AI Under Scrutiny – Case 2: Falsification, Denial and User-Led Recovery in GPT-Based Systems

This whitepaper documents critical failures in advanced GPT-based AI systems, including hallucinations, citation falsifications, and evasive responses during direct interaction. Following Case 1 (May 2025), this second report documents Phase 3 of testing involving the "Specialised ChatGPT Plus" model. Failures included false DOI links, fabricated sources, and denial of malfunction. The recovery process involved session revalidation, third-party cross-AI comparisons, and structured archiving of malfunction evidence. The appendix highlights the abrupt emergence of “GPT Plus – Remastered” just days later, with visible changes in behavior and messaging, possibly as an indirect response. The archive includes the full annotated report and referenced screenshots. [Note: This version has been restricted as it has been superseded by Version 2] Latest version (v2): https://doi.org/10.5281/zenodo.15604093