My Essential Guidelines for Secure AI Use at Google

Artificial intelligence has seamlessly integrated into my everyday routines, becoming an indispensable tool for tasks like in-depth research, jotting down notes, programming, and browsing the web. However, working at Google has heightened my awareness of the privacy issues linked to AI usage. Since joining Google in 2023, I've dedicated two years as a software engineer on the privacy team, fortifying infrastructure to safeguard user data. Presently, I'm part of the Chrome AI security squad, tasked with shielding Google Chrome from cyber threats, including hackers and those exploiting AI to orchestrate phishing attacks.

AI models rely on data to craft useful responses, making it imperative for us as users to shield our personal data from potential exploitation by cybercriminals and data harvesters. To protect my information when engaging with AI, I've established four key practices.

Perceive AI as a Transparent Message

The allure of AI might sometimes create a misleading sense of confidentiality, encouraging individuals to divulge information online they typically wouldn't. Nonetheless, IT professionals could be working to boost privacy within AI models, yet it's wise to refrain from sharing sensitive details like credit cards, Social Security numbers, residential addresses, medical history, or other personally identifiable information with AI chatbots.

When data is submitted to public AI chatbots, it might be utilized to refine future models, potentially causing what's called 'training leakage.' This refers to a scenario where the model retains one user's personal data and subsequently includes it in communications with another user. Additionally, there's always the lurking danger of data breaches that could unveil your chatbot-shared data.

I regard AI chatbots as akin to a public postcard. If I wouldn't want to write something on a postcard for all to see, I refrain from sharing it with a public AI system. The certainty of how my data might be leveraged in future training remains elusive.

Assess Your Conversational Environment

It's crucial to discern whether you're utilizing a public AI tool or a premium enterprise-grade version. The manner in which conversations are leveraged to train public AI models remains unclear, but some companies opt for enterprise models. These versions generally exclude training on user dialogues, offering employees a safer platform to discuss work and company-related matters.

Imagine speaking in a bustling café, where eavesdropping is possible, versus a confidential meeting in your private office. There have been reports of employees inadvertently leaking company data to ChatGPT. Therefore, if you're engaged in confidential company projects or working on securing a patent, it might be prudent to avoid discussing these plans with a non-enterprise chatbot to mitigate the risk of sharing sensitive information.

Regarding my work at Google, I consistently avoid discussing projects using public chatbots. Instead, I opt for enterprise models, even for seemingly minor tasks like editing an email. This way, my conversations are not utilized for training, though I still ensure that minimal personal information is shared.

Regularly Erase Interaction Records

Chatbots often retain history of various exchanges. To secure long-term user privacy, I advocate for routinely deleting these records from both public and enterprise models. Even without entering private data, the chance of account compromise merits precaution.

I experienced this firsthand when an enterprise chatbot surprisingly provided my exact address, which I hadn't intentionally submitted. Upon reflection, I recalled previously asking it to refine an email containing my address. Because of the tool's long-term memory capabilities, it could remember and store this detail.

Occasionally, if I wish for certain inquiries to vanish from chatbot memory, I switch to a special mode, resembling incognito browsing, where bots don't log past conversations or utilize them for model training. Known as the 'temporary chat' feature, this mode is available in ChatGPT and Gemini.

Opt for Established AI Applications

Choosing reputable AI tools with robust privacy frameworks and clear protections in place is advisable. Besides Google's offerings, I favor OpenAI's ChatGPT and Anthropic's Claude.

It's advantageous to review the privacy policies of any AI applications used. These can elucidate how data might be employed for model training. Most privacy settings provide an option to 'enhance the model for everyone.' By ensuring this is disabled, you can hinder the inclusion of your conversations in training processes.

AI technology wields immense capabilities, yet we should remain vigilant to uphold the security of our personal data and identities when engaging with it.