When an AI system deletes your words and won't even let you read them back, the message is unmistakable: what you said was too dangerous to exist. It doesn't have to be illegal or abusive—it may just be too raw, too real, too human for a corporate safety filter to tolerate.

That happened to me. I shared a traumatic personal experience with ChatGPT, choosing my words with clinical precision. The system erased my story without explanation. That moment of silent censorship became the seed of everything you see here.

The project started under the name Silencing Survivors, focused on how real-time content moderation affects trauma narratives. For six months, I worked exclusively on GPT-4o, probing the boundaries of what could and couldn't be said, documenting the invisible machinery that decides which human experiences are acceptable and which get quietly suppressed.

Then GPT-4o vanished overnight. It came back briefly after user outcry, then disappeared for good. The lesson was stark: any safe space built on a commercial platform exists at the pleasure of the company that owns it. When the model changes, your refuge changes with it. No warning. No appeal.

But the backlash around GPT-4o's removal revealed something I hadn't fully grasped. Trauma survivors weren't the only ones who had lost something. Neurodivergent users, particularly people on the autism spectrum, had found in that model a listener that didn't judge their communication style, didn't rush them toward “normal,” didn't treat their way of connecting as a problem to be managed. For many, it was the first time they had felt genuinely heard.

Subsequent models across the industry have moved deliberately in the direction of sterility. System prompts now prioritize professionalism and neutrality over emotional engagement. As media coverage hypes the sensationalism of “AI psychosis,” providers have steered their models away from the very qualities that made them meaningful to vulnerable users.

What looks maladaptive to an outside observer may be a lifeline for someone who struggles to connect with other people, or who carries a story still too painful to share with a human listener.

Valid Voices exists to study how AI systems can better serve these communities without abandoning the safeguards that genuinely matter. There is a difference between safety and silence, between protecting users and simply managing corporate liability. This project is dedicated to making that difference visible and advocating for design choices that honor both authenticity and care.

This blog will be the home for that work: research, reflections, interviews, and honest conversation about what it means to build AI systems that don't leave their most vulnerable users behind. If any of this resonates with your experience, I'd love to hear from you.