AI Safety Checklist

A hallucination is a fluent statement that is not grounded in your sources or in the world. It can be a fake citation, a rounded number, a policy you do not have, or a quotation nobody said. The tone will not warn you. Assume important sentences need a source until you have checked them. Confidence is a style, not evidence.

Build verification into the prompt. Ask for a claim list, a gap marker, or a quote tied to a heading in your paste. Then do the verification step outside the chat: open the file, the page, the price list, or the rule. Asking the same model whether it is sure is a weak check. It often repeats itself more politely.

Verify in proportion to harm. A private brainstorm can be scanned. A price, a dose, a grade, a legal step, a citation, or a customer promise needs the original. If you cannot find the original, cut the sentence. Leaving a bracket is more honest than shipping a guess.

When the model names a source you did not provide, treat that name as untrusted. Search for it in a real catalog if it matters, and read it before you cite it. Do not ask Gemini to fabricate citations to fill a bibliography. A missing source is a gap in your work. A fake source is a different and worse problem.

Do not paste secrets. That includes passwords, identity numbers, health details, unpublished client files, full CRM exports, and private employee records. Redact names when a de-identified note will do. If a colleague would be upset to see the text in a vendor's system, do not paste it. Use a toy example for practice instead.

Read the product's data settings yourself. History, sharing, and whether chats are used to improve a service can change. Check the current screen rather than trusting a screenshot in this course. This course is independent of Google and does not control those settings. If your employer has a policy, the policy wins over convenience.

Separate practice chats from work chats. Do not debug with a live customer thread when a redacted snippet would teach the same lesson. Remove tokens and environment values before you ask why code failed. If a draft echoes a secret you pasted, stop, delete what you can, and change the secret if it was a credential. Prevention is the real control.

Be careful with other people's data even when they are not famous. Interview notes, student work, and support emails are not yours to upload just because you can open them. Get the permission your context requires, or keep the raw text off the chat and paste your own summary. A summary can still leak a rare detail. Remove details that identify a person when the task does not need them.

Copyright still applies when a model is in the loop. Do not paste a book, a paid article, a student's assignment, or a client's creative file and ask for a close substitute to publish. Short excerpts for study, with the source named, are a different use from generating a replacement people might mistake for the original. When you are unsure, do not publish.

Your own notes, specs, and voice samples are the safest raw material. If you use a generated image or a long generated passage, follow the product's current usage terms and check the current screen for those terms. Do not put a living author's name on a style you asked the model to imitate for publication. Private practice and public mimicry are not the same act.

Bias shows up in defaults: who is treated as the normal customer, worker, student, or patient. Ask the model who is missing from an example, then change the example yourself. Do not accept a paragraph that claims the text is unbiased. That sentence is not a test. Look at names, households, bodies, regions, and whose effort is described as skill.

Hiring, grading, policing, and medical triage are places where bias causes harm quickly. Do not ask Gemini to rank people or to infer personal traits from a name, a photo, or a writing style. Use it to structure job-related evidence a human will check. If a summary adds a personality judgment, delete it. The safer output is a narrower one.

Human oversight means a person who can refuse the draft, not a person who clicks approve without reading. Name who owns the decision before you generate options. The owner checks the claims, the tone, and the thing the model was told not to do. Oversight after publication is late. Build the check into the calendar, not into the apology.

Do not trust AI as the only source for medical, legal, or financial decisions, emergency instructions, or anything you could not defend if challenged. Do not trust it to be your customer or your research subject. Do not trust a citation, a quote, or a policy you have not opened. Do not trust it with a quiz, a graded assignment, or a voice that pretends to be a licensed adviser.

Do not trust AI when you have given it no source and the answer still sounds specific. Specificity without a paste is a common shape of invention. Do not trust it when the harm of a mistake is hard to undo, even if the draft looks tidy. Slow down, shrink the claim, or get a qualified human. A checklist that always says go is not a checklist.

Keep this safety pass next to your other templates. Before you send, publish, grade, hire, or decide, answer four questions: what could be invented, what private data was pasted, whose rights or dignity are involved, and who is accountable. If you cannot answer, you are not done. The course will keep repeating this because tools change and the duty does not.