson travels well beyond the newsroom. The same habits the professionals use, checking the source, searching for earlier versions, asking who benefits from your believing it, remain available to everyone, AI or not. The tools are getting better. So, unfortunately, is everything they are built to catch, and the contest between the two is likely to define the information environment for years, as ongoing coverage in the News section will continue to document.
