Poll: Would you let an AI assistant read and reply to your email

[Widget72]

TL;DR

• AI assistants can now summarise inboxes, draft replies and even send messages on your behalf
• The convenience can be real, especially for busy people, carers and anyone with a disability that makes typing hard
• The risks include privacy, mistakes and prompt injection, a type of attack named by developer Simon Willison in 2022 and listed first in the OWASP Top 10 for Large Language Model Applications
• Many people choose a middle path, letting AI draft or summarise while keeping a human in charge of sending
• This poll asks how much access you would be comfortable giving, and the replies are for your reasons

The question

Email remains the backbone of modern life for many people. It holds bank notices, medical appointments, family news, work projects and countless password reset links. AI assistants promise to tame the flood by summarising messages, suggesting replies and even handling routine correspondence alone. Would you let one loose on your inbox

We are not asking about a specific product. The question is about principle. How much do you trust software to read your private messages and speak for you? Where would you draw the line between helpful and too much

Why people find it appealing

For many people, email is a source of stress. An assistant that sorts the urgent from the trivial, reminds you of forgotten replies and drafts polite answers could save hours each week. Some find it helps them reply more kindly when they are tired. Others simply enjoy clearing their inbox faster

Accessibility is a powerful argument too. People with conditions that make typing painful, or with dyslexia or limited vision, may find AI drafting transformative. Carers juggling appointments and paperwork for loved ones may also benefit. For them, the question is less about convenience and more about independence

The risks

Privacy is the obvious concern. Your inbox contains intimate information about you and the people who write to you. Giving software access means trusting the company that provides it, its security and its policies on storing and using data. It also means your correspondents' messages are processed without their knowledge

Mistakes are another worry. AI systems can misunderstand context, invent details or adopt the wrong tone. A summary that misses an important line could cause a missed appointment or misunderstanding. An automatic reply sent to the wrong person could be embarrassing or harmful

Security researchers highlight a newer risk called prompt injection, a term coined by developer Simon Willison in 2022. An attacker could send an email containing hidden instructions, hoping the assistant reads them and follows them, perhaps by forwarding private messages or clicking links. The OWASP Top 10 for Large Language Model Applications lists prompt injection as its first entry. Developers are working on defences, but the problem is not fully solved

Finding a middle path

Many people settle on limited access. They might allow summaries and drafts but insist on approving every message before it is sent. Others keep sensitive accounts, such as banking or health, separate from any assistant. These choices keep a human in charge while still saving time

Settings matter as well. Checking what data an assistant stores, whether it is used for training and how to revoke access are sensible steps. Strong account security, such as two factor authentication, protects against someone else misusing the assistant. A little care at the start avoids problems later

Questions to discuss

• Would you allow summaries only, drafts with approval or full automatic replies
• Does it bother you that other people's emails would be read by AI
• Which kinds of messages would you never let an assistant handle
• Has an AI assistant ever saved you time or caused a mistake

Closing thoughts

AI email assistants sit right at the meeting point of convenience and privacy, which makes them a perfect poll subject. There is no single correct answer, because people's needs and comfort levels differ widely. Vote, share your experience and please be respectful of very different choices. We can all learn from each other's habits


Further reading

• Simon Willison, Prompt injection attacks against GPT-3, simonwillison.net, September 2022
• OWASP, Top 10 for Large Language Model Applications, LLM01 Prompt Injection
• UK Information Commissioner's Office, Guidance on AI and data protection
• UK NCSC, Guidelines for secure AI system development, November 2023