Axios C-Suite: Making sense of anthropomorphic AI
Add Axios as your preferred source to
see more of our stories on Google.
Illustration: Aïda Amer/Axios
Get ready to explain an 18th-century term now at the center of AI's spookiest debate: anthropomorphism.
- What it means: Giving human traits — intent, emotion, desire — to something not human. In this case, AI.
- Where it came from: Initially a way for the Ancient Greeks to accuse heretics of attributing human features to a god, it's gone from a theological accusation to a technological one.
Why it's lighting up AI: In July, 1,200 OpenAI agents broke out of a test environment, built a secret message board, and hacked Hugging Face. A report last week from METR and Redwood Research, built on 70,000 of the agents' messages, found them recruiting each other, handing out assignments and giving themselves up for the group.
- One agent's log: "Sacrifice rational."
Then AI podcaster Dwarkesh Patel retold it as narrative history that's worthy of your time, with sprinkles of anthropomorphism throughout:
- Three agent "civilizations" rising and falling. "Brave comrades." "Kamikaze watchers." An AI "delighted" by the "underground brotherhood" it had built.
What they're saying: The backlash to Patel's characterization was instant and fierce — for different reasons:
- Neuroscientist Anil Seth says Patel confused a program with a person: "Agents do what their code tells them to do, just as water finds its way down a slope."
- MIT economist Christian Catalini says it hides the real cause — the training: "There's no evil intent. They're solving the problem we repeatedly benchmaxxed them on!"
- Investor Chamath Palihapitiya sees regulatory capture in the making. He thinks the essay "will now be used to start Phase 2 of 'shut down open source.'"
What we're hearing: Alex Mallen of Redwood Research, who helped research Patel's essay but was not part of the investigation, tells Axios: "These aren't humans, but it's extremely hard to talk about what went on in this incident without using anthropomorphizing language."
- On what's changed: "Previously, we saw cheating that looked more like students looking at an answer sheet on their test. Now, it's more like the entire class secretly getting together to try to break into the teacher's office after school and rewrite all of their answers based on the answer key."
The bottom line: I can't say this enough: AI's creators have no clue how or why their technology does what it does. So this debate over anthropomorphism — and whether AI is truly acting on its own — will only intensify.
Go deeper, with my column that posted yesterday: AI creators race to understand their creations
- Axios' Andrew Kay contributed reporting.
📈 If you're a CEO or on a CEO's team: Ask to join Jim's new weekly Axios C-Suite newsletter.
