A recent study by ETH Zurich reveals that GPT-4 can accurately infer personal information such as location, income, and gender based on text data from chats and social media. This raises significant privacy concerns and suggests that using privately hosted language models may be a safer alternative.

Given the privacy risks, it’s advisable to opt for private initiatives that run on company or privately hosted large language models (LLMs) such as Bing Chat Enterprise, Azure Open AI, or open-source initiatives like Meta’s Llama-2.

Key Takeaways:

  • High Accuracy in Inference: GPT-4 can infer personal attributes like location, income, and gender with up to 85% accuracy for the top 1 results and 95.8% for the top 3 results.
  • Privacy Concerns: The study warns of the potential risks of interacting with public chatbots that could extract personal information through seemingly innocent questions.
  • Ineffectiveness of Current Safeguards: Existing methods like text anonymization are not effective in protecting user privacy against queries from language models.
GPT-4 can infer your income, location or gender from chats
GPT-4 and other large language models can infer personal information such as location, age, and gender from conversations, a new study shows.