Work

Signal ≠ Value: Does Pressure Bend the Human Feedback That Trains AI?

Year: 2026
Type: preprint
Study design: Study Protocol
Abstract: AI systems are trained and evaluated on human feedback: ratings, preference votes, thumbs-up. These are treated as direct readings of what people prefer. This paper argues that each label is better understood as a measurement taken under conditions, and 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 direction is what matters. Random error averages out as more feedback is collected. A consistent lean does not: if it reaches training, more data teaches the lean more firmly. The paper sets out a pre-specified, within-person experiment (H1). Sixty to one hundred adults choose between two short answers to 48 questions twice, at least a week apart: once under a personal time limit plus a memory load, once untimed. One answer in each pair has a pre-named “attractor” style: unhedged confidence, or agreement with the asker. The primary measure, D, is each person’s shift toward the attractor between the two sessions. The paper shows that D counts only departures from the person’s own reflective choice. The smallest effect of interest is 3 percentage points, and the outcome is read from a decision table with six results, fixed in advance. Before any data, the full study was simulated on synthetic participants. At 60 people, the design detects a 6-point drift 99.7% of the time, with a 5.3% false-alarm rate. The simulation found four flaws in the design of the author’s earlier paper, and they are fixed here. Further checks show two things. The blinded sample-size rule does not inflate false alarms (4.8–4.9%). Randomising the attractor’s screen side prevents side bias from creating a false result. A second simulation compared remedies. The earlier paper’s weighting formula removed about 7% of the lean. Improving working conditions at source (capped sessions, breaks, hourly pay) removed about 30%, with no data lost and no group penalised. That ranking held across every assumed size of improvement tested, but the 30% figure is assumption-dependent. The paper therefore revises Coherence-Weighted Human Feedback (CWHF) into a staged programme: record the conditions of each label, prevent drift at source, weight only where it demonstrably helps, and audit always. It also gives a minimum schema for recording conditions and governance rules against using condition records to monitor workers. Six linked tests are specified (H1–H6), including an observational study of public Chatbot Arena votes that needs no new participants. The paper also includes a decision brief for data and alignment leads, a table of alternative explanations, the exact sentences the results paper will use for each outcome, a pre-mortem, and notes on ethics, data protection and equity. What this paper does not establish: no human data have been collected; reflective choices are a stated reference, not anyone’s true values; and no claim is made that correcting a lean improves reward models. Every claim is labelled Established, Hypothesis, Simulated, Proposed or Retired. This is a new paper, not a revision. It continues and supersedes the claims of Rao (2026), “The Coherence-Weighted Human Feedback (CWHF)”, doi:10.5281/zenodo.19235970, and lists every claim withdrawn from it. Preprint, not peer reviewed. Version 1.1, 12 November 2026. Companion lab: https://twopointsingularity.com/lab/ (mirror: https://cwhf.vercel.app/) Paper on the web: https://cwhf.vercel.app/paper/ Code and simulation: https://github.com/venramak/twopointsingularity-cwhf (simulation code Apache-2.0; text and data CC BY 4.0)
Author: Rao, Venkat
Language: English
Open Access status: green
OpenAlex ID: W7225864031
Cites:
Cited by:
Related to: 10