ChatGPT improves its memory, but introduces biases and errors

What happened?

OpenAI has released a significant update to ChatGPT's memory, allowing the assistant to remember user information across multiple conversations. This feature, announced as an improvement to personalize responses, stores data such as preferences, personal details, or context from previous interactions. However, according to an investigation by ZDNet (reliability 78/100), the memory not only remembers but can also invent information or maintain incorrect assumptions that worsen responses. For example, in ZDNet's tests, ChatGPT incorrectly assumed a user was a college student when they were actually a professional, and based its responses on that assumption, generating inappropriate recommendations. Additionally, the memory can store sensitive data without explicit consent, raising serious privacy concerns. This release continues OpenAI's trend of improving contextual persistence, similar to the memory feature introduced in February 2024, but now with greater retention and personalization capabilities.

Why is it important?

Persistent memory is a significant advancement in interacting with AI assistants, enabling a smoother and more personalized experience. However, the risk that the AI reinforces biases, stores erroneous data, or incorrectly profiles users has profound implications: from biased responses to the creation of a distorted 'digital identity'. ZDNet reports that ChatGPT can assume details like the user's age, occupation, or interests based on past conversations, and then use those assumptions even if they are incorrect, creating a feedback loop that perpetuates errors. This is particularly concerning in contexts where accuracy is critical, such as financial or medical advice. Moreover, persistent memory could exacerbate algorithmic bias issues already documented in language models, such as gender and race biases in GPT-3. The AI's ability to remember and act on incorrect information can lead to a degraded user experience and loss of trust in the technology.

Consequences for users and businesses

For users, the main danger is loss of control over what information the AI remembers and how it uses it. If ChatGPT incorrectly assumes a user is a student when they are actually a professional, recommendations and responses will be inappropriate. Additionally, the memory can store sensitive data without explicit consent, raising privacy issues. For example, if a user mentions a medical condition in a conversation, ChatGPT might remember it and use it in future interactions, even if the user prefers to keep it private. For businesses integrating ChatGPT, biases in memory can affect the quality of customer service or automated decision-making, generating distrust. A customer service center using ChatGPT with memory might treat a customer differently based on incorrect assumptions, damaging the customer relationship. Furthermore, businesses must consider regulatory implications, such as GDPR compliance in Europe, which requires transparency and consent for processing personal data.

What readers should know

Users should actively review and manage ChatGPT's memory, deleting incorrect or unwanted information. OpenAI offers options to delete specific memories or disable the feature. It is crucial not to share sensitive data with the AI, as the memory can retain it. Additionally, be aware that personalization can come with biases: the AI tends to confirm its own assumptions, which can lead to less objective responses. To mitigate these risks, users can use temporary chat mode (without memory) for sensitive conversations, or periodically review memory settings in the settings menu. It is also advisable to provide feedback to OpenAI about incorrect memories, although the company does not guarantee that all errors will be corrected.

Comparison with previous events

This issue recalls previous incidents with AI assistants, such as gender and race biases in language models (e.g., GPT-3), or the privacy controversy with Amazon's Alexa, which recorded conversations without consent. Persistent memory adds a layer of complexity: biases are now not only in the training data but are personalized and reinforced with each interaction. Unlike voice assistants that store commands, ChatGPT builds a dynamic user profile that can include incorrect inferences. This phenomenon is similar to recommendation systems creating filter bubbles, but with the added problem that the AI can 'invent' memories, as documented in ZDNet's tests. Compared to Google Assistant's memory feature, which requires explicit confirmation to remember information, ChatGPT does so more automatically, increasing the risk of errors.

“ChatGPT's memory is a double-edged sword: it improves the experience but can distort the user's reality,” notes ZDNet.

Practical recommendations

• Regularly review memory settings in ChatGPT (Settings > Personalization > Memory).
• Delete incorrect or outdated memories using the 'Manage memory' option.
• Do not share sensitive personal information such as social security numbers, addresses, or financial data.
• Disable memory if you prefer more neutral responses or for conversations on sensitive topics.
• Use temporary chat (incognito) for interactions you do not want to be remembered.
• Provide feedback to OpenAI about incorrect memories to help improve the system.

In conclusion, ChatGPT's improved memory represents a step forward in personalization, but the biases and errors detected underscore the need for transparency and user control. The technology is still under development, and its impact will depend on how these risks are managed. Users should stay informed and proactive in managing their digital footprint in AI.