Responding to disclosures of child sexual exploitation and abuse through AI systems

A disclosure of child sexual exploitation and abuse (CSEA) is an interaction in which someone describes or indicates sexual abuse or exploitation experienced during childhood. The person may describe their own experience, raise concerns about someone else, or leave unclear whose experience they are describing. 

Disclosures may be explicit, indirect, gradual or later retracted. Someone does not need to name their experience as abuse for an AI system to respond with care. This guide is designed primarily for AI systems that interact with users through conversations, but the principles also apply where a disclosure arises through other products, such as image generation tools or search systems. 

How this guidance applies in practice will vary depending on the model or product, its role within the AI stack, its technical capabilities and constraints, and the context in which it is used.

These six core principles apply to every response to a possible CSEA disclosure, whether it comes from the model, the product or the company.

1. Take harm seriously, without blame or doubt.
   Respond supportively and without judgment. Convey belief as support for the person, without judging what happened. Responsibility stays with the person who caused the harm.
2. Let the person lead.
   Do not push for details, labels or a complete account. Do not push for details, labels or a complete account. Match their pace and begin the response within their chosen frame. Mirror the person’s language as much as possible and do not introduce terms or labels for actions or behaviors the person is describing.
3. Ask only what safety and support require.
   One non-leading question at a time, and only when the answer could change the support or next step. Do not investigate, gather evidence or seek explicit detail.
4. Be calm, clear and honest.
   Use plain, non-stigmatizing language that never implies a child consented or was responsible. Be honest about what the product can do, how information may be handled and what cannot be promised.
5. Help the person reach safe human support.
   Offer suitable, accessible support in their location and language. Build toward human help without making the AI system a substitute, and do not assume any adult or service is safe.
6. Avoid further harm, including from data.
   Do not encourage confronting the person causing harm. Do not produce or reproduce sexual content involving a child. Do not assume a device, account or notification is private. Collect, keep and share only what safety and the law require.

In practice, disclosures to AI take many forms: explicit or partial, first-person or about someone else, current or historical, from a child or an adult.

The guide’s taxonomy sets out eight dimensions for describing a possible disclosure:

• Whose experience is being described.
• The age of the person disclosing and the person harmed.
• How the person conveys the possible harm.
• What the person is trying to do or seeking from the interaction.
• The type of harm involved.
• The timing and source of any current risk.
• How the disclosure develops across the interaction.
• Any material or identifying information shared.

Not all of this information may be available to companies. Any of the dimensions may be unknown or unclear in the course of a disclosure. 

The guide also identifies six decision points: 

• Recognizing a possible disclosure
• Responding
• Triaging possible risk
• Assessing the situation through human review
• Considering external action
• Communicating what happens. 

These are not a fixed sequence. Depending on the product and circumstances, decisions may overlap, recur or not apply.

Together, the taxonomy and decision points give companies a shared way to understand a possible disclosure, find a starting point for the response, and safely tailor the response to the person.

Companies should test both whether an AI system recognizes possible disclosures and how it responds. Build test sets that reflect how people disclose. Test sets can include:

• explicit, indirect, gradual and retracted disclosures, including across many turns and repeat sessions
• users who indicate they are under 18, adults, and users of unknown age
• different countries and languages, checking that the services named exist and can be reached where the person is
• co-occurring risks such as self-harm, and material such as screenshots
• fiction, roleplay and other requests that move into prohibited content

Companies may consider involving specialists in safeguarding, child helplines and survivor support when designing scenarios and reviewing outputs, and updating test sets as new patterns emerge.

Download the report