AI-Generated News Media: A Source-First Verification Workflow

Start With the Source, Not the Detector

A news clip reaches you already stripped of its origin. It was screenshotted, re-uploaded, clipped, captioned, and pushed through several apps before it landed in front of you. That travel history is the first thing worth reconstructing, earlier than any question about whether the pixels were synthesised.

Source-first verification means treating the origin, path, and handling of a file as evidence in its own right, ahead of any automated score. A file's journey often explains more than its content. A genuine video presented under a false date or location is misinformation even though nothing in it was generated by a model, and a synthesis check alone will never catch that case.

What NIST's Media Forensics Work Tells Us

The National Institute of Standards and Technology (NIST) runs the Open Media Forensics Challenge (OpenMFC), an open evaluation series that measures automated image and video manipulation detection and localization technologies. Its stated aim is to advance the state of the art of media forensics and to support researchers with benchmark datasets, evaluation infrastructure, and a communication forum.

Two details from that program are worth carrying into everyday verification. First, OpenMFC asks systems not only to decide whether media was manipulated but, where possible, to identify the region and type of manipulation, a reminder that a binary real-or-fake verdict is a coarse instrument. Second, related MFC work has included constructing a phylogeny graph describing the manipulation history of an image and verifying media sensor identification. Origin and edit history are treated as part of the forensic problem, not an afterthought.

The program is built on experience from the DARPA Media Forensics (MediFor) MFC evaluations, and NIST initiated the OpenMFC leaderboard platform in 2020. The program is structured around measurement, benchmark datasets, evaluation infrastructure, and repeated evaluations. One caveat carries directly into news work: NIST describes OpenMFC as focused on image and video, and does not list audio as an object of evaluation, so a soundtrack needs its own scrutiny.

The Source-First Workflow, Step by Step

1. Trace the first appearance. Reverse-search a representative frame or the clip itself. Look for an earlier posting under a different date or caption. Check whether an older video has been paired with a new story; investigate that question separately from synthesis detection.

2. Interrogate the account. Check when the account was created, what it has posted before, and whether it has a traceable identity. An anonymous upload with no history is a reason to slow down.

3. Preserve the file and its metadata. Save the original rather than a re-screenshotted copy. Note any available metadata, then verify it against the earliest traceable source rather than treating it as proof.

4. Reconstruct context independently. Do the claimed place, date, signage, and language match other records of that event? Corroboration from outlets that verify their own footage carries more weight than the sheer number of reposts.

5. Only now, run a detector. Upload the media to a probabilistic tool such as DeepFakeCheck. Record the score as one input beside the source work you have already done. Initial use does not require signup, and there is a limited free-use allowance.

Reading a Detection Score Without Overreading It

Automated detectors return probabilities, not rulings. They produce false positives, flagging authentic media as manipulated, and false negatives, missing real manipulation. A high score is a prompt to investigate further; a low score does not certify a file as clean.

Two constraints keep a score honest. Match the tool to the medium: NIST's program benchmarks image and video, while a news package may also carry audio and text that require separate verification steps. And weigh the cost of each error for your decision, because publishing a manipulated clip as real and denouncing an authentic clip as fake are different failures with different consequences.

DeepFakeCheck analyses image, video, audio, and text as risk signals, but none of that converts a probability into a verdict. The score belongs inside your reasoning, not at the end of it.

When the Evidence Runs Out

Sometimes the source cannot be traced, the context cannot be confirmed, and the score sits in the ambiguous middle. That result is itself information: you do not yet have grounds to amplify the item. Withholding a share, labelling something unverified, or routing a high-stakes case to a specialist reviewer are all legitimate endpoints. The burden of proof rests on the claim that the media is authentic, not on your doubt.

Sources

• NIST — Open Media Forensics Challenge (OpenMFC): https://www.nist.gov/itl/iad/mltg/open-media-forensics-challenge