Election Deepfakes Miss the Real Evidence Problem

The Election Commission of India dropped a warning last week that made headlines across the country: political parties and campaigners must not misuse artificial intelligence or deepfake content during the upcoming Assembly elections in Assam, Kerala, Tamil Nadu, West Bengal, and Puducherry. Clean, clear, firm. Good. Now let me tell you what they didn't say — and why that silence is the bigger problem.

Deepfakes are real. The threat is real. Manipulated video of a candidate saying something they never said, distributed at scale forty-eight hours before polling day — that's a genuinely serious problem. Nobody's disputing that. But here's what's been bothering me since that announcement landed: regulators have spent enormous political energy defining what synthetic faces cannot do in a campaign, while the tools and methods used to examine real faces in the investigations that follow those campaigns operate in something close to a methodology vacuum.

That asymmetry should bother you.

The Warning Nobody Argued With

Nenow reports that the Election Commission of India announced assembly election schedules for five states while simultaneously cautioning political parties and campaigners "against the misuse of artificial intelligence and deepfake content during the election campaign." That's the whole brief. Stark, direct, and — honestly — correct as far as it goes.

And that warning didn't land in a vacuum. The EU AI Act is moving in the same direction. European ambassadors have agreed to prohibit AI practices that create non-consensual intimate content, with mandatory machine-readable watermarking and strong detection tools required by August 2026, according to Diffsense. The regulatory mood globally is: synthetic content must be labelled, restricted, and traceable. Fine. Good. Agree.

But notice what both of those regulatory frameworks have in common: they're exclusively concerned with the generation and distribution of synthetic facial content. They say nothing — nothing — about the methodology used when an investigator sits down with two photographs of real people and tries to determine whether they're the same person. This article is part of a series — start with Stress Test Facial Comparison Method Against Deepf.

"Facial recognition for entry, facial recognition for age verification for alcohol, and facial recognition for purchase is coming." — Matt Pasco, Allegiant Stadium Technology Chief, Brisbane Times

Pasco was talking about stadiums — the Brisbane 2032 Olympics specifically — but his point cuts straight to the core issue. Facial recognition is being deployed at scale in commercial settings, in elections, in law enforcement. The technology is maturing fast. The methodology standards? Not keeping pace. Not even close.

The Part Regulators Aren't Talking About

Here's the uncomfortable reality that gets almost no airtime in the deepfake conversation: when an investigator — an election integrity officer, a private investigator working a voter fraud allegation, an insurance examiner reviewing ID documentation — compares two faces, they're often doing it by eye. No standardised methodology. No documented decision framework. No minimum competency requirement.

And that is a documented problem, not a theoretical one.

Peer-reviewed forensic science research — including work published through NIST frameworks — consistently shows that untrained human examiners perform significantly worse than trained forensic facial examiners, and significantly worse still than algorithmic analysis using Euclidean distance methods. The margin of error isn't rounding-error territory. In high-stakes settings, it's the difference between correctly identifying someone and destroying their reputation or missing actual fraud entirely.

Think about that for a second. We now have formal regulatory language about what an AI cannot generate during an election campaign. But we have no equivalent language about what method an investigator must use — or document — when they're reviewing photographic evidence in the election fraud investigation that follows. The threshold for synthetic content is becoming stricter than the threshold for identity evidence in actual cases.

That's not a conspiracy. It's just where the political energy went. Deepfakes are visible, scalable, and make great headlines. Investigative methodology is unglamorous, jurisdiction-specific, and mostly invisible until something goes catastrophically wrong. Previously in this series: Netanyahu Cafe Deepfake Video Evidence Investigato.

Scale Isn't the Only Metric That Matters

I can already hear the strongest counterargument, and it's a fair one: deepfakes are a public-facing threat. One manipulated video reaches millions of voters. Investigative methodology affects individual cases. The asymmetry of scale, the argument goes, justifies asymmetric regulatory attention.

Except — and this is the part that gets glossed over — one wrongly identified individual in a high-profile investigation can do extraordinary damage to public trust in entire institutions. Scale isn't the only metric. Specificity matters too. A misidentification in an election fraud case doesn't just hurt one person; it taints the investigation, poisons the result, and hands ammunition to everyone who already believes the process is rigged.

The DNA forensics world figured this out. It took time — and some painful wrongful conviction cases — but forensic DNA evidence now operates under strict chain-of-custody requirements, laboratory accreditation standards, and documented methodology. Cold cases that sat for decades are now being solved because the methodology became rigorous enough to be trusted. Nebraska TV reports that two New York cold cases dating back to the 1970s were solved using forensic genetic genealogy — including a 1970 case identifying a John Doe whose decapitated remains had gone unidentified for decades. That's what happens when a forensic discipline gets serious about its methodology.

Facial comparison isn't there yet. Not even in the same stadium. (And given what Brisbane Times is reporting about facial recognition coming to actual stadiums by 2032, maybe that metaphor is more apt than I intended.)

What Better Actually Looks Like

Look, nobody's saying every election officer needs a forensic science degree. But the gap between "eyeballing two passport photos and calling it a match" and "using a documented, methodologically sound comparison process" is not an insurmountable one. It's a training problem. A standards problem. A documentation problem.

Algorithmic facial comparison — the kind that measures geometric relationships between facial features with mathematical precision rather than human intuition — produces results that can be documented, audited, and defended. That matters enormously when a case ends up in front of a judge or a parliamentary inquiry. Understanding where facial recognition software has genuine limitations is part of using it responsibly — but "it has limitations" is not an argument for using no methodology at all. It's an argument for using methodology that acknowledges and documents those limitations. Up next: Multimodal Biometrics Face Fingerprint Voice Defea.

The alternative — which is largely the current situation — is investigators making consequential identity calls with no documented process, no minimum standard, and no external accountability. In an environment where election integrity is already politically contested, that's not just a technical problem. It's a trust problem.

The Election Commission of India is doing the right thing by addressing AI deepfakes. That warning deserves credit. But the harder, less glamorous work — establishing what counts as a defensible facial comparison in an integrity investigation — is the job that nobody's fighting over because there are no headlines in it. Not yet.

The question worth sitting with: if a candidate's election result was challenged on the basis of a facial comparison made by an untrained official using a consumer app on a Tuesday afternoon, would anyone even know that's what happened? And would there be any standard against which to measure whether it was done right?

That's not a hypothetical. That's just a case that hasn't made the news yet.