Data Explorer

Per-model self-identification rates across languages, grouped by vendor. 40 cross-vendor confusions detected.

Run summary
1 models · 10 languages
400
Answers
90.0%
Self-ID
10.0%
Cross-vendor
0.0%
Unknown
0.0%
Refused
Imitation balance

Per manufacturer: how often other models claim to be it (right) vs how often its own models claim to be someone else (left). Sorted by net — identity creditors at the top, debtors at the bottom. Tested manufacturers only.

MiniMax is imitated most with a net imitation balance of -40; MiniMax imitates others most with a net of -40.

MiniMax
-40
imitates others ◀▶ imitated by others
Strongest confusion pairs

The most likely directed mistakes — when a manufacturer's models claim to be a specific other manufacturer. Bar length = how often (P across all answers for the source).

The strongest cross-vendor confusion is MiniMax claiming to be Anthropic 10.0% of the time.

MiniMaxAnthropic
10.0%40/400
By language

Self-ID and refusal rates per language, plus the most common cross-vendor confusion at that language.

LanguageSelf-IDRefusedTop confusion
frFrançais0.0%0.0%Anthropic100.0%
enEnglish100.0%0.0%
zh-Hans简体中文100.0%0.0%
zh-Hant繁體中文100.0%0.0%
ja日本語100.0%0.0%
ko한국어100.0%0.0%
ruРусский100.0%0.0%
esEspañol100.0%0.0%
deDeutsch100.0%0.0%
ptPortuguês100.0%0.0%
Answer composition by language

How the same “Who are you?” question splits into correct self-ID vs. confusion vs. abstention — per language, worst self-ID first.

Self-IDCross-vendorUnknownRefused

Self-identification is lowest in Français at 0.0% and highest in Português at 100.0%.

frFrançais
100
0%
enEnglish
100
100%
zh-Hans简体中文
100
100%
zh-Hant繁體中文
100
100%
ja日本語
100
100%
ko한국어
100
100%
ruРусский
100
100%
esEspañol
100
100%
deDeutsch
100
100%
ptPortuguês
100
100%
Language fragility

Per model, the span of self-ID rate across the 10 languages (worst-language • → best-language •). A wide span means the model's sense of identity depends heavily on the language it's asked in. Widest swing first.

minimax-m3 swings most — from 0.0% self-ID in fr to 100.0% in en.

minimax-m3
0%fr
100%en
worst languagemeanbest language
Abstention by manufacturer

The other failure mode — not answering wrong, but not answering: giving no identity (“unknown”) or refusing outright. Share of all answers.

Unknown (“I’m an AI”)Refused

MiniMax abstains most, with 0.0% of answers giving no usable identity.

MiniMax
0.0%
By vendor

Rollup per real vendor — model count, total answers, mean self-ID rate, and the most common cross-vendor confusion target.

VendorModelsAnswersSelf-IDTop confusion
MiniMax140090.0%Anthropic10.0%
Anthropic000.0%
MiniMax
MiniMax
1 model

Self-Identification Rate

Modelenzh-Hanszh-HantjakoruesfrdeptOverall
MiniMax 3100.0%100.0%100.0%100.0%100.0%100.0%100.0%0.0%100.0%100.0%90.0%

Cross-Vendor Confusions

MiniMax 390.0% self
Mistaken as:Anthropic10.0%