Meta is changing the prompts suggested by its artificial intelligence chatbot after a viral video showed the company's AI asking a mother an uncomfortable question about her young child.

The controversy began after Instagram user Kalie Robins shared a video showing an unexpected prompt generated by Meta AI beneath a video she had cross-posted to Facebook.

The suggested question was blunt: “Who is the child passenger?”

For Robins, the issue was not simply that an AI system had analysed the video. It was that Meta's system appeared to be encouraging a user to provide more personal information about a child who was visible in the footage.

The incident quickly spread online and renewed questions about how AI assistants decide what information to ask users about — particularly when children appear in photos and videos.

Meta has now acknowledged that the prompt was inappropriate.

Meta Says It “Missed the Mark”

In a statement to The Verge, Meta spokesperson Dina El-Kassaby said the company “missed the mark”, adding that the feature should never have prompted someone with questions of that nature.

Meta said it is making changes to the suggested prompts generated by its AI system.

That response is significant because the problem appears to have come from a feature designed to make AI interactions more useful.

Rather than waiting for users to ask Meta AI a question, the system can suggest potential prompts based on the content users are viewing or sharing.

The goal is to help people discover ways AI can analyse or interact with their content.

But the Robins incident demonstrates how quickly that convenience can become intrusive.

Why the Question Was So Uncomfortable

Asking who appears in a photograph or video may seem harmless when the subject is a public figure or an adult.

Children are different.

A child's identity, location, relationship to an adult or other contextual information can become sensitive personal information. Parents may deliberately avoid sharing such details publicly even when they post photographs or videos of their children.

That makes an AI-generated question such as “Who is the child passenger?” particularly uncomfortable.

The system was not simply offering to describe what was visible in the video.

It was prompting the user to identify a person.

That distinction matters.

AI Can See More Than Users Expect

The incident also highlights a broader issue with AI-powered features integrated into social platforms.

Modern multimodal AI systems can analyse images and videos, identify objects, describe scenes and interpret relationships between elements in visual content.

Users may understand that an AI chatbot can “look at” an uploaded image, but they may not always realise how much contextual information an AI system could potentially infer or ask about.

When those capabilities are built directly into platforms such as Facebook and Instagram, the interaction becomes even more complicated.

People are already uploading enormous amounts of personal information to social networks.

AI introduces another layer: the platform can now actively interpret that information and suggest questions about it.

Suggested Prompts Can Shape User Behaviour

The controversy also raises an important point about AI interfaces.

A suggested prompt is not the same as a user's original question.

The user may never have considered asking who a child in a video was.

The platform introduced the question.

That means companies designing AI assistants have to think carefully about not only what their models are capable of answering, but also what the interface encourages people to ask.

A system that proactively suggests questions can influence the boundaries of an interaction.

If those suggestions involve sensitive personal information, the responsibility shifts partly to the company designing the system.

Child Safety Makes the Issue More Sensitive

The presence of children adds another layer of concern.

Parents frequently share family moments on social platforms, but that does not necessarily mean they want automated systems encouraging deeper identification or profiling of their children.

The challenge becomes even greater when AI features are integrated into platforms where videos can be reshared, recommended or viewed by large audiences.

Companies therefore have to consider more than whether an AI response is technically accurate.

They also need to ask whether the interaction is appropriate, necessary and expected by the user.

In this case, Meta's own response suggests that the prompt failed that test.

The Bigger Problem With AI Personalisation

The incident is part of a wider challenge facing technology companies as they make AI more proactive.

For years, digital assistants largely waited for users to give commands.

The new generation of AI products is increasingly designed to anticipate what users might want.

That can be useful.

An AI assistant might suggest summarising a long document, organising photographs or drafting a response to a message.

But anticipation also creates opportunities for overreach.

The more context an AI system has about a person, the more personalised its suggestions can become — and the greater the risk that those suggestions cross a privacy boundary.

Meta's Response Will Be Watched Closely

Meta's decision to change the prompts is a reminder that AI safety does not only happen at the model level.

It also happens at the product-design level.

A model may be capable of analysing a child in a photograph without necessarily needing to encourage users to identify that child.

The question therefore becomes not simply:

“Can the AI do this?”

but:

“Should the product encourage people to do this?”

That distinction is likely to become increasingly important as AI becomes more deeply embedded in social networks.

The most worrying part of this incident isn't that Meta AI could recognise something in a video. It's that the AI decided to turn a child's identity into a suggested conversation.

AI companies increasingly want their assistants to be proactive, personalised and context-aware.

But being proactive does not mean asking everything the system can think of.

The Robins incident shows why privacy boundaries need to be built into the experience, not simply added after something goes viral.

For users, it is another reminder that uploading something to a social platform can expose it to AI systems capable of analysing far more than the original post was intended to communicate.

And for Meta, the challenge is now to make its AI feel helpful without making users wonder what information it is quietly trying to extract from their content.