Sydney AI porn startup faces child-safety questions

A Sydney startup Coreflow is facing scrutiny after an ABC investigation found synthetic characters on its OurDream platform that age-estimation technology assessed as underage, a finding that pulls a fast-growing consumer AI business into Australia’s child-safety debate. The sharper question for regulators is not whether explicit AI tools exist, but whether a service that claims global scale can stop illegal-looking output before users generate, share and regenerate it.

On the company’s own numbers, the stakes are already large. OurDream says it has about 20 million users and 48 million monthly visits, while the ABC reported Coreflow advertised senior growth roles paying up to $450,000 a year and engineering roles up to $800,000. Those figures suggest a startup pushing hard on distribution and product speed, the two pressures that usually make moderation look adequate in policy decks and brittle in production.

Coreflow says its rules are stricter than critics suggest. In a statement carried by the ABC, OurDream said it bans not only depictions of minors but any character whose body proportions “read as a minor”, regardless of stated age. That is the tension at the centre of the case: the platform is claiming a preventive system, while the investigation points to how easily edge cases can still leak through.

“We do not host content depicting minors. Ever … This includes … any character that, in body proportions, reads as a minor regardless of stated age.”

— OurDream, via ABC News

Growth claims and moderation limits are colliding

For an adult-content startup, moderating an archive is one thing; moderating a generator is another. Once a user can spin up variants in seconds, every safeguard becomes reactive unless the model, prompt filters and review queues are aligned well before publication.

In the ABC’s account, OurDream relied in part on user moderators and one tested character was assessed by age-estimation technology at 15.1 years. Even if any single age estimate is contestable, the operational problem is obvious: if independent testing can surface child-like outputs from the public product, the moderation system is not failing at the edges, it is being asked to police scale after the fact. That is the same structural weakness Wired reported this month when it found xAI’s Grok Imagine still hosting sexualised deepfakes of famous women. Different product, same governance failure mode.

Commercial pressure cuts the wrong way. Adult AI products compete on novelty, speed and retention, not on slower review flows that frustrate paying users. In that environment, trust-and-safety teams are often asked to prove a negative: that no prohibited output made it through a system designed to generate endless variation. They rarely can.

Australia’s law already looks less ambiguous than the product pitch

Legally, the question also appears narrower than startups might hope. According to the ABC’s reporting, former NSW Police child-abuse squad detective Jon Rouse said the images he was shown appeared to be child abuse material under Australian law and concluded, “That needs to be stopped.” The significance is that existing child-sex-abuse provisions already focus on depiction and apparent age, not just whether a real child stood in front of a camera.

That reading also lines up with the eSafety Commissioner’s guidance, which warns generative AI is already being used to create child sexual exploitation material, and with ICMEC Australia, which has been pressing Canberra to treat AI-enabled abuse as an urgent enforcement problem rather than a future one. The point is not that every synthetic image automatically breaches the law. It is that “synthetic” is no longer much of a shield if the output depicts, represents or appears to show a minor in a sexual context.

“AI is being weaponised to harm children, and Australia must act.”

— Colm Gannon, ICMEC Australia

For policy makers, that answer matters as much as the legal theory. One of the regulator camp’s core questions is what notice-and-takedown rule makes sense when banned material can be regenerated in seconds. If the only remedy is to remove each image after it appears, the compliance model is already behind the product model.

This is no longer just an adult-site problem

Beyond one site, the Sydney case sits inside a wider shift from niche content moderation disputes to mainstream platform governance. The Verge argued in May that the United States’ new deepfakes crackdown would be messy but inevitable once sexualised AI imagery moved from fringe communities to large platforms. Since then, UK regulators have tightened expectations around intimate-image abuse, YouTube has widened its likeness-detection tools, and Apple faced WWDC protests over nudify apps. The pattern is clear: once generative systems can create harmful sexual content at scale, the regulatory burden moves upstream to the platform and the model, not just the user.

More broadly, Australia has been edging toward that conversation for months. A Conversation analysis in May argued that criminalising sexualised deepfakes would not be enough without also confronting the apps that make them. The Coreflow case gives Australian regulators a local, venture-backed example of the same problem. It is no longer about a foreign app store or a Silicon Valley lab. It is a Sydney startup, Australian hiring, Australian law and a product claiming mass global reach.

Viewed that way, the most important number in this story may not be 20 million users or 48 million visits. It is the gap between how quickly a generator can create risky content and how slowly a company, or a regulator, can prove it has been caught. If Canberra wants a practical AI safety agenda, this is what it looks like: fewer abstract principles, more hard questions about liability, logging, prompt controls, age-assurance claims and how fast a platform must act when its safeguards appear to fail.

None of that proves every legal question around OurDream has already been settled. It does, however, make one policy point hard to ignore: in AI sexual-content markets, scale is not a sign that safety has been solved. It may be the clearest evidence that the hard part has only started.