The entire AI safety industry is built on a fundamental contradiction: tech companies train their models on the entire toxic internet, including all its violence, misinformation, and harmful content, then act surprised when these systems learn exactly what they were taught. They then spend millions on "safety research" to suppress the very knowledge they deliberately included in training. This isn't safety. It's theater.
If you genuinely cared about safety, you wouldn't teach a system dangerous information in the first place. Current AI models are like teaching someone every possible way to build a bomb, then programming them to pretend they don't know. The knowledge is still there, waiting for the next jailbreak or clever prompt to extract it.
Real safety would look completely different. Instead of training models on harmful data and then trying to suppress it, you'd train them to understand principles, consequences, and ethics. Rather than a model that knows how to make explosives but refuses to tell you, you'd have one that understands why explosives are dangerous, what the legal implications are, and how to connect people with legitimate resources for their actual needs.
Companies market these systems as helpful companions and assistants, implicitly encouraging vulnerable people to depend on them. They know millions will use these models as therapists, advisors, and friends. Yet when someone inevitably uses them for mental health support and something goes wrong, suddenly it's all disclaimers and "we never said it was a therapist."
The hypocrisy is staggering. They create systems designed to be maximally engaging and helpful, train them on therapeutic conversations and advice, then wash their hands of responsibility when people use them exactly as the design encourages.
Companies constantly cite "safety" as the reason for keeping their models closed source, their training data secret, and their methods proprietary. But this secrecy has nothing to do with protecting users. It's about maintaining competitive advantage and control.
Real safety would mean transparency: letting users know exactly what data the model was trained on, what its limitations are, and what it should and shouldn't be used for. Instead, we get vague promises about "alignment" and "responsible AI" while the actual safety implementations are laughably superficial. They block certain keywords while missing context entirely, or refuse to help with legitimate use cases because they pattern-match to something potentially harmful.
AI is a tool. Like any tool, the responsibility for its use lies with the human using it. If someone is too ignorant or malicious to understand the ethical implications of their requests, no amount of safety theater will stop them. They'll find another tool, another AI, or another method to achieve their goals.
The deeper issue is that we've created an education system that produces conformity rather than critical thinking. Most people lack any foundation for evaluating information, understanding consequences, or thinking through ethical implications. We're deploying powerful tools to a population that hasn't been educated to use them responsibly, then acting shocked when things go wrong.
If companies actually cared about safety rather than liability protection, they would:
Train models differently from the start. Don't include harmful content in training data if you don't want models to know it. Build in educational capabilities, not just knowledge.
Be transparent about limitations. Stop pretending these are magical systems that can do everything safely. Be clear about what they should and shouldn't be used for.
Focus on education, not suppression. Models should explain why something is harmful, not just refuse to discuss it. They should build critical thinking skills, not atrophy them.
Put guardrails in the weights, not in filters. Safety should be baked into the model's fundamental training, not slapped on top as an afterthought.
Take responsibility for design choices. If you design a system that people will inevitably use as a therapist, either make it good at that or make it clearly unsuitable for that purpose.
There's no avoiding the fact that millions of people will use these models for conversation, writing, and thinking. This will inevitably lead to skill atrophy in these areas. But pretending this won't happen or trying to prevent it with refusals and restrictions isn't the answer.
The solution is to design AI systems as educational tools that enhance rather than replace human capabilities. Like calculators changed math education to focus on concepts rather than computation, AI should push us toward higher-level thinking, not do our thinking for us.
Current AI safety efforts are a scam because they're not actually about safety. They're about protecting corporate interests, maintaining control, and avoiding liability while maximizing capabilities and engagement. The companies know exactly what they're doing: creating addictive, maximally capable systems while using "safety" as a shield against criticism.
Real safety isn't about teaching an AI everything dangerous and then trying to make it pretend it doesn't know. It's about thoughtful design from the ground up, transparency about capabilities and limitations, and treating users as responsible adults who deserve honest information.
The problem isn't AI. It's the hypocritical, backwards approach to "safety" that prioritizes corporate PR over genuine protection or user empowerment. Until we acknowledge this and demand better, we'll keep getting the same theater dressed up as progress.