AI is part of how we build The Wagon Method, and we want you to see exactly how it is used. This page explains which AI models we work with, what they do, what they are never allowed to see, and the human checks that happen on the way to your screen.
Our ground rules
- No child PII ever leaves the platform. Child names, dates of birth, and caregiver names are stripped or replaced with opaque identifiers before any AI call.
- No training on your data. Our AI providers are contractually prohibited from using your inputs to train their models, and we do not opt in to any training program.
- Safety validators on every output. Every AI response passes through our safety layer before it reaches you.
- Human editorial review. News articles and published course content are reviewed by a real editor before publication.
- Opt-in. AI-powered features are off by default and turned on only after you read a short consent screen and agree.
- Transparency. Content that was AI-assisted is clearly labeled on the page where you encounter it.
Which models we use
Anthropic Claude
We use Claude for most of the text-generation work on the platform.
- The triage tool, "Where Do I Go From Here?", uses Claude in a tool-calling loop to suggest next steps based on anonymized summaries of the situation you describe.
- The news feed and course drafts are generated with Claude, then reviewed by a human editor.
- Community question classification uses Claude to tag posts with WAGON components so answers surface to the right people.
- Brand voice copy drafts and microcopy suggestions run through Claude with a voice validator scoring the output.
Google Gemini
We use Google Gemini image generation for illustrations across marketing and educational pages. Every generated image is reviewed for brand fit, safety, and accessibility, and is accompanied by descriptive alt text. No child photos, no realistic depictions of trauma, no clinical imagery.
How the safety layer works
Every AI interaction on the platform is wrapped in several layers of protection.
- Sanitization. Before we send any request, a dedicated utility strips child PII, caregiver names, and location data. Names become opaque identifiers. Dates of birth become age ranges.
- Crisis detection. If your message contains language suggesting someone is in danger, we stop and route you to crisis resources instead of calling an AI. Your message is never sent to the model in that case.
- System prompts. Every agent receives a trauma-informed safety system prompt that sets the tone and enforces our content rules.
- Output validation. Once the model responds, a safety validator scans the output for clinical language, deficit framing, forbidden phrases, and unsafe suggestions. Failing outputs are rewritten or replaced.
- Brand voice check. For user-facing copy, a voice validator scores the output against the brand style guide and rejects anything that does not pass.
- Logging. We keep a short audit log of AI interactions for safety review. The log contains no child data, only the prompt class and the validation result.
What a human always reviews
AI speeds us up, but a person still signs off on anything published at scale.
- Every news article, even short ones, is read by a human editor and checked against our content policy before it goes live.
- Course modules and quiz questions are reviewed by a subject-matter editor before release.
- Illustrations are checked for safety, representation, and tone before being saved to the image library.
- Answers from the triage tool are educational guidance, not clinical advice, and carry a clear disclaimer every time.
Your choices
You are always in charge of how much AI assistance you want.
- AI-powered features are opt-in. The first time you open one, you see a short consent screen explaining what is sent, what is returned, and what is not.
- You can turn AI features off at any time from Settings. Turning them off does not affect the rest of the platform.
- You can request that AI-assisted content be labeled in your feed so you always know what a human wrote and what an agent drafted.
- You can give feedback on any AI output with a single click. That feedback is used to improve prompts, never to train the underlying model on your data.
Limits of AI
AI can help draft, classify, and surface content, but it cannot replace a therapist, a caseworker, or the wisdom of your own team. Models make mistakes. They sometimes sound confident about things they get wrong. Please treat any AI suggestion as a starting point, not a prescription, and trust your own judgment about the child in front of you.
Questions and feedback
If something an AI feature produced feels wrong, please tell us. Write to ai@thewagonmethod.com with the details and we will investigate. We take this seriously and adjust quickly when we get it wrong.