_Part 16 of 18 in the Digital privacy for board directors series. The fifth and last of the children-focused posts._

When I drafted the children-focused posts in 2023, the AI section was a short paragraph in the digital-footprint post. Eighteen months later, it warrants a post of its own. Generative AI has changed what an image of a child can mean, what evidence can mean, and what parents should be alert to in a way that the existing safeguarding literature has only begun to catch up with.

I am writing this in 2024, with the children-AI conversation now visible in mainstream media but still working out its practical shape. Some of what I say here will age. Most of the structural points will not.

What is genuinely new

Three things.

Image generation that produces fake images of real children, at scale, from a small training set. A determined adversary, with as few as ten photographs of a child from public social media, can now generate plausible-looking false images of that child in any setting, including settings that are deeply harmful. The technical capability is widely available, the cost is near-zero, and the speed is near-instant. This was a research-grade capability in 2022; it is a commodity capability in 2024.

Voice generation that produces convincing fake audio from short samples. A few seconds of a child's voice — extracted from social media, gaming voice chat, school assembly recordings — is sufficient to clone the voice well enough to fool family members. The grandchild kidnapping scam has been adapted to use this. The voice on the phone is the child's voice. The instruction give me the money is not theirs.

The school-friend deepfake. A pattern that has emerged in UK schools through 2023 and 2024: a teenager generates deepfake intimate images of a classmate using publicly-available photographs and an AI tool. Distribution is among the school year. The harm to the depicted child is severe. UK law is catching up but the criminal offence around the creation of such images is still being worked through.

Each of these is new in the sense that the practical accessibility of the capability is new. The principles of safeguarding are not new. The actions parents and children can take are recognisable from the earlier children-focused posts in this series. The thresholds have changed.

What has not changed

A useful counterpoint. Several things parents worry about with AI have not really changed.

The information available to an adversary. A determined adult with bad intent has always been able to find ways to harm children. AI changes the speed and the type of artefact, not the underlying motivation. The protections that worked before — the relationship, the conversation, the platform settings — still work.

The school's safeguarding role. Schools have processes for the disclosure of harmful imagery, the involvement of CEOP, and the support of children affected. The processes are imperfect, but they exist and they handle AI-generated material under the same broad categories as photographic material.

The platform takedown process. Mainstream platforms — TikTok, Snapchat, Instagram, Discord — respond to reports of child sexual abuse material regardless of whether the material is photographic or AI-generated. The reporting routes have not changed. The volume the platforms are responding to has.

The legal options. The Internet Watch Foundation and the NCMEC Take It Down service accept reports of AI-generated intimate imagery of minors under the same processes as photographic material. The Childline Report Remove tool accepts the same.

What the conversation with the child looks like

I am going to be specific because the alternative is to be vague at exactly the point where vagueness fails.

For pre-teens and early teens, the conversation has two parts.

The image-of-you part. People can now make pictures of you doing things you have not done. If you see a picture of you that you know is not real, or a friend sees one, the action is to come and tell us. You will not be in trouble. We will deal with it. The picture is not what people will remember; how we deal with it is. The reassurance that coming to the parent is the right move is the protection. The frequency with which children do not come to parents because they are afraid of being blamed is the harm we are trying to prevent.

The voice-of-someone-you-know part. If you get a call or a message that sounds like grandma or grandpa or someone in the family, and they are asking for something unusual — money, a meeting, a password — even if the voice sounds exactly right, the rule is to call back on the number you already know. Voices can be faked now. Numbers are harder to fake.

For older teens, the conversation adds a third part.

The friend-of-yours part. If a friend of yours is the target of a deepfake image, the action is to stop the spread and to support them. The action is not to share it ‘just so people know’. The action is also not to keep it on your phone ‘for evidence’. The route is to tell a parent, a teacher, or the platform. The platforms will take it down.

These conversations are awkward. They are also short, and once had, they sit in the relationship.

What parents can do operationally

Five things, in addition to the conversations.

One: thin out the public footprint. The advice from post four — public-share versus family-share, no children's faces in public posts, no first names in captions — is the substantive protection against image-generation targeting. The training set the adversary uses is whatever they can collect. The smaller it is, the worse the result.

Two: enable platform-side protections. TikTok, Instagram, and Snapchat all have features that limit who can take screenshots or save content involving the child's account. Enable them. They are imperfect; they raise the bar.

Three: agree the family code-word for voice verification. A short phrase known only to immediate family that, on a phone call, confirms the caller is who they claim to be. Grandma needs help, what is the code-word? — if the caller cannot answer, the call ends and the family verifies separately. This sounds silly and works.

Four: know the reporting routes. Childline Report Remove for under-18s. The IWF reporting portal. CEOP for grooming or serious concerns. The routes work; the friction is being unsure where to start.

Five: keep updating the conversation. The AI capability set is changing faster than any single guidance document can keep up with. The conversation with the child should be revisited every few months rather than treated as a one-off.

The harder thing about deepfake harm

A note worth being honest about. The published guidance on AI-generated harm tends to focus on the content — the image, the audio, the platform takedown. The lived harm is, in many cases, the social experience of being the depicted person while the content is in circulation. Friends, classmates, teachers may see the material before takedown. The takedown does not undo the seeing.

The protection here is the same protection as for any form of bullying: the relationship, the school's pastoral team, the child's friends who close ranks. Practically, that means knowing in advance which teacher at the school is the right one to contact, knowing the school's policy on this kind of incident, and being the parent whose child knows the parent is on their side.

I cannot fix this with technical advice. I can only say that the parents who have had this conversation in advance handle the actual incident, when it arrives, dramatically better than the parents who have not.

What this month looks like

Two short pieces of work.

One: the conversation with each child you have, of an age where it is appropriate. Twenty minutes. The three parts above.

Two: the family code-word. Pick one. Tell everyone. Use it once, casually, in the next week so it is normal.

In five weeks: the synthesis post. Pulling everything in the series into a posture, not a list.