Research · paper Signal ≠ Value (2026) · design pre-specified · not yet tested
Signal ≠ Value: Coherence-Weighted Human Feedback
A thumbs-up is a signal. It is not the value.
AI systems learn from ratings, votes and thumbs as if each were a direct reading of what people prefer. A rating is better understood as a measurement taken under conditions: how rushed, distracted or tired the person was is part of it. CWHF asks one testable question: when people judge under pressure, do their choices move away from the choices the same people make on reflection, in a consistent direction rather than at random? The primary measure is that signed reflective–pressured choice gap (D); the study tests condition-dependent preferences over answers, not anyone’s moral values or correct preference.
The direction is what matters. Random error cancels as more feedback is collected. A consistent lean does not: more data teaches a model the lean more firmly.
Read the paper (PDF)Read it on the web ↗Open the Signal ≠ Value LabCode on GitHub ↗
Status
| Paper | Signal ≠ Value: Does Pressure Bend the Human Feedback That Trains AI? 12 November 2026 · preprint, not peer reviewed · supersedes the claims of the March 2026 paper (doi:10.5281/zenodo.19235970) · technical supplement (PDF) |
| Study design | Three studies specified: H1 (within-person experiment), H2 (weighting versus prevention), H3 (public vote data). Simulated before any data collection; code is open |
| Registration | Earlier H1 at OSF.IO/W7NYH; the revised H1, H2 and H3 are to be registered before any data are collected |
| Data | None yet. The simulation tests the design, not the hypothesis |
| Conflict of interest | The author has a commercial interest and will not serve as principal investigator. |
| Licence | Text and data CC BY 4.0; simulation code Apache-2.0 |
How strong each link in the wider story is
| People rely on each other’s attention and judgement (distributed cognition, shared intentionality). | Well established |
| Human social networks are layered in circles of rough size. | Established pattern; exact numbers disputed |
| A person’s state (fatigue, mood, context, framing) changes their answers and choices. | Well established |
| Current AI training treats human feedback as state-free. | Largely true of current practice |
| The state-dependent error is directional and does not average out with scale. | Hypothesis · H1 tests this |
| Being well held moves feedback closer to the value it stands for. | Proposal · unproven |
How to cite
Rao, V. (2026). Signal ≠ Value: Does Pressure Bend the Human Feedback That Trains AI? Preprint. Two Point Singularity. https://cwhf.vercel.app/paper/