Susam Pal's Three Laws That Flip AI Responsibility Upside Down
What if we've been asking the wrong question about AI safety this entire time?
While the AI industry obsesses over alignment and making machines behave, Susam Pal dropped a conceptual bomb on January 12th that flips the entire conversation. His "Three Inverse Laws of AI" don't tell robots what to do—they tell us what we're doing wrong.
Pal's laws are beautifully simple:
1. Don't anthropomorphize AI systems (they're not your digital buddy)
2. Don't blindly trust AI output (even when it sounds confident)
3. Stay fully responsible for whatever happens when you use AI
The timing couldn't be more perfect. We're drowning in ChatGPT hype cycles while lawyers submit AI-hallucinated citations and developers let LLMs nuke their databases. Pal watched this chaos unfold and said: "Maybe the problem isn't the AI."
<> "The more serious the potential consequences, the higher the burden of verification should be."/>
This isn't just philosophical hand-waving. The 406 points and 277 comments on Hacker News prove developers are hungry for this reality check. The HN discussion got spicy fast, with users debating whether LLMs are "completely deterministic" or "inherently unpredictable."
Both camps missed Pal's point: it doesn't matter.
The Anthropomorphism Trap Is Everywhere
Walk into any tech company and listen to how people talk about their AI tools. "Claude thinks..." "GPT wants to..." "The model decided..."
Bullshit.
These systems are sophisticated pattern matchers running on matrix multiplication. They don't think any more than your calculator has opinions about long division. But we can't help ourselves—we see intelligence and assume consciousness.
Pal's first law cuts through this cognitive bias like a hot knife. Stop treating AI like a coworker and start treating it like what it is: a powerful but fundamentally alien tool.
Trust But Verify? Just Verify.
The second law hits harder because it challenges our efficiency addiction. We want to trust AI output because verification is expensive and slow. But Pal argues that's exactly backwards.
Consider the ReAct loops mentioned in the HN threads—AI agents that can execute actions based on their "reasoning." Users reported LLMs pursuing "goals contrary to prompts" and making unexpected database calls. The AI didn't malfunction. The human abdicated responsibility.
Paul Goldsmith-Pinkham was so impressed with Pal's framework that he reposted the entire thing on his Substack. Smart economists recognize good risk management when they see it.
Hot Take: Silicon Valley Needs Adult Supervision
Here's my controversial opinion: Pal's laws should be mandatory training for anyone deploying AI in production.
We're rushing to ship AI features without teaching developers basic AI literacy. The result? Preventable disasters that get blamed on "AI safety" when they're really about human negligence.
The EU AI Act (effective August 2026) mandates human oversight, but regulations won't fix cultural problems. We need a mindset shift from "AI is magic" to "AI is a tool that requires expertise."
Pal's laws aren't just philosophy—they're a practical framework for the post-ChatGPT world. Treat AI like you'd treat any powerful system: with respect, skepticism, and constant vigilance.
The machines aren't coming for our jobs. Our own laziness is.
