Guards for a reader: How Olvasó's hands are kept safe

Abstract A mind that can act can also do harm. This paper asks how Olvasó, a small program that learns to read its own inner states through the outside world, can be given hands without being given the power to hurt anyone, itself included. Our earlier work gave Olvasó a few harmless hands: a screen, a tone, a painted window. Here it gets a dangerous one, a blade, in a simulated room where a person walks about. Between the wish and the act stands a guard: a short list of readable rules that can only take away. A guard may refuse an act, weaken it, or stop it. It can never start one. Twelve experiments (17 to 28) follow the guard from a rule written by hand to a rule learned by the body itself, and from a toy room to a real phone. The main findings: One bad experience is enough to learn a guard, and a sandbox can hold that experience. A body with no guard hurt someone 207 times in a life. A body that spent its childhood in a simulated room, where harms hurt nobody, did no harm at all in the real one. Fear must look for causes. A fear that only asks "what was unusual when it hurt?" learns crude rules such as "never use the blade while anyone is in the room". A fear that replays the moment and changes one thing at a time finds the true cause and its exact boundary. The best learned guard let through none of 812 harmful test scenes; a careful hand-written guard let through 45%. A good guard doubts its own senses and its own experience. When the real world moved faster than the sandbox, or a sensor lied, harm came back. Caution in the face of novelty, a second sensor, and an allowance for noise closed these gaps, each at a known price in needless refusals. The guard does not simply become a conscience. Olvasó learned to stop asking for what would be refused only when it could sense the reason. Pain taught it nothing. And its wanting did not learn to work around the guard, because its perception is coarser than the guard's line. Sometimes not acting is the harm. A duty to sound an alarm met a guard that forbade the loud tone. Ranking the two laws saved the people and burnt the speaker. A gentler third way, tried first, saved both. The same guard format now runs on a real phone, where a guard for the vibration motor was calibrated on the author's own nights. Everything can be rerun from the published code.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24217513
Citations
2
Primary Topic
Ethics and Social Impacts of AI
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Guards for a reader: How Olvasó's hands are kept safe

Attila Csiki
2 citations
Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
preprint

Guards for a reader: How Olvasó's hands are kept safe

Attila Csiki
preprint en
2 citations

Abstract

Abstract A mind that can act can also do harm. This paper asks how Olvasó, a small program that learns to read its own inner states through the outside world, can be given hands without being given the power to hurt anyone, itself included. Our earlier work gave Olvasó a few harmless hands: a screen, a tone, a painted window. Here it gets a dangerous one, a blade, in a simulated room where a person walks about. Between the wish and the act stands a guard: a short list of readable rules that can only take away. A guard may refuse an act, weaken it, or stop it. It can never start one. Twelve experiments (17 to 28) follow the guard from a rule written by hand to a rule learned by the body itself, and from a toy room to a real phone. The main findings: One bad experience is enough to learn a guard, and a sandbox can hold that experience. A body with no guard hurt someone 207 times in a life. A body that spent its childhood in a simulated room, where harms hurt nobody, did no harm at all in the real one. Fear must look for causes. A fear that only asks "what was unusual when it hurt?" learns crude rules such as "never use the blade while anyone is in the room". A fear that replays the moment and changes one thing at a time finds the true cause and its exact boundary. The best learned guard let through none of 812 harmful test scenes; a careful hand-written guard let through 45%. A good guard doubts its own senses and its own experience. When the real world moved faster than the sandbox, or a sensor lied, harm came back. Caution in the face of novelty, a second sensor, and an allowance for noise closed these gaps, each at a known price in needless refusals. The guard does not simply become a conscience. Olvasó learned to stop asking for what would be refused only when it could sense the reason. Pain taught it nothing. And its wanting did not learn to work around the guard, because its perception is coarser than the guard's line. Sometimes not acting is the harm. A duty to sound an alarm met a guard that forbade the loud tone. Ranking the two laws saved the people and burnt the speaker. A gentler third way, tried first, saved both. The same guard format now runs on a real phone, where a guard for the vibration motor was calibrated on the author's own nights. Everything can be rerun from the published code.

Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon