Our children's photos are training AI

It starts with a photo on the school Facebook Page

Every week, at thousands of schools across Australia, a teacher picks up a phone and takes a photo. Sports Day. Book Week. A classroom activity. A group of kids holding up artwork. The teacher uploads it to the school's Facebook Page with a caption about how proud everyone is. Parents see it in their feed and smile. Grandparents share it. The school community feels connected.

This has been happening for over a decade. It is one of the small, good things about school life. It is the reason schools set up Facebook Pages in the first place: to keep families in the loop, to celebrate what children are doing, to build community.

But the world that photo enters in 2026 is nothing like the world it entered in 2016. And almost none of us were told about the change.

What happens to a photo on a public Facebook Page

A school Facebook Page is public. In NSW, Department of Education policy PD-2011-0418 requires it. The policy states that school social media accounts "must not restrict access or be set as 'private' or 'closed'." That means every photo is visible to anyone on the internet, with or without a Facebook account. It is indexed by Google. It can be downloaded by anyone. And it is accessible to every automated system that scrapes public web content at industrial scale.

That last point is the one that changed everything.

Meta told the Australian Senate it scrapes every public post to train AI

On 11 September 2024, Meta's global privacy director Melinda Claybaugh appeared before the Australian Senate Select Committee on Adopting Artificial Intelligence. Under questioning from Greens Senator David Shoebridge, she confirmed that Meta has been scraping every public Facebook and Instagram post from Australian users since 2007 to train its AI models, including Llama and Meta AI.

Every public photo. Every public status update. Every public comment. Nearly two decades of content from 19 million Australian users. All of it fed into machine learning systems that learn to recognise, generate, and manipulate visual content.

When Shoebridge asked whether photos of children posted on adult accounts are included, Claybaugh confirmed they are. Meta excludes data from accounts belonging to users under 18, but a school Facebook Page is not a child's account. It is an organisational account run by adults. Every photo of every child on every public school Facebook Page is available to Meta for AI training.

"Meta has just decided that you will scrape all of the photos and all of the text from every public post on Instagram or Facebook that Australians have shared since 2007, unless there was a conscious decision to set them on private." Senator David Shoebridge, Australian Senate Select Committee on Adopting AI, September 2024

When Shoebridge asked Claybaugh directly to confirm the 2007 start date, she responded with a single word: "Correct."

Senator Tony Sheldon pressed on whether parents would have any idea this was happening. He noted that parents don't typically read privacy policies and would have no knowledge that photos of their "13-year-old daughter" or "nine-year-old son" were being harvested to train AI systems. Claybaugh did not dispute the point.

Europeans can opt out. The General Data Protection Regulation (GDPR) forced Meta to provide an opt-out mechanism and pause launching AI products in Europe while regulatory uncertainty persisted. Australians have no equivalent protection. When asked why, Claybaugh was direct: "The specific option that we're offering in Europe is in response to a very specific legal framework." In other words, Meta will only offer privacy protections where the law compels it. Australia's laws don't. So our children get nothing.

Clearview AI scraped 50 billion photos for facial recognition

Meta is not the only company harvesting children's photos from public Facebook Pages. Clearview AI, a US-based facial recognition company, has scraped more than 50 billion photos from Facebook, Instagram, LinkedIn, and other public sources to build the world's largest known facial recognition database. The company has sold access to this database to law enforcement agencies across the world. It has been searched nearly a million times.

Children's faces are in the database. Any face that appeared on a public Facebook Page is in the database.

In November 2021, the Australian Information Commissioner found that Clearview AI breached the Australian Privacy Act 1988. The Commissioner found that Clearview had collected Australians' sensitive biometric data without consent and ordered the company to destroy all data collected from Australians and cease collecting it. Regulators in France, Italy, the United Kingdom, and Greece imposed fines exceeding $100 million.

None of the fines have been collected. Clearview has not demonstrated compliance with the deletion orders. The database has grown since the orders were issued, from 30 billion images at the time of the OAIC finding to more than 50 billion today.

The consequence for children is permanent. Once a child's face is in a facial recognition database, every future photo of that child, taken anywhere, can be matched back to their identity. A single school photo creates a biometric anchor that enables lifetime trackability. It does not expire. It cannot be revoked.

362 Australian children found in an AI training dataset

In July 2024, Human Rights Watch published findings on LAION-5B, one of the largest open AI training datasets in the world. LAION-5B contains 5.85 billion image-text pairs scraped from the public internet. It has been used to train Stable Diffusion and other generative AI models.

Researchers examined less than 0.0001% of the dataset. In that tiny sample, they found 362 identifiable Australian children. The children's names were visible. Their ages were visible. Their school names were recorded in the metadata. Some of the photos came from school uploads. Some came from personal blogs and photographers' websites. Some of the original photos had already been taken down by the time the researchers found them in the dataset. It didn't matter. Once scraped into the dataset, the data cannot be removed.

LAION's public response was to blame parents for making photos "publicly available." The organisation took the dataset offline temporarily but has since made it accessible again.

The scale implied by the sample is staggering. If 362 Australian children appeared in less than 0.0001% of the dataset, the total number of identifiable Australian children in LAION-5B is likely in the millions.

Separate research from the University of Utah and Carnegie Mellon University analysed 18 million Facebook posts from school Pages and found that 726,000 posts included students' first and last names alongside photos and location data. The researchers concluded that schools represent "the largest existing collection of publicly accessible, identifiable images of minors" on the internet.

Consider what that means. Not social media profiles created by teenagers. Not family holiday albums posted by parents. Schools. Official school accounts, posting in an official capacity, under policies that require them to be public.

Deepfakes are already targeting Australian schoolchildren

In June 2024, AI-generated explicit images of more than 50 female students at Bacchus Marsh Grammar, a school in Melbourne, were created using the students' publicly accessible photos and circulated online. The girls were in Years 9 to 12. In February 2025, a similar incident was reported at Gladstone Park Secondary College, also in Melbourne. In January 2025, reports emerged from schools in Sydney.

The images were generated from ordinary photos. The kind of photos that appear on school Facebook Pages every week. A school Page makes this particularly efficient: the photos are high quality, children are often identified by name in the caption, the Page is fully public, and there are multiple photos of each child over time, giving AI models more material to work with.

The eSafety Commissioner reports that deepfake reports are doubling year on year. Victoria has criminalised deepfake pornography, with a maximum penalty of seven years' imprisonment. But criminal penalties apply after harm has occurred. They don't prevent it.

Nobody consented to any of this

When parents enrol their child at a NSW public school, they are given a "Permission to Publish" form. It offers a single binary choice: consent to all public publishing (website, newsletter, Facebook, newspapers, promotional materials) or consent to none of it. There is no middle ground.

There is no option to say "yes to the school newsletter, no to public social media." There is no mention of AI training. No mention of facial recognition. No mention of data scraping. No mention of deepfakes. No mention of third-party companies downloading and storing the photos permanently. The form was written before any of this existed.

The consent has no expiry date. It remains in effect "until I advise the school otherwise." A form signed in good faith in 2015, when Facebook was a place to share updates, is still being relied upon in 2026, when Facebook is a source of training data for AI systems and facial recognition databases. The child whose parent signed that form has no voice in the matter. They were never asked.

The irony at the centre of all of this

In December 2025, Australia's Online Safety Amendment (Social Media Minimum Age) Act took effect. Platforms are now required to take reasonable steps to prevent under-16s from holding accounts. Meta, TikTok, Snapchat, Reddit, and X removed a combined 4.7 million accounts belonging to Australian minors. The law passed with bipartisan support. As a country, we decided that social media is not safe for our children.

But our schools are still publishing identifiable photos of those same children on those same platforms. Every week. On Pages that are public by policy. Accessible to anyone on the internet. Scraped by the very systems we just decided children need protection from.

The law protects children from being on social media. Nothing yet protects them from being published on it by the institutions entrusted with their care.

NSW requires public accounts. Victoria doesn't.

The NSW Department of Education's social media policy, PD-2011-0418, requires school Facebook accounts to be public. The policy explicitly states accounts "must not restrict access or be set as 'private' or 'closed'." Schools have no discretion. A principal who wants to protect students by switching to a private account is in breach of departmental policy.

Victoria has no equivalent requirement. Victorian schools can and do use private Facebook Groups to share school photos with families. The experience for families is identical: the same photos, the same updates, the same community connection. The only difference is that strangers, scrapers, and AI companies cannot access the content.

One administrative policy update in NSW would close the gap. No legislation required. No budget allocation needed. One change to PD-2011-0418, allowing or requiring private settings, would immediately reduce the exposure of every child at every public school in the state.

What needs to change

First: NSW must update its school social media policy. PD-2011-0418 was written in 2011, before Meta confirmed AI scraping, before Clearview AI built a 50-billion-photo facial recognition database, before Human Rights Watch found Australian children in AI datasets, before deepfake attacks on Australian schoolgirls. The policy should allow or require private settings for school social media accounts. The Permission to Publish form should disclose AI risks and offer granular consent. Schools should receive guidance on the current threat landscape.

Second: P&C associations should formally recommend their schools act now. Every P&C in NSW can pass a motion recommending their school move its Facebook presence from a public Page to a private Group. This is within the P&C's role, it is free, and it can happen at the next meeting. It does not require policy change from above. It is a recommendation from the parent community to the school, backed by evidence.

Third: Australia needs a national framework for children's images and AI. The Children's Online Privacy Code is due by December 2026 and must explicitly address children's images in AI training datasets. The Privacy Act reform must establish that public availability does not constitute consent to AI training. The regulatory gap between what companies can do with children's photos and what parents believe is happening must be closed.

Where regulation is heading

The direction is unmistakable. The Social Media Minimum Age Act passed with bipartisan support and near-universal public approval. The Children's Online Privacy Code is due by December 2026, with binding obligations on companies that collect or process children's data. The Privacy Act is undergoing its most significant reform in decades, including a proposed statutory tort for serious invasions of privacy that would give individuals the right to sue.

Schools that act now are ahead of what is coming. Schools that wait will eventually be required to make changes anyway, after years of unnecessary exposure, without the benefit of having acted on their own terms when the evidence was clear.

Every school that acts makes it harder for the policy to stay unchanged

Photos that have already been scraped cannot be unscraped. AI models that have already trained on children's faces cannot unlearn them. That harm is done. But every day a school keeps posting children's photos on a public Facebook Page, the exposure grows. New photos are scraped. New faces enter facial recognition databases. New source material becomes available for deepfake generation.

The first step is practical, free, and immediate: move the school's Facebook presence from a public Page to a private Group. A private Group removes children's photos from public web access. Third-party scrapers can no longer reach them. Search engines can no longer index them. Strangers can no longer browse or download them. The school community keeps everything it values. The only difference is that someone needs to be approved to join.

This does not stop Meta entirely. Meta still has access to content posted within its own platform, including private Groups. Getting off Facebook entirely is the more complete step. But the private Group is the first meaningful action a school community can take, and it costs nothing.

Every school that acts protects its children. Every school that acts creates a precedent that other schools can follow. Every school that acts makes it harder for the NSW Department of Education to leave PD-2011-0418 unchanged. And every school that acts sends a clear signal: parents looked at the evidence, and they decided this was not acceptable.