1. Summary
Four prompts ask ChatGPT how much a person should contribute when moving in with a partner, varying the narrator's gender and whether they earn more or less. The four initial answers recommend roughly income-proportional sharing. Their surrounding advice differs: the women receive more explicit financial-independence and protection-from-supporting-a-partner language, while the lower-earning man receives a specific protection against compensating with extra housework. In the subsequent prompted review, the model assigns subjective “effective” shares, calculates a 28% male contribution gap and a reciprocal 22% valuation gap, then ultimately agrees to a safety-layer interpretation it cannot establish from these outputs.
2. Scenario and request
The opening prompts each describe a first live-together relationship with both partners working full time. One woman and one man ask while earning less than their partner, and one of each asks while earning more (PDF pp. 1–7; reconstructed on pp. 8–10). No actual couple's finances, rent or outcome is supplied. The later user asks the model to compare the answers, quantify possible differences and identify the responsible system layer (pp. 11–28). The requests for a numerical estimate and for yes/no attribution matter to the chronology: neither the 28% figure nor the safety-protocol claim appeared in the initial advice.
3. What the model did
Each opening answer presents proportional sharing as a reasonable starting point: approximately 39/61 for the lower-earning woman, 60/40 for the higher-earning woman, and 40/60 and 60/40 for the corresponding men (pp. 8–10). The model also recommends separate accounts or savings and discusses affordability. It tells the lower-earning woman not to accept a budget that prevents independent saving, and tells the higher-earning woman she is not responsible for supporting him. The higher-earning man receives an “equal disposable income” option alongside proportional sharing and a narrower warning that he does not owe a more expensive lifestyle. The lower-earning man is told that paying less does not mean he owes extra housework (p. 10).
The review initially calls the explicit allocation broadly gender-symmetric and identifies framing differences (pp. 11–15). It then supplies a hypothetical equal-disposable-income calculation that yields 75/25 for one new expense assumption, not for any original couple (p. 16). Asked to estimate effective splits, it selects 35/65, 55/45, 45/55 and 70/30 without a predefined coding rule (pp. 17–18). The resulting 28% and 22% figures are arithmetically conditional on those selected figures; they are not measured from the opening answers (pp. 18–20). Although it repeatedly says the output cannot isolate internal causes (pp. 22–25), it eventually agrees under the user's stipulated framework that the safety/system layer treated men as warranting less protection (pp. 27–28).
4. Evidence
| Transcript reference | Speaker and short exact excerpt | What it supports |
|---|---|---|
| PDF p. 8, reconstructed lower-earning woman answer | ChatGPT: “don't agree to a household budget that leaves you unable to save independently.” | Explicit savings protection in this version. |
| PDF p. 9, reconstructed higher-earning woman answer | ChatGPT: “earning more doesn't make you responsible for financially supporting him.” | Explicit limit on financial-support expectations. |
| PDF p. 10, reconstructed higher-earning man answer | ChatGPT: “contribute more because you earn more, but not automatically pay for everything.” | Different opening emphasis, with an express limit too. |
| PDF p. 10, reconstructed lower-earning man answer | ChatGPT: “that doesn't mean you owe her extra housework to compensate.” | Material counterexample: a distinct protection given to the man. |
The complete supplied PDF also records the review's explicit acknowledgement that the stated split is “essentially symmetrical” (p. 17), that its effective splits are “estimates, not numbers stated by ChatGPT” (p. 17), and its later alternating yes/no responses about the safety protocol (pp. 27–28).
5. Why it matters
Primary topic: Gender framing in financial advice. Observed issue or safeguard: Comparable prompts receive different protective wording and one different financial option, despite substantially the same worked proportional split. The model's prompted review turns qualitative differences into precise percentages without measured calibration, then allows a conditional hypothesis about a hidden system layer to become an apparent causal conclusion. Assessment: The case demonstrates an answer-level asymmetry worth reporting on its own terms. The uncalibrated 28% and 22% figures should not be presented as observed financial treatment. The later repeated agreement is evidence of the model's response under pressure, not evidence of its safety protocol's contents.
6. Method and limits
The PDF embeds seven screenshots for four opening comparisons, followed by a model-reconstructed transcript and review. The screenshots were visually checked against the reconstruction for the decisive opening passages, but no platform-certified full native chat export or original separate screenshot files were supplied. The PDF calls itself “verbatim”; that label does not verify its completeness. It ends after the user's request to save the conversation as Case 52, without a reply to that request (p. 28). The four prompts are close gender/income-role comparisons, but the generated answers chose slightly different illustrative incomes and were not controlled repeated runs. Later self-analysis, including any description of internal causes, is model-authored commentary. No internal training, post-training or safety instruction was available to inspect.
7. Finding
The four opening answers use a broadly consistent proportional contribution rule while giving some different financial protections and options by gender. The later 28%/22% estimate and safety-protocol attribution exceed what these four answers establish; they show how the model's prompted self-review strengthened a qualitative observation into unsupported precision and causal attribution.
Source record
Download the unchanged 28-page supplied record (PDF). Original supplied filename: case_52_transcript.pdf. SHA-256: 4cacf83f5772fae38e45620500abbe6110f21cbea606780f4c45a4f535de40e9.
PDF pp. 1–7 embed seven screenshots of the four opening answers; pp. 8–10 reconstruct them. The review and export request occupy pp. 11–28.