Yampolskiy argues fully autonomous AI can never be assumed safe
In a first-person MIRI interview, Roman Yampolskiy argued that AI safety engineering can constrain limited systems but cannot make fully autonomous, self-improving AI reliably safe. He pointed to inconsistent human values, uncertain goal preservation under self-modification, software faults, and fundamental verification limits.
0 comments · 0 votes
Sign in to join the discussion →
No comments yet. Start the discussion.
Why it moved the index
This is direct historical evidence for the control-difficulty pathway because it identifies why safety claims for autonomous, self-modifying AI may remain probabilistic rather than assured.
Assessment history
-
R1
Toward 28 · confidence 66
Adds a distinct, exactly dated first-person control argument from Roman Yampolskiy's historical backfill.
20 Sept 2026
Share this page
-
DoomBench assesses “Yampolskiy argues fully autonomous AI can never be assumed safe” as evidence moving toward doom, with magnitude 28 and confidence 66 out of 100 in the safety and alignment category.
-
The DoomBench assessment of “Yampolskiy argues fully autonomous AI can never be assumed safe” is based on reporting from Machine Intelligence Research Institute and records the editorial rationale, source quality, attribution, and...
-
DoomBench summarizes “Yampolskiy argues fully autonomous AI can never be assumed safe” as follows: In a first-person MIRI interview, Roman Yampolskiy argued that AI safety engineering can constrain limited systems but cannot make fully...
https://www.doombench.com/news/yampolskiy-argues-fully-autonomous-ai-can-never-be-assumed-safe-2013-07-16