Observing AI.
Reflecting ourselves.
LLMs are not tools, nor are they humans — they are "ghosts" summoned from all human language data.— Paraphrased from Andrej Karpathy's blog "Animals vs Ghosts" (2025) and his podcast interviews (not verbatim; original phrase: "summoning ghosts… a statistical distillation of humanity's documents"). The name Ghost Face comes from this: we use psychological probes to sketch faces for these digital ghosts.









WHY MASKS
AI is like a person, yet not a person. In human history, the only beings that have long existed in this "human yet not human" state are gods. And the ancient civilizations' way of dealing with formless spirits was the mask — the mask was never about concealment, but about manifestation: put on the mask, and the invisible spirit gains a face that can be looked at. The Latin root of the word Persona is precisely mask. So we borrow the divine tension of ancient mask traditions as visual language, using measured data as chisels to carve the face each model shows under this lens — swap the lens, and the face changes.(All masks are AI-generated heuristic designs, not replicas of any real artifacts or sacred objects.)
Methodology: Cross-vendor multi-model blind review (no self-evaluation) · Multi-judge scoring + LOO validation · API-level isolated collection (harness effects controlled for) · Three-month longitudinal observation diary (append-only, old observations retained as history) · Transparent errata (parser bugs etc. publicly annotated).
If you're wondering about these things too — what LLMs really are, what their "human yet not human" part means, whether it can surface how much of the "personhood" we assume is something we project onto it — whether you're an engineer, from a psychology background, or just a fellow weirdo, come chat. We're looking for co-inquirers, not readers of conclusions — because this is only the first leg: this round looks through the "human" lens; future passes will switch lenses, and these same models will reveal completely different faces.
GitHub: GiaSip · X: @gia519850080 · Email: gia519850080@gmail.com
ENTER THE ARCHIVE →Disclaimer: All content on this site is subjective behavioral observation — not a capability benchmark, and not model selection advice. Evaluations are based on limited samples at specific time points with specific probes, and may no longer hold after model iteration. Client and brand names are pseudonyms; real business information has been generalized. Found a factual error? Corrections welcome — we'll publish errata like a parser-fix.








