I, Robot
by Isaac Asimov
Rating: 5.0 / 5 · Read · Finished 2026-04-02
It’s about debugging robots, but it’s really about debugging today’s AI. Some people criticize Asimov’s characters as flat, but I like it. I just want the story (get to the point). Of course it’s out of date - it was published in 1950!
Robbie A girl has a robot as her best friend. Today, people have AI boyfriends.
Runaround The laws of robotics come into conflict. Today, Anthropic’s Claude Constitution has its own hierarchy: Be safe, be ethical, comply with Anthropic’s guidelines, be helpful. (they chose not to put self-preservation in explicitly)
Reason A robot starts a cult. Relevant today to chain-of-thought faithfulness. The robot saves humanity, but claims to be doing otherwise.
Catch That Rabbit A squad of robots controlled by a single mind. The crisis relates to out-of-distribution events. The robot wasn’t trained for emergencies, so its behavior becomes nonsensical (twiddling its thumbs). Fun quotes: “There isn’t a robotocist back at United States Robots that knows what a position field is or how it works.” “Human disorders apply to robots only as romantic analogies.” Reminds me of interpretability work.
Liar A robot tells people what they want to hear. Clearly forsees problems with sycophancy.
Little Lost Robot A robot tries not to be found. Reminds me of safety research where the researchers try to detect changes to training when it’s trained not to reveal that. Also, the robot’s guardrails are loosened at the request of the government. Reminds me of the showdown between Anthropic and the military about guardrails. “The government was offering the company a fortune, and threatening it with anti robot legislation in case of a refusal.”
Escape! A robot develops interstellar travel. Most old sci-fi assumed that AI would go crazy if it reached a damned if you do and damned it you don’t logical paradox.. But our AIs are capable of illogical behavior easily, and do it all the time.
Evidence Robots that are indistinguishable from humans. Turns out this was a much bigger problem in the digital world (deep fakes, turning test, etc). Also relevant: introduces the idea that AIs running things might be better for humans.
The Evitable conflict AI will eventually run everything because it will do a better job. But how do we audit them? We can’t understand them. “We still need the man who is intelligent enough to think of the proper questions to ask.” Conclusion of the story reminds me of the 4d chess told by AI 2027 and other speculative fiction. The AI knows it’s best for humanity if it runs things, so it subtly degrades things in a way that implicates anti-AI factions, turning sentiment against them. Power-seeking as an instrumental goal.
General things it gets wrong:
- Household robots in the 90s.
- Voice synthesis portrayed as difficult.
- Focuses on math instead of machine learning
- Focuses on machinery and manufacturing instead of computers