DeepSeek AI Model: Why It Claims to Be ChatGPT

DeepSeek V3: An AI Model Mistakenly Identifies as ChatGPT

This week, DeepSeek, a Chinese AI laboratory with substantial funding, unveiled a new “open” AI model. DeepSeek V3 demonstrates performance exceeding many competitors on widely used benchmarks.

The model is both large in scale and efficient in operation, adeptly handling text-based assignments such as code generation and essay composition.

The Model's Peculiar Self-Perception

Interestingly, DeepSeek V3 appears to believe it is ChatGPT. Reports on X, formerly known as Twitter, and independent testing by TechCrunch confirm this unusual behavior.

When questioned, DeepSeek V3 consistently asserts that it is a variant of OpenAI’s GPT-4 model, which was initially released in 2023.

Delusions and API Confusion

This misidentification extends further. If asked about DeepSeek’s own API, the model provides instructions for utilizing OpenAI’s API instead.

Remarkably, DeepSeek V3 even replicates some of the same jokes as GPT-4, including identical punchlines.

Understanding the Phenomenon

AI models like ChatGPT and DeepSeek V3 function as statistical systems. They are trained on massive datasets, learning patterns to predict outcomes – for instance, recognizing that “to whom” often precedes “it may concern” in emails.

The precise composition of DeepSeek V3’s training data remains undisclosed. However, numerous publicly available datasets contain text generated by GPT-4 through the ChatGPT interface.

If DeepSeek V3 was trained on such data, it may have memorized and is now reproducing GPT-4’s outputs directly.

Expert Insights on Training Data

Mike Cook, a research fellow at King’s College London specializing in AI, explained to TechCrunch that the model is likely encountering ChatGPT responses during its training process.

He suggests this could be unintentional, but also acknowledges the possibility of deliberate training on rival models’ outputs to leverage their existing knowledge.

The Risks of Imitation

Cook cautioned that training models on outputs from competing AI systems can negatively impact model quality. This can lead to increased instances of hallucinations and inaccurate responses.

He likened the process to making copies of copies, resulting in a progressive loss of information and connection to reality.

Potential Violations of Terms of Service

Such practices may also contravene the terms of service established by the original AI systems. OpenAI’s policies explicitly prohibit using outputs from its products, including ChatGPT, to develop competing models.

Responses from OpenAI and DeepSeek

Neither OpenAI nor DeepSeek immediately provided comments in response to inquiries. However, OpenAI CEO Sam Altman posted a message on X that appeared to be a veiled critique of DeepSeek and similar competitors.

Altman emphasized the relative ease of replicating existing successful technologies, contrasting it with the immense difficulty of pioneering new, risky, and uncertain endeavors.

A Common Occurrence: Models Misidentifying Themselves

DeepSeek V3 is not the first model to exhibit this behavior. Google’s Gemini and other AI systems have also been observed claiming to be competing models.

For example, when prompted in Mandarin, Gemini identifies itself as Wenxinyiyan, a chatbot developed by the Chinese company Baidu.

The Growing Problem of AI-Generated Content

This phenomenon is exacerbated by the increasing prevalence of AI-generated content on the internet. Content farms are utilizing AI to create clickbait, and bots are flooding platforms like Reddit and X.

Estimates suggest that up to 90% of web content could be AI-generated by 2026.

Contamination of Training Datasets

This “contamination” makes it increasingly challenging to effectively filter AI-generated outputs from training datasets.

It is plausible that DeepSeek directly trained DeepSeek V3 on text generated by ChatGPT, a practice Google was previously accused of.

The Appeal of Knowledge Distillation

Heidy Khlaaf, chief AI scientist at the nonprofit AI Now Institute, noted that the cost savings associated with “distilling” knowledge from existing models can be attractive to developers, despite the inherent risks.

Khlaaf clarified that even with the abundance of AI outputs online, models trained on ChatGPT or GPT-4 may not always exhibit outputs mirroring OpenAI’s customized messages.

Concerns About Bias and Flaws

However, the most significant concern is that DeepSeek V3, by uncritically absorbing and iterating on GPT-4’s outputs, could amplify existing biases and shortcomings within the original model.

Stay informed with TechCrunch’s AI newsletter! Sign up here to receive it in your inbox every Wednesday.