The current crop of AI image-authenticity tools tends to do one thing. You upload a photograph, the tool returns a number. 73% likely AI-generated. 0.21 confidence of manipulation. The number is supposed to settle the question. It doesn’t, of course, because the number is a statistical probability that the image was entirely or partially generated by an AI. That’s it.
Real fact-checking has never worked like that. When a verification desk takes apart a viral image from a war zone, the work is an investigation. Cross-referencing GPS data against known geography. Pulling satellite imagery for the claimed location at the claimed time. Checking weather records against the lighting in the frame, sun position against shadow angles. Querying public tracking databases for vessels and aircraft. Reading witnesses in multiple languages. Searching for other photographs of the same event that should exist if the event actually happened. Asking who took the picture, who distributed it, and what story its existence is meant to tell. Context matters because 90% of misinformation is not AI-generated images but real images used with different, false contexts.
The issue is that it’s slow, expensive, expert work, and it lives in maybe a hundred specialists collectively across the world’s wire services. Most images in circulation never get this treatment. Most images that get questioned get the probability-score treatment instead, which is to say they don’t really get verified at all.
The question worth asking right now is whether an AI agent could conduct the actual investigation. Chain the heterogeneous structured sources. Read the witness statements. Pull the satellite tiles. Run the deductive checks. Build the evidence dossier.
The AI agent as investigator
The answer is starting to look like yes, and it’s worth paying attention to what that means.
Consider what investigative verification actually involves in practice. Someone publishes a photograph claiming to show a particular vessel at a particular location on a particular date. Public ship-tracking data is accessible. The agent queries it: Where was that vessel on that date? If the logs place the ship a thousand miles from the photograph’s claimed coordinates, the agent’s role is to make the contradiction visible. Three hypotheses remain. The image is misattributed. The tracking data is wrong, since these systems can be spoofed. The timestamp is wrong. The agent weighs those hypotheses and surfaces the evidence chain that produced them.
This is exactly the kind of work investigative open-source teams have been doing manually for years on the highest-profile cases. It’s painstaking, methodical, multi-source. And it’s structurally well-suited to agents, because what agents do well is chain queries across heterogeneous structured sources at speed. The investigation that takes a researcher three weeks could plausibly run in three minutes if the infrastructure is wired correctly.
That’s the conceptual shift. Most of the AI-authenticity discourse is about better detectors. The more interesting development is the move toward investigators.
A Glimpse into the future
A tool worth looking at as an indication: Henk van Ess’s ImageWhisperer, which launched out of beta in February. Already in use at major wire services. Its design is explicit about the move it’s making. Instead of returning a probability score, it runs 41 different checks — reverse image search against a database of debunked photographs, anomaly detection across four specialized models, OCR on embedded text, QR code parsing, short-URL tracing, social media archeology, and LLM interpretation that weighs the evidence and explains the reasoning. The output reads like a structured investigation, color-coded by confidence, with the evidence chain visible and the reasoning explained.
Van Ess describes it as a first-pass filter, with the explicit caveat that human judgment remains the load-bearing layer. That’s the right framing. ImageWhisperer is doing maybe ten percent of what a full investigator-agent could do, and it’s already operating at scale in newsrooms, handling some of the most contested images in the world. The trajectory from here is what matters.
A fully developed investigator-agent would extend the chain. Public tracking databases for vessels and aircraft. Satellite imagery against claimed locations. Weather records against lightning and atmospheric conditions. Registries, maritime, corporate, customs. Witness synthesis across many languages and platforms. Photographic triangulation across multiple sources of the same event. Motive modeling: why would this image exist as presented, who benefits from its circulation, and what narrative does it serve? Intent projection: given what we know about the photographer and the distributor, what’s the most plausible story behind the artifact?
That last layer is where the agent stops being a pattern matcher and starts being an argument-maker. Asking why an image exists is an example of abductive reasoning. Generating plausible hypotheses and ranking them by evidence is what language models do natively, given the right context. The capability is there. What needs to be built around it is judgment about how to use it.
The hard questions
This is where the harder questions start. An agent that models how one government shapes images should equally model how others do the same, allies and adversaries, on the same terms. The motive library is the agent’s political theory. There is no neutral set of priors here. Whoever builds the investigator-agent decides whose motives get modeled and whose don’t, and that decision shapes every verdict the system produces. The technical architecture can be neutral. The governance can’t. And this is where the next cyber war could unfold: by influencing fact-checking agents.
The other shift is what the output looks like. A probability-score detector returns a number that travels easily into headlines and downstream automation. An investigator-agent returns an evidence dossier, closer to a court filing than a pass-fail check, an argument that can be interrogated and contested. Newsrooms and platforms that consume the dossier as journalism input will use it well. Newsrooms and platforms that consume it as journalism output, treating the brief as the verdict, will create exactly the confidence-laundering machine that the move away from probability scores was supposed to escape.
Then there’s the substrate question. Even the most capable investigator-agent has to anchor its reasoning somewhere. Cryptographic provenance at the point of capture, signatures embedded by the camera, watermarks that survive platform stripping, give the agent a privileged read that isn’t itself the output of another model. Without that anchor, the investigation is recursive: representations checked against representations, models compared against models, with no purchase on physical reality. The provenance infrastructure becomes more important as the agent becomes more capable, because it’s the only signal in the chain not itself produced by AI.
The investigator-agent and the provenance layer are the same project from different sides. One asks how the image came into existence. The other asks what story the image is being made to tell. Both have to function together for visual authenticity to mean anything at an industrial scale.
It’s when
There’s a version of the next two years where most images in public circulation get investigator-agent treatment as a matter of course. Each image gets treated as a case file. The agent runs the first-pass evidence brief. The editor decides what the brief means and whether to act on it. The verification specialist transforms from the person who conducts the investigation to the person who audits what the agent did, a different professional identity, one that requires exactly the institutional expertise the agent was supposed to obviate. And now, he can do it at scale within hours, if not minutes, instead of weeks or months.
The harder version of the same question: who runs the investigator-agent? A newsroom-controlled agent serves that newsroom’s editorial standards. A platform-controlled agent serves as a moderation policy. A state-controlled agent serves the state’s truth regime. The same technical artifact produces radically different epistemic regimes depending on deployment. This is the governance question that should occupy newsroom leadership and standards bodies right now, and it’s largely not being discussed because everyone is still focused on detector accuracy.
The detector question is mostly settled. The investigator’s question is wide open. Someone needs to notice.
Author: Paul Melcher
Paul Melcher is a highly influential and visionary leader in visual tech, with 20+ years of experience in licensing, tech innovation, and entrepreneurship. He is the Managing Director of MelcherSystem and has held executive roles at Corbis, Gamma Press, Stipple, and more. Melcher received a Digital Media Licensing Association Award and has been named among the “100 most influential individuals in American photography”