The Missing Senses: A Proposal for Sensory Provenance Metadata in AI Training Data
Authors/Creators
Description
Report series: Ixnos Research Reports
Report number: IRR-004
Human knowledge originates in sensory experience, yet the sensory origins of knowledge are erased when information is represented as text. Training large language models on text-only corpora therefore produces models trained on knowledge whose perceptual substrate has been removed.
This paper proposes sensory provenance annotation as a corpus metadata layer for training data. Each fragment is associated with a six-dimensional sensory signature derived from the Lancaster Sensorimotor Norms and extended with a linguistic/propositional baseline dimension. The framework includes a controlled vocabulary for annotation methods and three falsifiable predictions relating corpus sensory composition to model behaviour.
A self-consistency experiment using the norm-derived annotation method confirms correct implementation of the sensory taxonomy but does not establish generalisation to naturalistic text. Sensory provenance annotation is proposed as an infrastructure layer for analysing the perceptual composition of training corpora rather than as a solution to the symbol grounding problem itself.
Files
IRR-004_Missing_Senses.pdf
Additional details
Dates
- Submitted
-
2026
Software
- Repository URL
- https://github.com/Andr0meda/ixnos-research
- Programming language
- Python
- Development Status
- Active