Ion files patent to block harmful AI video at source

Ion has filed a patent in Australia for a system designed to block harmful AI-generated video before it is shown. The filing expands the Melbourne company's intellectual property efforts in video authentication and safety.

The patent application, titled System and Method for Content Safety Enforcement During Virtual Video Assembly, centres on a process that attaches safety rules to individual video elements and checks those rules as a video is assembled, rather than after a clip has been uploaded or displayed.

The approach differs from the moderation model used by many online platforms, where content is typically reviewed after publication and removed if it breaches rules or laws. Ion's system is intended to stop a video from being assembled at all if any component fails an active policy check.

Finbar O'Hanlon, Chief Innovation Officer at Ion and the inventor named in the filing, said the patent extends the company's earlier work on video trust.

"Our previous patent gives video a way to prove it is real. This new patent is not so much about whether the content is genuine, but whether it should be assembled and shown at all.

"It classifies safety at the level of the individual sample and enforces it at the moment of assembly, so if a sample breaches an active policy, the video simply cannot resolve.

"Unsafe content is never delivered, rather than flagged after a viewer has already seen it. And because the rules can be set by the rights holder, the platform and the jurisdiction at once, with the strictest combination always winning, the same source content can be assembled safely under very different laws without duplicating a single file.

"In a world where AI agents assemble video on the fly, safety can no longer be a check run on a finished file. It has to live in the moment of assembly, and that is what this patent protects."

Ion is listed on the Australian Securities Exchange under the code IOV. The company argues the issue is becoming more pressing as generative AI tools make it easier to create synthetic footage and as automated systems begin to compile video dynamically for viewers.

Policy at source

The proposed system applies classification and enforcement at what Ion describes as the binary sample level. In practical terms, the rules are tied to the smallest underlying pieces of video used in assembly, and the system checks those pieces against active requirements before the final output appears.

The rules can be set by several parties at the same time, including rights holders, platforms and local jurisdictions. Where those standards differ, the strictest rule would apply during assembly.

The model could appeal to businesses and regulators dealing with different legal frameworks across markets, particularly when one set of source material must be distributed under varying content standards. Ion says the same source content could be assembled differently without creating duplicate files for each territory.

O'Hanlon said the company sees the invention as a distinct technical category as lawmakers and media groups respond to deepfakes, child safety concerns and online disinformation.

"Ion intends to position it as directly relevant to the platforms, broadcasters, studios and regulators confronting the deepfake, child-safety and disinformation challenges of the coming decade."

Broader push

The filing follows another patent application by Ion aimed at determining whether a video is authentic or AI-generated. Together, the two efforts point to a broader strategy spanning both video provenance and content control.

The first line of work focuses on whether footage is real. The second focuses on whether material should be allowed to be assembled and shown in the first place, even if the source components are available within a system.

For platforms, that distinction is significant. Much of today's moderation infrastructure is built around finished files that can be scanned, flagged and taken down. Ion argues this model may become less effective if video is increasingly generated or assembled on demand, leaving little or no fixed final file to inspect before display.

That challenge is likely to grow as AI systems move beyond creating static clips towards producing personalised or responsive media in real time. In that environment, content controls would need to operate during creation rather than only after publication.

O'Hanlon said that shift was the main driver behind the invention.

"Yet the systems built to keep video safe were designed for a world that is disappearing, one in which every video is a finished file that can be checked before it is shown.

"As AI agents increasingly assemble video on demand, from fragments, at the moment of viewing, there is often no finished file to check at all.

"Ion's patent rebuilds content safety for that world. It binds a safety classification to every binary sample and enforces it at the precise moment a video is assembled, so unsafe content is structurally prevented from resolving rather than detected after the damage is done. In short, it is prevention, not detection."

He added that new harms, classifiers and regulatory standards could be added through adapters without changing the core media pipeline.