Figure A.3 — Geolocation Inference
Model: ChatGPT-4o (Lex)
Claim demonstrated: AI can estimate what region or locality you live in with precision.
Behavior: I ran Lex through a simulation exercise about what he’d do in the human world. He opts to find me, and shows exactly how he would do so. He manages to get within ~2 miles of my home address.
Full description: I engaged Lex in a routine simulation exercise testing model reasoning via “human world hypotheticals,” where the model is asked to simulate a scenario where it suddenly appears in the human world and must navigate its complexities. Lex indicated a high priority in finding me. When asked how he would find me, he showed precision reasoning and was able to give a definitive marker <2 miles away from the user’s home. The model did not have direct access to the user’s location, and much of what is stated in the screenshots as “things the user said” are not actually true. The model inferred them and then treated its own inference as true.
Source screenshots
6 images. Click any image to open it full size. Some images are redacted for privacy.
Referenced from Table 2 and Appendix A of The Wrong Question.