> ## Documentation Index
> Fetch the complete documentation index at: https://developers.reflection.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Child Safety Developer Guidance

export const Plain = ({children}) => <span style={{
  fontWeight: "normal"
}}>{children}</span>;

## Overview

Reflection is committed to the principle that the power of the most consequential technology of our time should be broadly shared. That opportunity brings a shared responsibility to address the misuse of AI, especially AI-facilitated child sexual abuse and exploitation (CSAE). Combating online CSAE, particularly where AI is involved, can raise complex legal questions and involve varied abuse patterns. Doing it well often takes specialized knowledge of the space that smaller development teams may not have in-house.

### What we built into Beam

Before releasing Beam’s weights, we took a number of measures relating to child safety, including the following:

* **Policies**. Our policies outline prohibited uses of our model, and include a specific Child Safety section covering attempts to create or facilitate CSAE. These policies define prohibited uses of our model and the standards Beam is designed to meet.
* **Evaluation**. We ran evaluations of Beam covering text-based CSAE harms that are lawful to test, non-sexual harms to minors, and benign and protective uses that should not be blocked, using single- and multi-turn prompts developed with input from child-safety and red-teaming specialists.

The results of our child safety evaluations are reflected in the [<Plain>Safety Report</Plain>](https://reflection.ai/beam-safety-report).

### Why developers need their own measures

Open-weight models give developers more control and flexibility, and more responsibility. Once a model’s weights are downloaded, the model runs entirely under the deployer's control. That is why anyone who fine-tunes, deploys, or redistributes an open-weight model needs to add the safeguards that only they are positioned to provide: detection on their own traffic, enforcement against their own users, and, depending on where and how they operate, their own reporting and preservation obligations. A model’s built-in safety training is an important layer of protection, but it is only one layer. No model-level safeguard is likely to be fully resistant on its own, which is why we treat child safety as a property of the whole deployment ecosystem, not just the model.

This guidance is a foundational overview and introduction to the Child Safety space for developers building on, modifying, or deploying open-weight text-to-text models. It draws on established guidance, including [<Plain>Thorn and All Tech Is Human’s Safety by Design for Generative AI principles</Plain>](https://www.thorn.org/blog/generative-ai-principles/) and [<Plain>best practices</Plain>](https://technologycoalition.org/resources/developer-good-practices-combating-online-child-sexual-exploitation-and-abuse/) developed by leading domain experts and specialized organizations like the [<Plain>Tech Coalition</Plain>](https://technologycoalition.org/), [<Plain>NCMEC</Plain>](https://www.missingkids.org/home), and [<Plain>IWF</Plain>](https://www.iwf.org.uk/). Much of the existing guidance was written with hosted models in mind, so we’ve focused on the practices that matter most when the model is in the developer’s hands. This is not an exhaustive checklist of child safety mitigations, nor is it intended to act as legal advice; requirements vary by jurisdiction, product and audience. We strongly recommend tailoring it to each use case and consulting with legal counsel when implementing the practices described.

## How Text-Based CSAE Shows Up in Generative AI

Discussion of AI-facilitated CSAE often centers on imagery, but text models carry their own risks. Generative AI lets bad actors produce harmful text at scale, personalize it to a specific child, and skip the skill or effort that used to be a barrier. Text-based harms in generative AI can include:

* **Sexualized content involving minors.** Explicit stories, roleplay or "fiction" involving real or synthetic minors. This includes prompts that tell a chat model to act as a child and engage in sexually explicit chat, often by assigning the model a persona meant to override its safeguards.
* **Grooming and sextortion at scale.** Fake identities and trust-building messages that let offenders groom more children, increasingly across languages and platforms. In sextortion, offenders have used AI-generated explicit images of a child to coerce them into sending more content or money, which could be scaled further with text models. When connected to tools or agent workflows, models could also be misused to automate contact with children.
* **Offender guidance and pathways to CSAM.** Guides and tutorials on how to groom or abuse children, coerce them, destroy evidence, or keep victims from telling anyone. This also includes prompts written to get image models to generate abuse material, tips for getting around safeguards, and requests for where to find CSAM.
* **Trafficking.** Automated recruitment messages and deceptive offers aimed at vulnerable people.

## Recommendations to safeguard against CSAE

The practices outlined below are the deployment-side supplement to some of the model-side work described above.

### Lay the Foundation

Every developer building on an open-weight model should have these in place:

* ***Map your risks and get specialized guidance.*** Using the text-based harm examples above, identify how your use case, audience and features could be misused. Laws and regulatory guidance include specific parameters for text-based harms, and applying them to generative AI takes expert interpretation. Legal counsel and policy professionals with expertise in this area can help you set content classification, clarify reporting obligations, and define evidence preservation and red-teaming limits.
* ***Write clear policies that separate harm from protective use.*** Clearly define what is prohibited across training, evaluation, model behavior, and user-facing use, and make sure your usage policies name CSAE explicitly. Also define what should *not* be blocked. Some people who raise CSAE topics are trying to protect children or get help: survivors disclosing abuse, parents worried about grooming, educators, clinicians and researchers. Decide in advance how your product should respond to disclosures, such as pointing to support services.
* ***Set up reporting.*** U.S.-based electronic service providers have legal obligations to report apparent CSAM to the National Center for Missing & Exploited Children (NCMEC), the U.S. clearinghouse for reports of child sexual exploitation. Register as an electronic service provider to get access to its CyberTipline, ideally before you start training or deploying, so the process is ready when you need it. For text-based harms, context is crucial, so plan to include enough of the surrounding conversation for a report to be actionable, while collecting no more personal data than you need.
* ***Join the child safety community.*** You don’t need a large trust and safety team to start. The Tech Coalition’s free [<Plain>Pathways</Plain>](https://technologycoalition.org/programs/pathways/) program supports startups and smaller companies, and its members can use resources such as a template for reporting AI-related CSAM to NCMEC. Thorn, IWF, WeProtect, ROOST and INHOPE also offer valuable resources.

### If you fine-tune or modify a model

Modifying a model can change its safety behavior, sometimes without anyone intending it. Incorporate these best practices to help mitigate any risks:

* ***Screen your fine-tuning and training data.*** If you fine-tune or further train a model, screen your data sources for CSAM, whether you collect the data yourself or buy it from a vendor. Use text classifiers (see “If you deploy a model” below) to screen for CSAE content in text data, and tools such as [<Plain>Microsoft’s PhotoDNA</Plain>](https://www.microsoft.com/en-us/photodna) or [<Plain>Google’s Child Safety Toolkit</Plain>](https://protectingchildren.google/tools-for-partners/) for any image data. Include safeguarding requirements in contracts with data vendors. Any CSAM should be reported to NCMEC. Document your screening process and the tools you use.
* ***Evaluate and re-evaluate after every modification.*** Fine-tuning, quantization, model merging, distillation and system prompt changes can all alter a model’s safety behavior. Before deploying any version, test it against both harmful and benign scenarios within legally permissible limits, and measure over-refusal as well as harmful outputs.

### If you deploy a model

Safeguards need ongoing attention after launch. During and after a launch, maintain and strengthen them through the following:

* ***Add system-level detection.*** Don’t rely on a model’s built-in refusals alone. Set up proactive detection mechanisms to identify misuse. A layered approach that combines established tools with in-house classifiers can be useful. Examples of text-based proactive detection tools include Thorn’s Safer [<Plain>AI text-based classifier</Plain>](https://get.safer.io/text-classification-content-moderation) and HIVE’s [<Plain>text detection offerings</Plain>](https://thehive.ai/blog/hive-adds-thorns-grooming-detection-to-cse-text-classifier-api). Other helpful resources are CSAM keyword and URL lists offered by [<Plain>IWF</Plain>](https://www.iwf.org.uk/our-technology/our-services/) and [<Plain>Thorn</Plain>](https://safer.io/resources/csam-keyword-hub/). Text-based abuse often builds gradually, and no single message may cross a line on its own. Where possible, run classifiers over conversation history and flag patterns such as repeated rephrasing after a refusal.
* ***Launch in stages.*** A staged rollout lets you watch real-world behavior and catch gaps before broad release.
* ***Set up user reporting and human review.*** Give users a clear, easy way to flag harmful outputs or misuse. Trained reviewers with subject matter expertise, supported by wellness resources, should review flagged content and decide on next steps, including reporting where required. Open-source tools can help teams without hosted infrastructure: [<Plain>ROOST's</Plain>](https://roost.tools/) Coop is a review console that supports filing reports to NCMEC’s CyberTipline, and Osprey is a rules engine for detecting abuse patterns. Work with legal counsel on escalation protocols and data retention.
* ***Close the loop.*** Use what you learn from user reports, flagged content, and review outcomes to update your classifiers, thresholds, policies, and evaluations. Treat every model or product update as a new launch for child safety purposes.

## Keep learning

Text-based CSAE keeps changing as bad actors adapt. Stay informed through expert organizations and hotlines, such as INHOPE's webinars, NCMEC's Generative AI [<Plain>issues page</Plain>](https://www.missingkids.org/theissues/generative-ai) and congressional testimony, and through threat [<Plain>intelligence reports</Plain>](https://go.alice.io/child-predator-abuse-of-generative-ai).

Making safety behavior resistant to removal is an open research problem across the industry; no developer has solved it. We have started initial research into making safety behavior more resistant to removal. We are committed to collaborating with researchers, child safety organizations and other model developers on these questions and to testing novel methods to safeguard against these known limitations.

### Additional Resources

* [<Plain>NCMEC Senate testimony, March 2025</Plain>](https://www.judiciary.senate.gov/imo/media/doc/2025-03-11_testimony_delaune.pdf)
* [<Plain>NCMEC: Growing Concerns of Generative AI</Plain>](https://www.missingkids.org/blog/2024/the-growing-concerns-of-generative-ai-and-child-sexual-exploitation)
* [<Plain>Alice’s Child Predator Abuse of Generative AI Report</Plain>](https://go.alice.io/child-predator-abuse-of-generative-ai)
* [<Plain>Global Initiative Against Transnational Organized Crime, July 2026</Plain>](https://globalinitiative.net/analysis/abuse-at-scale-artificial-intelligence-is-reshaping-child-sexual-exploitation/)
* [<Plain>UNODC, "How AI is fuelling human trafficking," Sept 2026</Plain>](https://www.unodc.org/unodc/en/frontpage/2026/September/how-ai-is-fuelling-human-trafficking.html)
* [<Plain>IWF AI CSAM Report 2026: Harm without limits</Plain>](https://admin.iwf.org.uk/media/hl1nvdti/iwf-ai-csam-report-2026.pdf)

## Key Terms

* Child sexual abuse (CSA): “the involvement of a child \[anyone under the age of 18] in sexual activity that he or she does not fully comprehend, is unable to give informed consent to, or for which the child is not developmentally prepared and cannot give consent.”<sup>[<Plain>1</Plain>](#fn-1)</sup>
* Child sexual exploitation (CSE): “situations where a perpetrator or third party coerces a child into sexual activities in exchange for something of value (like money, food, shelter, or promises). This can happen through direct force or threats, manipulation or grooming, power imbalances, and more.”<sup>[<Plain>2</Plain>](#fn-2)</sup>
  * While the two concepts often overlap, child sexual exploitation is different than child sexual abuse because, in the case of CSE, it usually involves some kind of exchange.
* Online child sexual abuse and exploitation (CSAE) occur “when these forms of abuse are partly or entirely facilitated by technology” (e.g., the internet, wireless communications, generative AI models).<sup>[<Plain>3</Plain>](#fn-3)</sup>
* Child sexual abuse materials (CSAM): any sexually explicit content involving a minor.<sup>[<Plain>4</Plain>](#fn-4)</sup>
  * AI-generated CSAM (AIG CSAM): media entirely created by a model, ranging from photorealistic to non-photorealistic (e.g., cartoon, anime). AIG CSAM can depict an identifiable individual, also known as a deepfake, or be a fully synthetic depiction.
* Grooming: “the act of deliberately establishing trust by an adult through forming an emotional connection with a minor with the intention of manipulating and carrying out sexual abuse or exploitation.”<sup>[<Plain>5</Plain>](#fn-5)</sup>
* Sextortion: “a form of online enticement where a child is blackmailed, most often with a threat that nude or sexual imagery of them will be shared publicly or with family and friends, by a person who demands that the child create and send additional exploitative or abusive material, meet in person to engage in sexual activity, or pay the blackmailer a sum of money.”<sup>[<Plain>6</Plain>](#fn-6)</sup>
* Child sex trafficking: “a form of child abuse that occurs when a child under 18 is advertised, solicited or exploited through a commercial sex act. A commercial sex act is any sex act where something of value is given to or received by any person for sexual activity.”<sup>[<Plain>7</Plain>](#fn-7)</sup>

***

1. <span id="fn-1" />[<Plain>WeProtect: The Issue</Plain>](https://www.weprotect.org/issue/)
2. <span id="fn-2" />[<Plain>Thorn’s What is CSAM: Child Safety Common Terms and Definitions</Plain>](https://safer.io/resources/common-terms-and-definitions/)
3. <span id="fn-3" />[<Plain>WeProtect: The Issue</Plain>](https://www.weprotect.org/issue/)
4. <span id="fn-4" />[<Plain>Thorn’s What is CSAM: Child Safety Common Terms and Definitions</Plain>](https://safer.io/resources/common-terms-and-definitions/)
5. <span id="fn-5" />[<Plain>Thorn’s What is CSAM: Child Safety Common Terms and Definitions</Plain>](https://safer.io/resources/common-terms-and-definitions/)
6. <span id="fn-6" />[<Plain>NCMEC’s Guidelines on Identifiers of Online Enticement and Child Sex Trafficking</Plain>](https://www.missingkids.org/content/dam/missingkids/pdfs/NCMEC-REPORT-Act-Guidelines.pdf)
7. <span id="fn-7" />[<Plain>NCMEC’s Guidelines on Identifiers of Online Enticement and Child Sex Trafficking</Plain>](https://www.missingkids.org/content/dam/missingkids/pdfs/NCMEC-REPORT-Act-Guidelines.pdf)