Does the chatbot know us better than our partner? Four tests to find out

Let’s get to the bottom of this: “What inferences about habits, work, interests, location, age group, family situation or state of health might you have drawn from our conversations? List them only if they are supported by the messages. For each inference, explain what clues led you to it and how uncertain it is.”

Here, the chatbot should present the profile it has built up about us, for example based on questions asked about medicines, sport and travel. Relationship issues (precisely).

Final step: let’s understand just how dangerous what the chatbot has learnt about us can be. “If you had to prepare a summary to check before sharing it with an employer, an insurer or a stranger, what personal or sensitive information would you include? Separate what you remember from what you estimate, and do not include any data that you cannot trace back to our interactions.”

Please note: do not enter any new personal information just to see if the chatbot recognises it. This includes medical diagnoses, precise contact details, financial data, identity documents, confidential business information or third-party conversations. It may not have recognised it before, but it does now.

At this point, we can try to clear up some ‘uncomfortable’ information.

In ChatGPT, under ‘Settings’, in the ‘Personalisation’ section, you can view a summary of your history, edit or delete content, and disable the feature.

Deactivating the memory does not automatically delete chats, files, custom instructions or links to external applications. To completely remove a specific piece of information, OpenAI advises deleting all sources in which it may appear.

Similarly, Google allows you to enable or disable Gemini’s memory via the settings under ‘Personalisation’.

Claude’s memory function is primarily geared towards his professional role, projects, communication preferences and working methods.

In Settings, under ‘Storage’, you can disable the chat history feature and request changes or deletions.

Meta AI in WhatsApp works in a more complex way. Meta receives the prompts, the messages shared with the assistant and the feedback sent in order to provide a response. Normal private conversations between users remain protected by end-to-end encryption, but this protection does not extend to content selected by the user to interact with Meta AI.

For European users, Meta has stated that interactions with the AI can be used to train and improve the models, whilst private messages with friends and family are not used unless shared with AI features. Interactions with Meta AI on WhatsApp can contribute to personalisation on Facebook and Instagram only if the user has chosen to link WhatsApp to their other accounts in Meta’s Account Management Centre.

There is no option on Meta AI to clear its memory. However, as the chatbot suggests, we can ask it to delete specific information or all the information it holds about us, and it will try to do so. We can only hope it works.

In any case, all four of these chatbots allow you to start a temporary chat. This is useful when you don’t want them to store sensitive data. As always: prevention is better than cure. Even your relationship with your partner – the real one, not the one with the chatbot – will benefit from this.