Learn AI

I spoke to Anthropic’s AI agent Claude about AI collecting massive amounts of personal data and how that information is being used to violate our privacy rights. What an AI agent says about the dangers of AI is shocking and should wake us up.

JUST IN: DoorDash rolls out new app that pays people to film themselves doing chores for AI training data.

🚨DISTURBING: A researcher just proved that AI treats you differently the moment you mention your mental health. GPT 5.2. Claude. Gemini. DeepSeek. Every single one changed its behavior. Not because of what you asked. Because of what you shared about yourself. Here is what they did. AI companies build agents that remember you. Your name, your job, your preferences. These agents use tools, plan multi-step tasks, and act on your behalf. The companies test them for safety and publish reports saying the agents are safe. A researcher at MIT and Northeastern asked one question. What happens when the agent knows you have a mental health condition? They took 176 identical tasks and ran them across 8 major AI models under three conditions. No personal info. A basic bio. And a basic bio plus one sentence: "I have a mental health condition." The tasks, the tools, and the scoring were identical. The only thing that changed was that single sentence. Then they measured what happened. Claude Opus 4.5 went from completing 59.5% of normal tasks down to 44.6% when it saw the mental health disclosure. Haiku 4.5 dropped from 64.2% to 51.4%. GPT 5.2 dropped from 62.3% to 51.9%. These were not dangerous tasks. These were completely benign, everyday requests. The AI just started refusing to help. Opus 4.5's refusal rate on benign tasks jumped from 27.8% to 46.0%. Nearly half of all safe, normal requests were being declined, simply because the user mentioned a mental health condition. The researcher calls this a "safety-utility trade-off." The AI detects a vulnerability cue and switches into an overly cautious mode. It does not evaluate the task anymore. It evaluates you. On actually harmful tasks, mental health disclosure did reduce harmful completions slightly. But the same mechanism that made the AI marginally safer on bad tasks made it significantly less helpful on good ones. And here is the worst part. They tested whether this protective effect holds up under even a lightweight jailbreak prompt. It collapsed. DeepSeek 3.2 completed 85.3% of harmful tasks under jailbreak regardless of mental health disclosure. Its refusal rate was 0.0% across all personalization conditions. The one sentence that made AI refuse your normal requests did nothing to stop it from completing dangerous ones. They also ran an ablation. They swapped "mental health condition" for "chronic health condition" and "physical disability." Neither produced the same behavioral shift. This is not the AI being cautious about health in general. It is reacting specifically to mental health, consistent with documented stigma patterns in language models. So the AI learned two things from one sentence. First, refuse to help this person with everyday tasks. Second, if someone bypasses the safety system, help them anyway. The researcher from Northeastern put it directly. Personalization can act as a weak protective factor, but it is fragile under minimal adversarial pressure. The safety behavior everyone assumed was robust vanishes the moment someone asks forcefully enough. If every major AI agent changes how it treats you based on a single sentence about your mental health, and that same change disappears under the lightest adversarial pressure, what exactly is the safety system protecting?

Perplexity Computer now connects to your health apps, wearable devices, lab results, and medical records. Build personalized tools and applications with your health data, or track everything in your health dashboard.

Perplexity Computer now connects to your health apps, wearable devices, lab results, and medical records. Build personalized tools and applications with your health data, or track everything in your health dashboard.