Deepfakes, Impersonation, and the Verification Habit

For a few years the worry about synthetic media ran in one direction: what somebody might fabricate about your candidate. That worry has kept its place and acquired a twin. The same tools that can put false words in your candidate's mouth can put your candidate's voice on a phone call to your own treasurer. Synthetic media is now cheap in both directions — outward, at your reputation, and inward, at your people — and the defence in both directions is the same unfashionable thing: a verification habit.

Verify before you amplify

The outward case arrives as temptation. A clip of the opposing candidate saying something disqualifying lands in your mentions during the last fortnight. Every instinct says share; the share button is faster than the sceptical mind. But a campaign that amplifies a fabrication torches its credibility at the exact moment credibility is its scarcest asset, and the correction never travels as far as the clip did.

The habit that protects you is provenance before forensics. Where did the clip first appear, and from what account? Who filmed it, and does anyone who was in the room confirm it? Does a longer or higher-resolution original exist? Is any outlet with standards reporting it, or only accounts that materialised last month? These questions are answerable in twenty minutes, and they resolve most cases before any detection tool is needed.

Australian law adds a useful tell. Electoral communication is required to carry an authorisation naming who approved it, and a synthetic smear carries none — its makers are, by design, nowhere to be found. But the law around synthetic content itself is still catching up, and no takedown arrives fast enough to matter in a campaign's final week. Discipline has to do the work that regulation cannot yet.

What detection can and cannot tell you

Detection services exist and are worth knowing: upload an image, a video or an audio file and receive a probability that it was machine-made; some now sit inside calls and meetings listening for cloned voices. Used well, they are one signal among several. Used as an oracle, they are dangerous.

The limits are structural. Detection is probabilistic, trained on yesterday's generators, and locked in an arms race with tomorrow's. Compression, cropping and re-uploading — precisely what happens to every clip as it spreads — degrade the evidence. A clean score is not authentication, and a flagged score is not proof. Treat the tool like a smoke detector: a reason to go and look, never a verdict to publish. The judgment about what your campaign says remains a human job, and the wise version of that judgment often says nothing until provenance is settled.

The attack that comes through the group chat

The inward direction is less cinematic and more likely. Campaign work runs on high-trust channels — encrypted messaging apps, personal email, the volunteer group chat — and generative tools have made impersonation on those channels fluent, personal and multilingual. The message that reads exactly like your campaign manager, at the hour she usually writes, asking the treasurer for an urgent transfer or a volunteer for the roster spreadsheet: that is the modern phish, and the old advice about spotting bad spelling is now advice about spotting attacks from a decade ago.

The defences are procedural, not technological. Any request touching money, credentials or supporter data gets verified on a second channel — a call to a number you already had, never one supplied in the message. Access to lists and accounts is trimmed to what each role needs, so one compromised volunteer cannot lose the whole database. And the culture matters most: a campaign where phoning to double-check is praised as professionalism, not resented as insult, is a campaign where the scam finds no purchase.

The final week

Now compress all of this into the week that decides things. A fabricated clip of your candidate surfaces on the Tuesday before polling day. Pre-poll has been open for days; thousands in your electorate have already voted and thousands more will before any correction lands. There is no time to design a response, which is why the response must already exist: who assesses, who contacts the platforms, who briefs journalists, who talks to the AEC, and who tells the volunteers — usually the hardest part — not to engage with the clip and feed its reach.

One preparation outweighs the rest: keep your own archive. Original recordings of speeches, debates and interviews, timestamped and retrievable within the hour, let you answer a fabrication with the authentic record rather than a denial. Denials sound like politics. Originals end arguments. A campaign that has practised producing them has turned verification from a crisis reflex into what it needs to be — a habit, held all year, so that it holds in the week it is tested.