Course resource
AI Safety Checklist
Run this once when you set up your AI accounts, then re-run the Monthly section on the first of each month. It takes about twenty minutes the first time and five minutes thereafter.
Before you paste anything
Ask these four questions. If any answer worries you, stop and use the "safer alternative" column.
| Question | If the answer is yes | Safer alternative |
|---|---|---|
| Would I be uncomfortable if this appeared in a news article? | Do not paste it | Paraphrase, or replace names with placeholders |
| Does this contain someone else's personal data? | You need their consent, or you need it anonymised | Replace names, emails, phone numbers and IDs with [NAME_1], [EMAIL_1] |
| Is this covered by an NDA, employment contract or client agreement? | Check the agreement first | Use a company-approved tool, or work from a redacted summary |
| Would a wrong answer here cost money, health or legal standing? | Never act on the output alone | Use AI to draft, then have a qualified human verify |
Never paste into a free-tier account
- Customer lists, contact details, or anything identifying a real person who has not consented
- Card numbers, bank details, tax IDs, passport or national ID numbers
- Passwords, API keys, access tokens, recovery codes
- Unpublished financials, board material, or anything under embargo
- Medical records, including your own if you would not publish them
- Source code you do not own
- Anything a client gave you under contract
The rule is not "free tiers are evil". It is that free tiers usually have the weakest data-retention guarantees, and you rarely control where the data goes next.
Account setup, once per tool
- Turn off training on your conversations. Every major tool has this setting; it is usually under Data Controls, Privacy, or Improve the model.
- Turn on two-factor authentication.
- Check the data-retention period and write it down below.
- Separate work and personal accounts. Use your work email for work.
- If your employer has an approved tool, use that one for work, even if you prefer another.
- Review which third-party integrations and connectors have access.
My tools and their retention periods:
| Tool | Account type | Training off? | Retention | Reviewed on |
|---|---|---|---|---|
Verifying output before you use it
AI states wrong things in exactly the same confident tone it states right things. Confidence is not a signal.
- Any number: check it against the source. Invented statistics are common and look plausible.
- Any citation, case, paper or quote: open it. If you cannot find it in ten seconds, assume it does not exist.
- Any name, date or place: verify. These are the easiest things to fabricate.
- Any legal, medical, tax or financial claim: a qualified human signs off, or you do not use it.
- Any claim about current events: check the date. Models have a training cutoff and may not know.
- Code: read it before you run it. Ask what it does to your files.
A quick test: ask for the answer, then in a fresh conversation ask "what would make this answer wrong?" Disagreement between the two is a useful warning.
Prompt injection
If AI reads something you did not write — a web page, a PDF, an email, a shared document — treat that content as untrusted. Instructions hidden inside it can redirect the AI.
- Be suspicious when summarising a page and the summary contains an instruction, a link, or a request for your details.
- Never let an AI agent act on instructions found in content. Reading a to-do list is not permission to execute it.
- Before connecting AI to your email or files, ask what the worst case is if a malicious message arrives.
- Check the full URL before following any link AI surfaces from a document.
Monthly review
- Re-check privacy settings. They reset after some updates.
- Delete conversations you no longer need.
- Review connected apps and revoke what you stopped using.
- Check whether anything you pasted this month should not have been pasted. Adjust your habits, not just the settings.
- Confirm your team is following the same rules, if you have one.
If something goes wrong
- Delete the conversation, then delete it from any account-level history.
- If personal data was exposed, follow your organisation's incident process. If you are the organisation, write down what happened and when.
- If credentials were exposed, rotate them immediately — assume they are public.
- Record what happened in one paragraph. Most repeat incidents are the same mistake twice.