As synthetic video floods social feeds, messaging apps and even corporate video calls, the question of how to stop a convincing fake before it does damage has moved from academic seminar to boardroom priority. An Australian company reckons it has an answer, and it wants a patent to prove it.
Ion, an Australian firm working in the trust and safety space, has filed a patent for technology it says can block harmful AI-generated video at the source rather than chasing it around the internet after the fact, according to a report by SecurityBrief Australia. The pitch is deceptively simple: instead of asking platforms to detect and take down deepfakes once they are already circulating, intervene at the point of creation or upload so the malicious clip never gets a clean run at an audience.
Why the timing matters
The deepfake problem has stopped being hypothetical. Generative video tools have improved so quickly that a passable fake of a public figure, a chief executive or an ordinary person can be produced in minutes, often from a handful of reference images and a short audio sample. The results are being used for non-consensual intimate imagery, financial fraud, political disinformation and reputational attacks.
The scale is what worries regulators. Detection has become an arms race in which the defenders are perpetually a step behind: every improvement in a detector is quickly answered by a more convincing generator. Worse, detection is reactive by design. By the time a fake is flagged, screenshotted and reported, it may already have been seen by thousands of people, and the original harm, whether that is a scammed retiree or a humiliated teenager, is done.
That is the gap Ion is aiming at. Blocking “at the source” implies moving the intervention upstream, closer to where content is generated, encoded or first uploaded, rather than at the messy downstream layer of feeds and shares where content moderation currently lives.
How the approach is meant to work
Public detail on the filing is limited, as is normal for a patent that has been lodged but not yet granted, so the precise mechanics remain closely held. But the broad category of “source-level” defences against synthetic media is well understood, and Ion’s approach sits within it.
Source-level systems generally lean on one or more of a few techniques. The first is provenance and content credentials: cryptographically signing genuine footage at the point of capture so that anything without a valid signature is treated as suspect. The second is watermarking, where AI generators embed a hidden, machine-readable marker in their output that downstream systems can read to confirm a clip is synthetic. The third is real-time analysis at the upload or transmission layer, screening video as it enters a platform or network and refusing to pass on content that matches known harmful patterns.
The commercial attraction of doing this at the source is that it does not depend on catching every fake in the wild. If the interception happens before publication, a single check can prevent thousands of downstream shares. It also potentially sidesteps the whack-a-mole economics of takedowns, which cost platforms enormous sums and still leave victims exposed during the lag between publication and removal.
Two ways to read it
Supporters of the source-first model argue it is the only intervention that scales. Downstream moderation, in this view, is structurally doomed because it forces defenders to inspect an essentially infinite volume of content after harm has begun. Shifting the checkpoint upstream, the argument goes, is how you get ahead of the curve instead of forever cleaning up behind it. A defensible patent in that space could also give an Australian company a genuine edge in a market dominated by much larger overseas players.
Sceptics counter that no single choke point captures the whole internet. Bad actors increasingly run open-source generative models on their own hardware, entirely outside the reach of any platform-level or vendor-level control. A source-level system works beautifully against content that flows through cooperating tools and platforms, and not at all against a determined operator generating fakes offline and distributing them through channels that never agreed to be screened. There is also the perennial patent question: filing is not granting, and a filing describes an ambition, not a shipped, independently verified product. Critics will want to see accuracy rates, false-positive numbers and evidence that the system does not quietly block legitimate content along with the harmful material.
What it means for Australia
For Australia, the story lands in the middle of an active policy debate. The federal government has spent the past two years tightening the rules around online harms, from the world-first move to keep under-16s off social media to an expanded remit for the eSafety Commissioner and ongoing work on mandatory guardrails for high-risk AI. Deepfakes sit squarely inside that agenda, particularly after Commonwealth laws criminalising the sharing of non-consensual deepfake sexual material.
Home-grown technology that can demonstrably reduce the flow of harmful synthetic video would give policymakers something they currently lack: a practical enforcement tool to sit alongside the legislation. Regulation can make an act illegal, but it cannot by itself stop a clip from being generated and shared. That is a job for technology, and an Australian solution keeps both the intellectual property and the safety expertise onshore rather than importing it.
There is a sovereign-capability angle too. Australia has repeatedly flagged concern about relying on offshore platforms to police content aimed at Australians. A local firm holding patents in AI safety infrastructure fits neatly with the push to build domestic capacity in trustworthy AI, an area where the country risks being a pure consumer of overseas technology rather than a producer.
What is next
The immediate milestones are procedural. A filed patent must clear examination before it is granted, a process that can take years and often narrows the original claims. Beyond the paperwork, the harder test is deployment: whether Ion can turn a filing into a working product that platforms, telcos or enterprises are willing to integrate, and whether it can publish the kind of independent performance data that turns a claim into a credential.
The bigger question hanging over the whole field is interoperability. Source-level defences only reach their potential if generators, platforms and networks agree on shared standards for provenance and watermarking. Ion’s filing is a signal that Australian firms intend to compete in that fight rather than watch from the sidelines. Whether it becomes a genuine line of defence or one more promising idea that struggles against determined bad actors will depend on what comes after the patent office.
Sources: SecurityBrief Australia.
