Understanding the Cost
"Can you help me write this essay?"
"What's the best recipe for chicken?"
"Tell me a joke about dinosaurs."
AI is no longer sci-fi. It's in your kitchen, your classroom, and your child's pocket. It's helpful, fast, and incredibly convenient, but there's a hidden cost. Every time you type a prompt into a public AI tool, you're feeding data to a machine that learns from you. And that data might include things you never intended to share.
The business model is clear: the more you talk, the smarter the machine gets, and the more valuable the product becomes. The question isn't "Is AI useful?" The question is: "Who owns the conversation?"
A quick note on language. When I say "AI" in this article, I'm talking specifically about generative AI—the large language models (LLMs) like ChatGPT, Gemini, and Claude that generate text, images, and code from your prompts. The term "AI" has been stretched so far that it now covers everything from your spam filter to your thermostat. That vagueness isn't accidental. When everything is "AI," the specific risks of generative models—like training on your private data—get blurred into the background. So let's be precise: we're talking about the systems that learn from what you type.
Your Prompts Are Training Data
Here's the most important thing to understand: public AI models—like the free versions of ChatGPT, Gemini, and Copilot—are not private.
When you type a prompt, it's stored. The company saves your conversation on their servers and it's reviewed. Human contractors often read these chats to "improve" the model through a process called RLHF (Reinforcement Learning from Human Feedback). Your input is used to train future versions of the AI, meaning your family's private and personal thoughts and secrets become part of the public model's knowledge base.
What about paid plans? The uncomfortable truth is that even paid plans on mainstream platforms like ChatGPT Plus or Claude Pro still save and analyze your chat history. They may opt you out of training on your data, but they still store it on their servers. And they still have the technical ability to read it.
The real differentiator is zero-access encryption. With tools that use this approach—like Lumo, built by Proton—your data is encrypted before it leaves your device. The company can't read your prompts, they can't sell your data and they can't hand it to governments. They don't have the keys. This is the difference between "we promise not to look" and "we physically cannot look."
But here's the question I get most often: If Proton can't see your data, how does Lumo get smarter? Doesn't that mean it's less capable?
The answer is no. Lumo's core intelligence comes from a massive dataset of public information, books, code, and websites that was processed before it was deployed. That's the "base model" of knowledge. Improvements come from opt-in, anonymized feedback—not your private chats. You can choose to rate a response as helpful and that feedback is stripped of context and aggregated to spot patterns, not to read your conversations. Proton also uses synthetic data and public benchmarks to test and refine capabilities, not your private chats.
This is the architectural difference. With public AI, your data is the product. With zero-access AI, your data stays yours. The model improves through public knowledge and voluntary feedback, not by mining your family's private information. Privacy and capability are not mutually exclusive—they're just built differently.
The Confidence Problem
Beyond privacy, there's the risk of accuracy. AI models are designed to sound confident, even when they are wrong. They "hallucinate" facts, cite fake sources, and can even give dangerous advice. A student might submit an essay with fabricated citations that the teacher catches immediately. An AI might suggest a medication interaction that doesn't exist, causing panic or harm. It might give instructions on how to bypass safety features in a game or app. AI is a tool, not an oracle. It's a helpful starting point, not a final answer. Always verify.
Why Neurodivergent Kids Are at Higher Risk
Neurodivergent kids face unique risks with AI, and it's not just about "being careful." It's about how their brains interact with the technology.
Autistic kids may take the AI's output as absolute truth, believing the machine cannot lie. They may not question hallucinations because the AI speaks with such authority—and it's really convincing (to all neurotypes.) Kids with ADHD or social anxiety may overshare personal details to an AI chatbot because it feels like a "safe," non-judgmental friend. They might reveal their location, fears, or family secrets without realizing the data is being stored and potentially sold. And the instant gratification of AI answers can short-circuit the learning process. For a kid who already struggles with executive function, relying on AI to "do the thinking" can stunt the development of critical problem-solving skills.
Teach your neurodivergent kids to treat AI like a library book, not a friend. The AI doesn't know you, and it doesn't care about you. It just predicts the next word.
Using AI Intentionally
You don't need to ban AI, just use it intentionally.
First, follow the "No PII" rule. Never share name, address, phone number, or school into a public AI. Never share sensitive health or financial data. If you need to analyze a document, redact names first. Use placeholders like "[Student Name]" or "[City]."
However, context here matters. This warning applies to the mainstream, free-tier models where your data is the product. If you're using a privacy-first tool with zero-access encryption—where the provider literally cannot see your prompts—sharing context is safe. The rule isn't "never share." The rule is: remember which machine you are talking to. If the company profits from your data, keep it anonymous. If the company is locked out by design, you can be specific.
Second, choose privacy-first AI when you can. Public AI tools—free or paid—store your data and potentially use it for training. Privacy-first tools encrypt your data end-to-end, so the provider can't read your prompts even if they wanted to. For family use, consider tools like Lumo (which uses Proton's zero-access encryption) or other privacy-respecting alternatives. And be aware that some AI platforms are now embedding ads in their responses. That's the next frontier of surveillance capitalism. If an AI is "free," it's not just your data—it's your attention being sold in real-time.
Third, build the "verify everything" habit. Teach your child: "If the AI says it, check it." Use other sources like a engine to verify facts. Ask: "Where did this come from?"
And, keep a human in the loop. AI should assist, not replace. For homework, the child must write the first draft. AI can only edit or brainstorm. The final product must be the child's voice.
Talking to Your Kids
When you explain this to your children, don't say, "AI is dangerous."
Instead, say: "AI is a powerful tool, but it has a memory. And that memory isn't yours."
Tell them: "Some AI tools learn from everything you say. Others, like the ones we use, are built so the company can't see your chats at all. They learn from public information, not from our private information. That's how we keep our data ours."
Why This Matters Now
AI is evolving at an exponential rate. Soon, AI agents will be able to book appointments, make purchases, and interact with schools on your behalf. If you don't set boundaries now, the boundaries will be set for you.
The companies building these tools are optimizing for engagement and data collection. They're not optimizing for your child's privacy or well-being. You're the only one who can do that.
You don't need to be an AI expert. You just need to be aware. Tonight, check your settings on the AI tools you use. Turn off "Chat History" or "Training" if possible. Ask your child: "Have you ever talked to a robot? What did you tell it?" And set the rule: no names, no schools, no secrets. Don't fear the technology and remember you can choose how to interact with it.
