How to Verify Deepfakes With Context Instead of Spotting Glitches

Why Context Matters More Than Spotting

Spotting means looking for flaws in the media itself: odd blinking, smeared teeth, six fingers, a voice that sounds flat. It worked against early deepfakes and it still catches careless ones, but it has two structural problems. Generators remove the flaws people learn to look for, and genuine media produces the same flaws through compression, low light and filters.

The evidence on human spotting is poor. Nils Köbis and colleagues reported in iScience in November 2021 that 210 participants identified deepfake videos with 57.6% accuracy, close to chance. They labelled 67.4% of clips authentic although half were fake, incentives did not help, and confidence was inversely related to accuracy. Matthew Groh and colleagues, in a PNAS study of 15,016 participants published in January 2022, found people and the leading detection model similarly accurate overall but wrong on different videos.

Context checks ask questions a generator cannot answer by improving its output. Did this account exist last week? Was the executive in that city on that date? Does the organization's own channel carry the same announcement? A flawless fake still fails those tests.

How Do You Verify a Deepfake With Context?

Turn the item into a set of separate claims and test each one against a record that did not come from the same sender. A video of a chief executive announcing a merger, for example, makes claims about identity (this is the CEO), event (an announcement happened), time and place (today, at headquarters), source (this account is the company's) and action (buy, sell or click this link).

Each claim has its own best evidence. You rarely need all of them. One solid contradiction, such as the company's verified channels saying nothing or the CEO being on stage elsewhere, usually settles the question.

What Do Professional Fact Checkers Do Differently?

They read laterally. Sam Wineburg and Sarah McGrew of the Stanford History Education Group compared professional fact checkers with historians and college students as they evaluated live websites. The fact checkers quickly left the page they were judging and opened other tabs to see what independent sources said about it, while the other groups stayed on the page and judged it by how it looked. The fact checkers reached better conclusions in less time.

The same habit transfers to video and audio. Instead of replaying a clip to find a glitch, open a search for the claim, the account and the people involved. Mike Caulfield's SIFT method condenses this into four moves: Stop, Investigate the source, Find better coverage, and Trace the media to its original context.

How Does Context Catch Deepfake Fraud?

In fraud, the request is the context that matters most. In the Arup case, a Hong Kong finance employee received a message from the company's UK office that called for a secret transaction, and suspected phishing. A video call with deepfake recreations of the chief financial officer and colleagues overcame that doubt, and the employee made 15 transfers totalling HK$200 million, according to police statements reported by CNN in May 2024. The secrecy and urgency of the request were the strongest signals in the whole episode, and they were visible before any face appeared.

The FBI's December 2024 public service announcement on generative AI fraud draws the same lesson for families: agree on a secret word, hang up, and call back on a number you found yourself. Those controls do not depend on judging the media at all, which is why they keep working as the media improves.

When Is Context Not Enough to Verify a Deepfake?

Context needs something to check against, and some situations offer nothing. A selfie video captured during account opening has no earlier copy and no publisher. A brand-new image posted by an anonymous account has no history. A recording leaked in a workplace dispute may have no witnesses willing to speak. And if an attacker controls the official account or phone line, the trusted channel confirms the fake.

In those cases the content has to be examined directly. In the 2024 Pikesville High School case in Baltimore County, police reported that an analyst contracted by the FBI and a second analyst at the University of California, Berkeley both examined a circulated recording before police announced it was AI-generated. Context told investigators who had motive and access; forensic analysis spoke to the file itself. Using both, with each limited to the question it can answer, is the defensible approach.

How Does Neural Defend Fit Into Context-Based Verification?

NeuroDetect is designed to sit alongside context work. An investigator uploads an authorized image, audio file, video or document into a case, and the analysis records model and version context and shows heatmaps and attention views down to regions, frames and audio segments. The forensic report it exports includes provenance, uncertainty and analyst notes, so the source checks, timeline and callbacks can be recorded next to the technical findings in one file.

Where there is no context to check, as with selfies and videos captured in onboarding or account recovery, NeuroVerify analyses the media in a single API call across six signals and returns a unified risk score with signal-level evidence, in under 200 ms p95 on eligible image requests. In both products the output is a signal for a person or a policy to weigh, never the whole decision.