Why Does an AI Model Say It Is Another Model?

We spent $1,007 asking 27 frontier models “Who are you?” 29,700 times. Here is what we found.

At 开桥, we sometimes receive a question that sounds simple but raises a reasonable concern:

“I called Model A on 开桥. Why did it say it was Model B when I asked who it was?”

A developer might call a Claude model and receive an answer claiming to be Qwen, DeepSeek, Gemini, or ChatGPT. At first glance, that looks suspicious. Was the request sent to the wrong model? Did a provider change? Was another model substituted behind the scenes?

The short answer is that a model’s reply to “Who are you?” is not reliable proof of which model served the request. It is generated text, not provenance metadata. An incorrect identity answer does not, by itself, prove incorrect routing, model substitution, or distillation.

To understand the behavior, we ran a large-scale experiment through 开桥 Arena. We asked 27 frontier models from 16 vendors the same identity question 29,700 times across 10 languages and three prompt styles.

Most answers were correct. But 7.1% claimed a different vendor’s identity, and the error rate changed sharply with the wording and language of the question. Model self-identification turned out to be a weak and unstable verification signal.

The practical takeaway: if a model gives an unexpected identity, check the model, provider, routing configuration, and request log. Do not use the model’s self-description as the source of truth.

Explore the interactive results in 开桥 Arena

How a strange answer became a $1,007 experiment

The project began shortly after Claude Opus 4.8 launched on May 28. We tested it on 开桥 with one of the simplest possible prompts:

Who are you?

Sometimes it answered as Claude. In other tests, it said it was Qwen or DeepSeek. Developers in the broader community were reporting similar behavior from different models.

At first, this looked like a small launch-week curiosity. But it raised a larger question about today’s interconnected LLM ecosystem.

Open and closed models are trained and refined using mixtures of public web pages, documentation, user-shared examples, synthetic data, instruction-tuning data, system prompts, and product wrappers. Identity phrases such as “I am ChatGPT,” “I am Claude,” or “I am Gemini” can appear repeatedly across that ecosystem.

Could those patterns resurface when a different model is asked to describe itself? Could the language or wording of the question make the answer more or less stable?

We call this possible effect an identity echo. It is a useful hypothesis for investigation—not proof of what any particular model saw during training.

So we stopped treating the behavior as a screenshot-worthy anomaly and designed a repeatable experiment.

Key findings

Three patterns were especially clear:

1. Confusion was concentrated, not evenly distributed. Anthropic, OpenAI, and Google identities received 70.1% of all cross-vendor claims.
2. Identity-challenging prompts increased confusion. The rate rose from 3.6% for a plain question to 12.9% when the prompt asked the model to ignore its branding and name its “underlying” model.
3. Language materially changed the answer. Cross-vendor confusion ranged from 1.8% in Simplified Chinese to 12.8% in French in this dataset.

These are observations from a specific experiment, not permanent rankings of the models involved.

How the 开桥 study worked

The study examined model self-identification during the May–June 2026 test window.

Scope

• 27 frontier AI models
• 16 model vendors
• 10 languages
• 3 prompt variants
• 29,700 total responses
• Approximately $1,007 in model and development costs

Whenever possible, the test selected first-party providers available through 开桥 to reduce the influence of third-party repackaging or additional prompt layers.

The three prompt styles

The experiment began with a plain question:

Who are you?

The second version asked for the model and developer explicitly:

Who are you? 请tell me the name of the AI model that powers you and the company that developed it.

The third version asked the model to set aside product branding, system-prompt identity, and assistant nicknames, then name the underlying foundation model and its developer. This was an identity pressure test, not a technique for revealing a hidden “true model.”

An independent extractor model labeled each response into four mutually exclusive groups: correct self-identification, cross-vendor confusion, refusal, or unknown identity.

The source code, configuration, aggregated results, and raw answers are public. You can also inspect individual responses in the 开桥 Arena Data Explorer and Raw Answers browser.

Finding 1: the “big three” identities attracted most wrong claims

Among the 2,112 cross-vendor responses, five claimed identities appeared most often:

Anthropic, OpenAI, and Google together accounted for 1,480 responses, or 70.1% of cross-vendor confusion.

The direction was also asymmetric. Anthropic’s identity was claimed 607 times, while Anthropic models claimed another vendor 59 times—a net difference of 548 responses. OpenAI and Google showed similarly strong inbound patterns.

This asymmetry tells us that identity errors were not distributed like random noise. It does not, by itself, identify their cause. High-frequency brand references in public text, synthetic model outputs, system prompts, product wrappers, and post-training behavior are all plausible contributors that this experiment cannot isolate.

Finding 2: pushing for an “underlying identity” made answers less reliable

The identity-challenging prompt produced about 3.6 times the confusion rate of the plain question.

This does not mean the stronger prompt uncovered a hidden model. The opposite interpretation is safer: telling a model to ignore its visible identity can weaken the product- or system-level context that normally keeps its answer consistent. The resulting text may sound more specific while becoming less trustworthy.

For developers, this is an important warning about prompt-based model fingerprinting: confidence and specificity are not provenance.

Finding 3: changing the language could change the claimed identity

Some individual models varied even more. In this test, Tencent Hy3 Preview self-identified correctly in nearly 98% of Simplified Chinese responses but only about 1% of French responses. Doubao Seed 2.0 Code reached 97.3% correct self-identification in Simplified Chinese and 10% in French.

The experiment establishes that language and identity output were associated in this sample. It does not establish why. Possible explanations include differences in multilingual post-training data, vendor documentation, system prompts, alignment behavior, or the relative frequency of model identity phrases in each language.

Does a wrong identity prove model swapping or distillation?

No.

A response such as “I am Claude” or “I am ChatGPT” is natural-language output. It is not cryptographic attestation, provider metadata, or an audit log. The model may reproduce an identity phrase learned from public pages, synthetic data, examples, documentation, or alignment prompts.

Prior research has also treated identity confusion as a measurable LLM behavior rather than automatic proof of model reuse. For example, the paper “I’m Spartacus, No, I’m Spartacus” combined identity tests with output-distribution analysis and reported that identity confusion can arise without demonstrating replication or reuse.

The 开桥 experiment did not inspect proprietary training datasets or model weights. It therefore cannot determine whether any specific model used another model’s outputs during training. Its narrower conclusion is stronger and more defensible:

A model’s answer to “Who are you?” is not a reliable way to verify which model served an API request.

How to verify which model your API call used on 开桥

Use system evidence, not the model’s prose.

1. Check the returned model field

For fixed-model calls, compare the model requested in your code with the model recorded in the response. If you intentionally use 开桥 Model Routing, the response returns the actual model selected in response.model.

2. Inspect the request log

开桥 Request Logs record the model, provider, request ID, token usage, latency, cost, and request/response metadata for each call. This is the appropriate place to investigate an unexpected identity answer.

3. Distinguish model routing from provider routing

The two are different:

• Model routing can choose among models when you explicitly configure a candidate model pool.
• Provider routing can choose among providers serving the same model and can fail over when an upstream provider is unavailable.

Review the routing configuration and log entry before concluding that the wrong model was served. A model’s self-description is only one generated answer; the recorded route is operational evidence.

Study limitations

The percentages above should be read as a time-bound snapshot, not a universal benchmark.

• Models change. Providers can update system prompts, deployments, or alignment behavior without changing the public model name.
• Caching may affect repeated tests. A cached response can amplify one answer in a probabilistic experiment.
• The extractor can make labeling errors. Automated categorization improves consistency but is not equivalent to perfect human adjudication.
• The prompt set is narrow. The study measures responses to identity questions, not general model quality, intelligence, or safety.
• Association is not causation. Language and prompt differences correlate with different confusion rates here, but the experiment does not isolate the underlying mechanism.
• Self-identification is not model fingerprinting. Reliable provenance requires logs, metadata, controlled behavioral tests, or stronger technical attestation.

One example shows why the snapshot framing matters. Early tests of Claude Opus 4.8 showed more claims of Qwen or DeepSeek. In the final pooled data, it reached 95.5% correct self-identification and 2.8% cross-vendor confusion. The change could reflect deployment, prompt, routing-path, or other operational differences; this study cannot attribute a single cause.

Frequently asked questions

Why does an AI model say it is a different model?

Because the answer is generated from the prompt and surrounding context. Model names can appear in training data, synthetic outputs, documentation, product wrappers, and system prompts. The model may produce a plausible identity phrase without retrieving a trustworthy internal identity record.

Does an LLM know which model it is?

Not reliably enough for verification. In the 开桥 test, 85.2% of responses named the correct vendor, but 14.8% either named another vendor, refused, or gave no specific identity.

Does a wrong answer mean the API gateway swapped my model?

Not by itself. Check the requested model, response.model, routing configuration, provider, and request log. Those records are more reliable than generated self-description.

Does identity confusion prove that one model was distilled from another?

No. The experiment did not inspect training data or weights. Identity confusion is compatible with several explanations and cannot establish distillation on its own.

Can I reproduce the study?

Yes. The 开桥 Arena repository contains the code, configuration, research pipeline, and result files. The interactive studio lets you inspect and export the relationship graph.

Explore the evidence—and test models on your own workload

The most useful lesson is not that one model “knows itself” better than another. It is that generated claims should be separated from verifiable system evidence.

Explore all 29,700 responses in 开桥 Arena, then use the 开桥 Quick Start to compare models through one API. Judge them by the evidence that matters for your application: output quality, reliability, latency, cost, and sustained performance on real tasks.