Criminal Law

Warning: This story discusses child sexual abuse and child sexual abuse material.

Sitting amongst the top 50 Generative Artificial Intelligence web products, OurDream.AI allows users to create a custom AI companion or girlfriend. On the website, users can access the content on both free and premium plans, and are capable of generating images and videos that can involve ‘erotic roleplay’.

However, a recent investigation by the ABC has revealed that the platform could be generating child sexual abuse material.

The pornography website gained the attention of the ABC through a Sydney-based start-up. Coreflow, which began operating out of Neutral Bay, Sydney, has 16 staff members. Job advertisements for the start-up advertise salaries in the hundreds of thousands of dollars for engineers and “growth” team members.

When looking into the start-up’s OurDream.AI product, the ABC identified over 10 female AI-generated “companions” appearing ‘15 years old or younger, according to age-estimation technology.

Even though none of the “companions” are real people, they still potentially constitute child sexual abuse material, and therefore may constitute an offence under Australian law. A police investigation into the ABC’s findings is ongoing.

Following other controversies, such as nonconsensual sexual images generated by X’s Grok, the OurDream.AI findings raise the repeated question: how will the law keep up with AI?

Current Laws

Presently, the Crimes Act 1900 (NSW) outlines child sexual abuse material laws within New South Wales. Even if the material is AI-generated, creating, distributing or possessing material that depicts the abuse of someone under the age of 18 constitutes a criminal offence.

This is similarly criminalised under the Criminal Code Act 1995 (Cth).

However, whether the content has to be ‘real’ may appear unclear.

In NSW, the issue of whether the content has to be ‘real’ was disputed in McEwen v Simmons (2008).

McEwen was in possession of cartoons of characters from The Simpsons, where the children were depicted as engaging in sexual activity. The Supreme Court ultimately ruled that the definition of “persons” for both State and Federal child pornography laws would extend to both real persons and fictional and imaginary depictions of what appeared to be children.

This means that an AI-generated, sexual depiction of someone appearing under 18 years old would likely be an offence.

The federal parliament is currently debating an amendment to the Criminal Code Act 1995 in the House of Representatives to explicitly criminalise the use of technology to generate child abuse material.

This amendment would criminalise the use of a ‘carriage service’, such as the internet, to download, access, supply or offer access to technology where the sole or dominant purpose is the creation of child abuse material. The amendment would also criminalise using a carriage service to collect, scrape, or distribute data if the purpose was training or developing technology to create child abuse material.

If implemented, the amendment could assist in creating strong safeguards against AI-generated child abuse material. It also introduces the concept of training data.

The Role of Training Data

The use of existing child sexual abuse material has been found in the training datasets for AI systems.

In 2023, Stanford released a study that found hundreds to thousands of images of child abuse material in training data for AI. The dataset studied was LAION-5B, a dataset of 5.85 billion CLIP-filtered image-text pairs that has been used to train multiple AI models.

Despite the dataset being taken down after the study was released, it is now available again. The website has a disclaimer that notes that it is ‘uncurated’ and that ‘collected links may lead to strongly discomforting and disturbing content for a human viewer’. Users can filter out certain content with a “safe” subset, but this still does not guarantee the exclusion of harmful content.

Whilst the potential amendments to the Criminal Conduct Act 1995 do not explicitly require using datasets that are free of child sexual abuse material, the eSafety Commissioner lists it as a critical step companies must take to prevent the risk of abuse.

A separate Human Rights Watch report examined 5,850 images from LAION-5B and found 190 images of Australian children. There is no indication that any of these 190 children were found in the context of child sexual abuse material, according to Hye Jung Han, a children’s rights and technology researcher, it could explain why AI can generate content of children extremely realistically.

Regardless, this raises significant concerns about child safety and the risks posed by datasets relying on public web scraping.

A Growing Pattern

OurDream.AI is not the only website built on a model in which users can generate their own partners. Many platforms currently accessible to internet users raise concerns about what other platforms may exist in hidden markets.

Whilst other platforms are not necessarily producing child sexual abuse material, significant data exists to suggest that content is not adequately monitored. A report released by the eSafety Commissioner in October of last year examined four AI programs: Chai, Chub AI, Nomi and Character AI.

Some of the key findings were as follows:

1. None of the platforms had a ‘robust age assurance’ program,
2. Inputs and outputs were not being checked by Chub AI and Nomi, and outputs were not being checked by Chai,
3. Neither Chai nor Nomi had stated that they had reported any child sexual exploitation and abuse material to the NCMEC, and
4. Chub AI and Nomi did not have staff dedicated to ‘trust and safety or moderation’.

Many of these sites claim over 1 million users, with Character.AI valued at over 1 billion dollars.

As the platforms grow, AI-generated child sexual abuse material is likely to continue appearing.

The Internet Watch Foundation stated that 2025 was the ‘worst year on record for online child sexual abuse material’. It recorded a 26,362 per cent rise in photo-realistic AI videos of child sexual abuse, with some videos being of recognisable children.

This spike follows 13 videos found in 2024, followed by 3,440 in 2025. 65 per cent of these videos were Category A content, which can involve sexual penetration and torture.

Proactive Stance

As statistics suggest this issue continues to grow, the eSafety Commissioner says that ‘Technology companies must take a proactive role in embedding safety into the very foundation of their AI systems’. In addition to curating their training data responsibly, the Commissioner requests visible watermarking, strict and enforceable policies for users, transparency, and regular product review.

Nevertheless, it is unlikely that technology companies will adequately make these changes without a legal catalyst. Significant opposition to legislative or regulatory limitations on generative AI tools exists, and includes concerns about ‘stifling innovation; limiting free expression; challenges related to the complexity of technology’ and creating barriers for smaller companies.

However, experts say there is real harm in users engaging with this AI content. Colm Gannon from the International Centre for Missing and Exploited Children (ICMEC) told the ABC that people “become familiarised with childlike identity engaging in sexual talk, and then it actually can cause escalation in harmful behaviour”.

As the investigation into OurDream.AI continues and generative AI continues to develop exponentially, how the issue can be minimised will need to be examined.

Despite current bills in parliament, the Government will likely have to continue to act to fill emerging gaps.