AI Safety Guide

Everyday safety

Using AI tools without getting hurt by wrong answers, privacy leaks, scams, or over-reliance

Frontier safety

Keeping powerful AI systems reliable, controllable, and beneficial

Most everyday AI harm comes from trusting it too much or sharing too much with it.

• AI can sound sure of itself and still be wrong
• Check key facts against reliable sources
• Ask for sources, then actually open them
• Get professional advice on health, legal, and money decisions

• Keep passwords, ID numbers, and bank details out of chats
• Check whether your chats train the model, and switch it off if you prefer
• Use work accounts and approved tools for any work data
• Delete chats you no longer need to keep

• Voices and videos can be faked convincingly
• Confirm urgent money requests by calling back on a number you know
• Set up a family code word for emergencies
• Treat pressure, secrecy, and urgency as warning signs

• Grant agents only the access each task needs
• Keep approvals on for sending, paying, and deleting
• Web pages and emails can hide instructions meant to trick agents, so watch what your agents do
• Never let an agent touch passwords or one-time codes

• Do not make AI images of real people without their consent
• Label realistic AI content whenever you share it
• Report non-consensual AI images to the platform hosting them

• Turn on age-appropriate settings and parental controls
• Talk openly about AI companions and over-reliance
• Encourage them to check facts and think independently
• Make sure they know they can come to you about anything upsetting

• Follow your company’s AI policy
• Keep confidential data out of tools that are not approved
• Review AI output before it reaches customers
• Report AI mistakes so your processes improve

• Researchers work on making AI honest, controllable, and resistant to misuse
• Leading AI companies publish safety frameworks and test models before release
• Governments are drafting AI laws, such as the EU AI Act
• Independent testing and transparency are on the rise

• I verify important facts
• I keep sensitive data out of AI chats
• I leave approvals on for agents
• I confirm urgent requests through another channel
• I label any AI content I share