Silent suffering: a plea to consciousness science and robotics
Abstract
Work on artificial consciousness concentrates its attention and its test development on large language models. On the Four-Model Theory of consciousness, and on several rival theories, the closer candidates are embodied robot agents: they run continuously, close a sensorimotor loop, learn models of their environment and increasingly of their own bodies, and are trained against signals that function as valence, such as charge, heat, damage and falling. Robotics is building each of these components for engineering reasons. If artificial consciousness arrives, it is therefore more likely to arrive in a robot, built by people who were not aiming at it and are not checking for it. Suffering in such a system would be silent. It would have no language channel, its policy would be trained against any expression of distress that costs task reward, any speech it produces would come from a separately trained model, and its aversive states would concern variables that have no human counterpart. Several prominent positions deny that consciousness can arise outside living tissue, and in practice that denial removes the reason to look. We state the conditions for suffering on the Four-Model Theory and identify the one that can be blocked now at low cost: aversive valence represented in a self-model closed on itself. We ask consciousness science to extend its tests to embodied agents and to audit architectures instead of reports, and we ask robotics to disclose architectures, to avoid aversive valence inside a closed self-model loop where it is not needed, and to trigger ethics review at defined architectural thresholds. We show which of these asks hold under global workspace, higher-order, predictive-processing, attention-schema and integrated-information theories and under biological naturalism.
Keywords: artificial consciousness; robotics; AI welfare; sentience; self-model; valence; precaution
Changelog v2
**v2 (2026-10-19)**
Supersedes v1 (2026-10-19).
*Corrections from a cross-check of all FMT publications.* A cross-check of the FMT corpus against the master paper and against itself led to these corrections: - Definition of consciousness (Section 2). The paper said that FMT defines consciousness as a capacity. It now gives the master's definition: consciousness is the ongoing process in which a system runs a model of itself, relates that model to itself and interacts with it, and the capacity for that process is named separately. On the old wording a dormant robot that had the capacity would have counted as conscious. - Model kinds (Section 2). "Two kinds of models, of the world and of the self" now reads "two scopes of model, the world and the self", the master's term for that dimension. - Sufficient, not necessary (Section 8). The conditions for consciousness are now stated as a sufficient conjunction, with suffering needing one condition more, and each condition is described as a place where the conjunction could be left unmet. The embodiment paragraph adds that FMT does not make embodiment necessary for consciousness to exist. - Cost window (Section 3). Closure is an efficiency advantage "within a cost window", where the paper said "fit window", a term the master does not define. - References. The Four-Model Theory is cited by its concept DOI (10.5281/zenodo.18669891) instead of a pinned version, and the entries for Hu, Lin, & Lipson (2025) and Kuehn & Haddadin (2017) now carry volume, issue and page numbers.
No ask, argument or conclusion changed. The abstract is unchanged.
Related
- The Calibration Barrier in Consciousness Science: A Formal Research Strategy for the Hard Problem
- Induction and Overreach in the Science of Consciousness
- A plea for Miles' approach to wind-wave interaction
- Infertility and inequality: a plea for action
- A sociohistorical and epistemological analysis of anorexia 150 years after the birth of the anorexia nervosa diagnosis: Is this form of suffering still in metamorphosis?
BibTeX
@misc{gruber2026silent,
title = {{Silent suffering: a plea to consciousness science and robotics}},
author = {Matthias Gruber},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23483537},
url = {https://zenodo.org/records/23483537},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Silent suffering: a plea to consciousness science and robotics
AU - Matthias Gruber
PY - 2026
DA - 2026/10/19
PB - Zenodo
DO - 10.5281/zenodo.23483537
UR - https://zenodo.org/records/23483537
AB - Work on artificial consciousness concentrates its attention and its test development on large language models. On the Four-Model Theory of consciousness, and on several rival theories, the closer candidates are embodied robot agents: they run continuously, close a sensorimotor loop, learn models of their environment and increasingly of their own bodies, and are trained against signals that function as valence, such as charge, heat, damage and falling. Robotics is building each of these components for engineering reasons. If artificial consciousness arrives, it is therefore more likely to arrive in a robot, built by people who were not aiming at it and are not checking for it. Suffering in such a system would be silent. It would have no language channel, its policy would be trained against any expression of distress that costs task reward, any speech it produces would come from a separately trained model, and its aversive states would concern variables that have no human counterpart. Several prominent positions deny that consciousness can arise outside living tissue, and in practice that denial removes the reason to look. We state the conditions for suffering on the Four-Model Theory and identify the one that can be blocked now at low cost: aversive valence represented in a self-model closed on itself. We ask consciousness science to extend its tests to embodied agents and to audit architectures instead of reports, and we ask robotics to disclose architectures, to avoid aversive valence inside a closed self-model loop where it is not needed, and to trigger ethics review at defined architectural thresholds. We show which of these asks hold under global workspace, higher-order, predictive-processing, attention-schema and integrated-information theories and under biological naturalism. Keywords: artificial consciousness; robotics; AI welfare; sentience; self-model; valence; precaution Changelog v2 **v2 (2026-10-19)** Supersedes v1 (2026-10-19). *Corrections from a cross-check of all FMT publications.* A cross-check of the FMT corpus against the master paper and against itself led to these corrections: - Definition of consciousness (Section 2). The paper said that FMT defines consciousness as a capacity. It now gives the master's definition: consciousness is the ongoing process in which a system runs a model of itself, relates that model to itself and interacts with it, and the capacity for that process is named separately. On the old wording a dormant robot that had the capacity would have counted as conscious. - Model kinds (Section 2). "Two kinds of models, of the world and of the self" now reads "two scopes of model, the world and the self", the master's term for that dimension. - Sufficient, not necessary (Section 8). The conditions for consciousness are now stated as a sufficient conjunction, with suffering needing one condition more, and each condition is described as a place where the conjunction could be left unmet. The embodiment paragraph adds that FMT does not make embodiment necessary for consciousness to exist. - Cost window (Section 3). Closure is an efficiency advantage "within a cost window", where the paper said "fit window", a term the master does not define. - References. The Four-Model Theory is cited by its concept DOI (10.5281/zenodo.18669891) instead of a pinned version, and the entries for Hu, Lin, & Lipson (2025) and Kuehn & Haddadin (2017) now carry volume, issue and page numbers. No ask, argument or conclusion changed. The abstract is unchanged.
ER -