As synthetic media floods inboxes, dating apps and video calls, one of the more promising defences against deepfake fraud may not be smarter software at all. It may be a better trained human eye.
New work from the Australian National University suggests that ordinary people, given the right coaching, can meaningfully improve their ability to tell a real face from one conjured by artificial intelligence. The ANU research lands at a moment when generated faces have become disturbingly good, and when the tools built to catch them are struggling to keep pace.
Why the human eye matters again
For most of the deepfake era, the assumption has been that machines made the mess and machines would clean it up. Detection algorithms, watermarking schemes and provenance standards have all been pitched as the answer. The trouble is that generative models improve faster than the detectors chasing them, and a classifier trained on last year’s fakes can be blindsided by this year’s.
The ANU angle flips that logic. Rather than treating people as the weak link, the research treats them as a defence layer that can be upgraded. The premise is straightforward: AI-generated faces still carry tell-tale artefacts, from oddities in the way light falls on skin to strange symmetry, mismatched earrings, warped backgrounds or teeth and hairlines that do not quite behave. Once someone knows where to look, the argument goes, their accuracy climbs.
That is a significant claim because the raw starting point is grim. A well-documented quirk of modern synthesis is that people frequently rate AI faces as more real, and more trustworthy, than genuine photographs. Generators tend to produce averaged, symmetrical, conventionally attractive faces, and human brains reward exactly those qualities. If training can push people past that instinct, it offers a cheap and scalable countermeasure that does not depend on any single piece of software staying ahead of the curve.
The news, and its limits
The finding is encouraging, but it is not a silver bullet, and the researchers are careful about that. Training lifts performance from a low base; it does not make anyone infallible. And there is a moving-target problem baked in: the artefacts people learn to spot today are precisely the flaws that model builders are racing to eliminate. Teeth, hands and reflections have already improved markedly across successive model generations. A cue that works in 2026 may be gone by 2027.
There is also the question of context. Spotting a dodgy face in a controlled test, where you know you are being tested, is very different from clocking a fake during a hurried video call from someone claiming to be your chief financial officer. Fraud thrives on urgency, authority and distraction, the conditions under which careful visual inspection tends to collapse.
Two schools of thought
The research sharpens a genuine divide in how to fight synthetic media. One camp argues the future is technical: cryptographic content credentials, hardware-signed cameras and industry standards such as the Coalition for Content Provenance and Authenticity, which attach a verifiable history to legitimate media so anything unsigned is treated with suspicion. On this view, asking humans to eyeball pixels is a losing game.
The other camp, into which the ANU work fits, holds that provenance systems will take years to reach ubiquity, will never cover every screenshot or re-encoded clip, and leave a vast gap that only human judgement can fill in the meantime. Media literacy, in this framing, is critical infrastructure. The two positions are not mutually exclusive. The most credible defence is layered: trained people, better detection tools and provenance standards working together, with none of them asked to carry the load alone.
What it means for Australia
For Australia, the stakes are concrete and expensive. Scamwatch and the National Anti-Scam Centre have logged deepfake-enabled investment cons, including fabricated video of prominent Australians spruiking crypto schemes, and voice-cloning scams that mimic family members or executives to authorise payments. Losses to scams run into the billions each year, and synthetic media lowers the cost and raises the plausibility of the con.
The identity-verification stakes are just as sharp. Banks, government services and the emerging digital-ID ecosystem increasingly lean on facial recognition and liveness checks to confirm who is on the other end of a transaction. If a generated face can satisfy those checks, the exposure is systemic. Human vigilance becomes a backstop for automated systems that can be gamed, and a trainable workforce, in call centres, bank branches and fraud teams, becomes part of the control environment.
Regulators are already circling. The federal government’s work on scam-prevention obligations for banks, telcos and digital platforms, along with the broader debate over AI guardrails, all assume that some mix of technology and human oversight will police synthetic content. Research showing that the human half of that equation can actually be improved is useful ammunition for policymakers weighing where to spend. It suggests that funding public education and staff training is not a soft option but a measurable intervention.
There is a sovereignty dimension too. Australia has leaned heavily on the argument that home-grown research and infrastructure matter in the AI era, and university labs such as ANU’s are a big part of that story. Detection and media-literacy expertise developed locally can feed into national standards, procurement rules and the training programs that banks and agencies roll out, rather than importing every safeguard from offshore vendors.
What is next
The obvious next step is durability. If artefact-spotting cues degrade as models improve, the value of any training program depends on how often it is refreshed and whether people learn transferable habits of scepticism rather than a fixed checklist of glitches. Expect follow-up work testing whether trained users hold their edge against the newest generators, and whether the skill survives outside the lab.
For organisations, the practical takeaway is immediate. Fraud teams and frontline staff can be coached now, cheaply, while the technical defences mature. For everyone else, the message is a familiar one dressed in new clothes: the confident, attractive, perfectly lit face on your screen deserves more suspicion than instinct wants to give it. The uncomfortable truth from this line of research is that the fakes have become good enough to feel more real than reality, and that closing the gap starts with knowing you are already fooled.
