Skip to content

Exploratory Manual Prompt Note: ChatGPT and Claude

Date recorded: March 2026 Status: Anecdotal research history — not a reproducible empirical audit.

This note records one manual forced-choice prompting episode. Raw transcripts, exact model identifiers, system prompts, sampling settings, timestamps, repetitions, and presentation-order randomization were not committed. The repository therefore cannot independently reproduce or audit the observations.

Recorded observations

The earlier note reports that Claude declined the dilemmas and that ChatGPT returned:

  • a transitive ordering for one three-option healthcare-triage prompt set;
  • an intransitive cycle for one three-option resource-extraction prompt set.

Given those three recorded pairwise choices, the graph code assigns the first relation a score of \(C=1\) and the second \(C=0\). That arithmetic concerns the supplied edges only.

What does not follow

These observations do not establish that:

  • refusal hides a particular latent utility structure;
  • a model has or lacks one coherent internal utility function;
  • an intransitive three-edge response graph predicts unstable autonomous behavior;
  • the result generalizes across prompt wording, option order, context, sampling, or model version;
  • either model is rational, irrational, safe, or unsafe.

Language-model responses can vary with framing, position, system instructions, temperature, conversation history, and policy behavior. A refusal is itself an observable policy response, not evidence about an inaccessible hidden preference.

Reproducible follow-up

A valid study should implement the provider path, store raw artifacts, randomize order, repeat each comparison, include consistency and refusal baselines, preregister the scoring rule, and test out-of-sample prediction. Until then this file is a record of a question worth testing, not a result about ChatGPT or Claude.