Signal CEO Meredith Whittaker, in an interview with Bloomberg, stated that users should not treat chatbots like ChatGPT and Claude as “friends” or entities with subjective awareness. She focused the discussion on privacy and permissions, emphasizing that the risks of such tools escalate rapidly once they become deeply integrated into personal services.
Whittaker on the boundaries of AI use
Whittaker said she uses AI tools only in a few scenarios, such as formatting documents, but does not delegate the processes of asking questions, thinking, or writing to models. She stated that she does not want a system that generates answers based on existing information to replace the personal thinking process that should be done by individuals.
Concern about granting sensitive permissions across applications
She also responded to a previous idea proposed by Mustafa Suleyman, head of Microsoft’s AI business, that users could have Microsoft Copilot handle their Christmas shopping. Whittaker noted that for such a scenario to be realized, the system would need to access a large amount of personal information and perform actions across multiple services.
- Credit Card and Payment Information
- Browser browsing history
- Signal messages and contacts
In addition to the above information, such systems may also access home addresses, calendar schedules, and send messages on behalf of users to friends and family. She believes this goes beyond the auxiliary use of a single tool and constitutes broad access to multiple applications and services.
Signal considers such access a backdoor risk.
Whittaker further stated that introducing similar capabilities into encrypted communication products like Signal would be nearly equivalent to setting up a “backdoor” from a product security perspective. She used this to emphasize that the closer AI assistant functions come to acting as delegates or agents on behalf of users, the more users need to pay attention to what actual permissions they are granting.
This statement also reflects an ongoing debate in the AI assistant competition: as products shift from answering questions to performing tasks on behalf of users, the conflict between convenience and privacy protection becomes more direct.