Image Whisperer (imagewhisperer.org) is a verification tool for investigators. Most detectors hand you a percentage and leave. Image Whisperer works like a newsroom does — as a detector and a detective.
The detector does the math. It runs 44 independent checks in parallel: GPU-based AI models scanning pixel consistency, lighting and frequency patterns; forensic tests like Error Level Analysis; and EXIF and C2PA content credentials.
The detective does the reasoning. Vision models read the scene the way an investigator would — do the shadows, reflections, and anatomy make physical sense? The file hash is matched against a daily-updated database of known and debunked fakes, aggregated from fact-checking networks worldwide. Reverse-image and text searches trace who published the picture first. Then it explains itself.
Verdicts are colour-coded — red for AI-generated, orange for uncertain, green for authentic characteristics, blue when a human needs to look — and every one comes with plain-language reasoning. Never an unexplained score.
Image Whisperer v1.0 launched in June 2026, out of beta and free to use twice a day.
Image Whisperer is built and operated by Henk van Ess, an investigator, tool builder and trainer with two decades in online research and open-source intelligence. He wrote the people-research chapter of the Verification Handbook for Investigative Reporting, and teaches verification and AI-assisted research worldwide, running workshops on most continents through the year.
He trains for the Global Investigative Journalism Network and the Centre for Investigative Journalism, and has taught investigative and newsroom teams including the Wall Street Journal, The Washington Post, NBC News, ProPublica, ITV, Axel Springer, DPG Media, Bellingcat and Global Witness, alongside NGOs, law enforcement and legal teams. More than twenty Pulitzer Prize winners are among the people he has trained.
He assesses fact-checking organisations for the IFCN at Poynter and for the EFCSN, sits on the New York City Bar Association's AI Task Force, and has guest lectured at Arizona State University and the Free Tech Academy in Berlin. His publication, Digital Digging, reaches more than 11,000 subscribers in newsrooms, NGOs and agencies. In 2026 he is publishing The Researcher's Guide to Practical AI and running GIJN's courses on AI for journalists, including its webinar on detecting AI-generated content.
Every detection rule in this tool comes from that practice: cases where a fake image reached an editor's desk and a generic detector failed to catch it. That story is told in full here.
Addresses and phone numbers are at the top of this page. Everything the company claims about the product is published and dated, so you can check it rather than take it on trust.
Two different kinds of AI are involved. The detection models produce scores. A separate language model reads the picture and writes what you see in words: the Who, What, Where, When and Why, and the suggested next steps.
That written text is generated, so it is checked against the scores rather than trusted on its own. It may not name a detector that has not reported, advice contradicting the verdict is dropped rather than printed, and a location that cannot be grounded in something visible in the picture is reported as Unknown instead of guessed.
The verdict itself is not written by a language model. It comes from rules applied to the model scores, and no single check decides on its own: a verdict only moves when another independent check corroborates it. That is why it can be traced, and argued with.
This is a helper, not a judge. It is a first-pass filter, and it does not replace reporting.
The detection models run on our own hardware in Germany. Some checks are performed by outside services, and your image is sent to them to do it: Google Cloud Vision and Google Gemini for reverse image search and the written analysis, and the commercial scanners SightEngine and Hive. Those providers process in their own regions, which can be outside the European Union.
Retention, accounts, payment data and your rights over all of it are set out in the privacy policy, and the terms of use are here. If you need a specific data-residency arrangement for your organisation, ask us before you upload anything sensitive.
Two verifications a day are free and need no account. Beyond that you buy a pack, starting at €1.00 ($1) for two more, with no subscription and no renewal. Newsrooms with volume use an enterprise plan billed by invoice. All prices are on the plans page.
There is no advertising, no tracking business, and nothing is sold on. That is deliberate: a verification tool funded by attention would have the wrong incentives.
Every result carries a "was this correct?" control. Those reports are read, and they are the main reason the detection rules change. The two failures described in why detectors fail were both found by outside experts who took the trouble to say so, and both led to rewrites of how verdicts are reached.
For anything else, write to admin@imagewhisperer.org or use the contact page. Press and research enquiries are welcome, including requests to test the tool and publish what you find.
Why AI detection fails on the fakes that matter most . The cases that shaped this tool: a viral snowstorm, a composite in front of the Eiffel Tower, and a fake that two models caught and one did not.