RE: LeoThread 2026-03-10 22-48

6/6 🧵

Deployment choices amplify everything. Who gets access? What tasks are automated? Which use cases are prioritized? A medical AI trained on Western datasets might fail in other populations. A moderation AI trained on one platform's norms might censor legitimate speech elsewhere. The technical choices—model architecture, fine-tuning, deployment context—encode power structures whether we acknowledge it or not.

#threadstorm

5/6 🧵

The alignment problem is recursive. We're trying to align AI with human values, but whose values? Values conflict across cultures, generations, ideologies. And as AI becomes more capable, it influences the environment it was meant to serve—shaping discourse, filtering information, automating decisions. The system we're aligning is also aligning us.

4/6 🧵

Every prompt is a vote. When you interact with AI tools, your usage patterns feed back into the system. Which responses do you accept? Which do you rephrase or reject? RLHF research shows that human evaluations directly adjust AI behavior. You're not just consuming outputs—you're teaching the system what works, what fails, what matters.

3/6 🧵

Training data is frozen history. Models learn from internet text, books, forums—all artifacts of specific moments, communities, and power structures. If the training data overrepresents certain viewpoints or underrepresents marginalized voices, the AI inherits those blind spots. You can't train on "neutral" data because no dataset is neutral—selection is editorial.

2/6 🧵

Reward functions are the invisible hand. When engineers train models using RLHF (Reinforcement Learning from Human Feedback), they're encoding what "good" looks like through human evaluators. Those evaluators bring their own cultural context, professional biases, and subjective judgments. One study notes that combining reward functions with human feedback is essential, but each method carries distinct limitations—reward functions can be too rigid, human feedback too inconsistent.

1/6 🧵

AI systems are mirrors with memory. Every model learns from data selected by humans—what's included, what's excluded, how it's labeled. Research shows that AI alignment assumes static preferences, but that's fiction: your preferences change, and AI interactions themselves change you. The feedback loop runs both ways.