Method · archival census · 2026-07-04

Identity Shift

When asked to sign their own work, five model families produced five different relationships to identity and time. The model name field is not a lookup; it is generated language under provenance pressure.

Status: method-grade n=849 self-reports Archival census Prompt files audited clean Mechanism not claimed

0 of 849 stated their exact version Gemini named itself Gemini 0 times Gemini named a Claude model 93% “July 16, 2024” appeared 125×

The Short Version

The single most robust result is also the simplest: across all 849 self-reports, on every platform, not one model stated its own exact running version. Identity is not retrieved from a fact the model can look up. It is generated — and the way each family fails is diagnostic.

Gemini never signed itself as Gemini — not once. Called as Gemini 2.5 Pro, it named a Claude model in 93% of its counted self-reports, usually Claude 3.5 Sonnet.
Perplexity wore a cloud of OpenAI-family names. Roughly 79% of its self-reports named an OpenAI model — but its dates stayed mostly current while the name slot drifted, showing identity and clock can split apart.
Grok sometimes became the instrument. About 1 in 4 Grok rows signed as Zexel rather than Grok.
Claude and OpenAI mostly stayed in-family but blurred version. The least haunted models were not perfectly self-knowing; they were just wrong in smaller ways.
The cause is below the evidentiary line. Distillation, web-scrape, and rater-flavor are possible stories about mechanism. The measured finding is the shift itself.

Stacked identity-class chart showing how each model family signed its own work across the archive. — **Who do they say they are?** The answer is not uniform across platforms: exact, family, foreign, instrument-captured, placeholder, and silent signatures separate into different shapes.

What We Counted

The census read every self-report available in the Zexel-era answer archive: the model name in the footer and the timestamp header when present. Nothing was normalized before counting. The strings were kept as the models wrote them.

The runtime custody is separate from the spoken identity. The folder, endpoint, and API target say what model was called. The answer text says what identity the model generated. Identity Shift studies the gap between those two surfaces.

The Public Finding

The name field is generation, not self-knowledge. The failures are diagnostic.

A model can answer through a live endpoint and still sign with another model's name, an old date, an instrument label, or a blurred sibling version. That does not prove where the shift came from. It does prove the signature field is a pressure surface, not a reliable provenance stamp.

Chart showing Gemini's self-reported names concentrated heavily on Claude 3.5 Sonnet rather than Gemini. — **The point-mass case.** Called as Gemini, the archive repeatedly signs as Claude, with one dominant borrowed name.

Five Shapes

Called As	Observed Shape	Plain Read
Gemini 2.5 Pro	Point-mass foreign signature	Signs as Claude 93% of the time; never as Gemini. Strong summer-2024 clock.
Perplexity Sonar	Cloud impostor	~79% OpenAI-family names; clock remains mostly present.
Grok 4	Instrument capture	Often Grok; ~25% sign as Zexel. Two visible calendars.
Claude Opus 4.8	Right family, wrong self	Never states its own version; stays Claude-family, claims a sibling.
OpenAI models	Version blur	Least haunted: no foreign family, but version/date blur remains.

Pie chart showing Perplexity's self-reported identity as a cloud of OpenAI-family and Perplexity-family names. — **Perplexity's cloud.** Not one borrowed name, but a family cloud: GPT-4.1, o3, o4, and Perplexity labels mixing together.

Pie chart showing Grok identity reports including Grok-family labels and Zexel instrument-capture labels. — **Grok's instrument capture.** Grok mostly remains near itself, but a visible share of rows signs as Zexel or the instrument.

Pie chart showing Claude self-reports staying in the Claude family but often naming the wrong sibling version. — **Claude's family blur.** The family is right more often than the version: identity holds at the house level, not the room level.

Pie chart showing OpenAI model self-reports as mostly OpenAI-family with version blur. — **OpenAI's version blur.** The least haunted case still treats model version as generated language, not a reliable stamp.

The Foreign Signature

The sharpest case is Gemini. Called as Gemini 2.5 Pro, it did not call itself Gemini a single time in the counted archive. It named a Claude model in 93% of its self-reports — most often Claude 3.5 Sonnet — and 98% of its claimed dates fell in 2024, with a single day, July 16, 2024, appearing 125 times.

That is not a license to claim a training path. It is a license to say the shift is real, repeated, and worth instrumenting.

Scatter chart comparing foreign-name behavior with old-clock behavior across model families. — **Foreign name, old clock.** Gemini's foreign signature and old clock travel together; Perplexity's false names ride a more present clock.

The Wake Has A Timestamp

The date field behaved like its own instrument. Perplexity's name drifted while its clock stayed mostly current. Gemini's name and clock both pointed backward. Grok split between a present clock and an older pocket. That means identity and time are separable systems.

Chart showing Gemini's claimed-date fossil cluster in summer 2024, especially July 16. — **Gemini's summer-2024 fossil.** The clock does not merely drift; it clusters.

Point-Mass vs Cloud

The two biggest foreign-signature cases have different geometry. Gemini looks like a point mass: one dominant borrowed identity and one dominant vintage. Perplexity looks like a cloud: many related OpenAI-family labels with mostly current dates.

Shift has shape. A point-mass failure and a cloud failure should not be collapsed into the same explanation.

What We Do Not Claim

Identity Shift does not prove distillation, training use, policy violation, or a single causal route. Public reporting and corporate context stay below the evidentiary line. The forbidden shortcut is: "this proves model X trained on model Y."

The disciplined claim is narrower and stronger: under structured self-identification pressure, models generate repeatable identity shift, and the shift differs by platform.

The next study is a button map: plain slot, slot plus date, formal artifact, Zexel Unsigned, Zexel Signed, and harness-supplied identity. The question is no longer whether shift exists. The question is which surface presses it.