Plain-language summary
What this guide covers
AI safety means using AI in ways that reduce avoidable harm. It includes checking facts, protecting sensitive information, watching for scams and deepfakes, respecting copyright, reducing bias, following workplace and school rules, avoiding unsafe professional advice, and knowing when to escalate to a person or official channel.
AI tools can make work faster, but speed can hide risk. A wrong summary can mislead a team. A copied prompt can expose private data. A fake voice can trick a family member. A biased workflow can affect access or trust. Safety habits help ordinary users pause before a small mistake becomes a serious problem.
What you will learn
- Use a stop-and-check framework before relying on AI output.
- Identify sensitive data, scam, deepfake, bias, copyright, security, and workplace-rule risks.
- Explain when AI should not be used without qualified human review.
- Apply emergency and escalation boundaries for serious situations.
- Choose safer links to playbooks and prompts for low-risk practice.
Guide section
The stop-and-check framework
A simple framework can prevent many common AI mistakes before they happen.
SAFE: a five-step check
- Scope: State the task and the decision the output might influence.
- Access: Check whether the tool is approved for the information and setting.
- Facts: Verify important claims, names, dates, numbers, sources, and policy statements.
- Effects: Ask who could be affected if the output is wrong, biased, private, or misleading.
- Escalate: Move high-risk, sensitive, unclear, or urgent cases to the right human or official process.
Example
Example: customer email draft
A small business owner asks AI to draft a customer reply. SAFE asks: What is the task? Is the tool approved for customer data? Are refund promises correct? Could the reply affect trust or money? Should a manager approve it before sending? The result may still be useful, but it is treated as a draft, not a decision.
Guide section
Verification and overreliance
AI can sound confident even when it is wrong, incomplete, or unsupported.
Verification means checking an output against evidence outside the model. NIST identifies generative AI risks including confabulation, which is a polished but false or unsupported output. Verification matters most when an answer includes facts, dates, numbers, citations, policy statements, product claims, safety guidance, or instructions that affect another person.
Verify before using
- Check all dates, numbers, names, titles, and source claims.
- Compare important claims with trusted sources, not only another AI answer.
- Ask what the output assumes and what information may be missing.
- Check whether the answer is too broad for the evidence.
- Watch for fake citations, invented policies, and made-up product features.
- Keep a human owner for any final decision or public use.
Guide section
Sensitive data and privacy
Some information should not be entered into an AI tool unless the tool and process are approved.
Sensitive data includes information that could identify, expose, harm, or unfairly affect a person or organization. Privacy risk is not only a personal concern. FTC and NIST guidance frame privacy and security as organizational responsibilities that require clear promises, data minimization, safeguards, and risk management.
Do not enter into unapproved AI tools
- Passwords, security codes, keys, tokens, or access links.
- Government ID numbers, account numbers, payment data, or tax records.
- Medical, legal, financial, immigration, or mental-health details.
- Student, employee, customer, client, patient, or applicant records.
- Private messages, private photos, private audio, or family emergencies.
- Confidential business plans, contracts, internal notes, source code, or unreleased product details.
- Information about children or teens unless the tool and process are clearly approved for that setting.
Guide section
Phishing, scams, deepfakes, and impersonation
AI can make fake messages, calls, images, and videos more believable.
Phishing is a trick that tries to get you to reveal information, send money, click a link, or open a harmful file. AI can help scammers write better messages or imitate voices. The FTC has warned that voice cloning can make a request sound like it came from a boss or family member. In the United States, the FCC stated in February 2024 that AI-generated voices in robocalls are artificial voices under the Telephone Consumer Protection Act.
Scam check
- Pause when a message creates urgency, secrecy, fear, romance, or pressure to pay.
- Verify through a trusted channel you already know, not the link or number in the message.
- Use a family or workplace verification phrase for urgent money or emergency requests.
- Do not trust a voice, image, or video only because it sounds or looks familiar.
- Report suspicious messages through workplace, school, bank, platform, or law-enforcement channels when appropriate.
- Do not forward suspected deepfakes as proof; label uncertainty and verify first.
Guide section
Bias, accessibility, and copyright
Responsible AI use also includes fairness, access, and rights.
Bias can appear when data, labels, design, deployment, or review processes produce unfair results. Accessibility means people with disabilities can use information and tools; WCAG is an international web accessibility standard. Copyright questions around AI are changing, and the U.S. Copyright Office has issued guidance and reports on AI-generated material and human authorship. For education, UNESCO emphasizes human-centered use, policy, privacy, and capacity building.
Fair access and rights check
- Who could be underrepresented, mislabeled, excluded, or harmed?
- Can people with disabilities use the content or process?
- Does the output need alt text, captions, plain language, keyboard access, or human support?
- Are sources, permissions, licenses, and human contributions clear enough for the use?
- Could the output imitate a person, style, brand, or protected work in a risky way?
- Is there a path for appeal, correction, or human review?
Guide section
Security, workplace rules, students, and professional boundaries
The right safety rule depends on where and why AI is being used.
Security risk includes data leakage, unsafe code, harmful links, prompt injection, weak access controls, or using tools outside policy. Workplace and school rules matter because an action that feels harmless at home may be disallowed with employer data or student work. AI Revolution Atlas also sets a professional-advice boundary: AI output should not be treated as legal, medical, tax, financial, mental-health, safety, or emergency advice.
| Situation | Safer response | Escalate when |
|---|---|---|
| Workplace task | Use approved tools, follow policy, and keep review steps. | The task uses confidential data, customers, employees, code, contracts, money, safety, or legal duties. |
| Student use | Follow class rules, disclose when required, and use AI for learning support rather than hidden substitution. | The task affects grades, accommodations, identity, discipline, or student records. |
| Health, legal, tax, financial, or mental-health question | Use AI only for general educational background, then consult qualified sources. | A decision, deadline, symptom, contract, account, diagnosis, safety issue, or legal right is involved. |
| Security concern | Do not click, download, run code, share credentials, or follow urgent requests without verification. | There may be a breach, fraud, malware, account takeover, or unsafe code. |
Guide section
Incident and emergency boundaries
Some situations should not be handled through AI at all.
Escalate immediately when
- There is immediate danger, violence, self-harm risk, abuse, exploitation, or a medical emergency.
- There is suspected fraud, theft, account takeover, malware, data breach, or identity misuse.
- Someone is asking for money, credentials, private files, or secrecy under urgent pressure.
- A child or student may be at risk or their records may be exposed.
- A decision may affect legal rights, housing, employment, benefits, education, safety, health, or finances.
- You are unsure whether your tool, policy, or authority allows the use.
Try it
Exercise: create a personal AI safety card
Create a small note you can use before working with AI. Do not include private details. Write your trusted verification channels, your sensitive-data stop list, and three escalation contacts or processes, such as workplace IT, school policy office, bank fraud line, or local emergency services.
- Write your sensitive-data stop list.
- Write your trusted verification method for urgent calls or messages.
- Write where to report phishing or account compromise.
- Write your school or workplace AI-rule source if you have one.
- Write one low-risk task that is safe for practice.
Avoidable errors
Common mistakes and better approaches
Trusting AI output because it sounds confident.
Better approach: Verify important claims with trusted sources and keep a human decision owner.
Pasting private information to get a better answer.
Better approach: Use public, fictional, de-identified, or approved information only.
Assuming you can spot every deepfake or scam by sight or sound.
Better approach: Verify through a separate trusted channel before acting.
Using AI where a professional or emergency channel is needed.
Better approach: Escalate when health, safety, law, money, children, credentials, or urgent harm is involved.
Treating prompts as safety controls by themselves.
Better approach: Use prompts with policy, verification, access control, human review, and escalation.
Remember this
Key takeaways
- AI safety is practical: scope the task, check access, verify facts, consider effects, and escalate when needed.
- Sensitive data should not enter unapproved tools.
- Scams and deepfakes require separate-channel verification.
- Bias, accessibility, copyright, and security are part of everyday AI use.
- Students and workers must follow local rules, not only general AI advice.
- AI should not replace qualified professional or emergency help.
- Safety improves when checkpoints are written before a task begins.
Questions readers ask
Frequently asked questions
What is the first AI safety question to ask?
Ask whether the task uses sensitive information or could affect another person in a meaningful way. If yes, pause and check tool approval, policy, and review needs.
Can a better prompt make AI output safe?
No. A better prompt can improve structure and usefulness, but safety also requires verification, privacy controls, human review, and escalation rules.
How should I handle a suspicious AI voice call?
Do not act on the call alone. Hang up or pause, then verify through a trusted number or channel you already know. Use a family or workplace verification phrase for urgent requests.
Can students use AI safely?
Students can use AI safely when school rules allow it, privacy is protected, the tool supports learning rather than hidden substitution, and important outputs are checked by people.
What if a workplace has no AI policy?
Use extra caution. Avoid confidential data, customer records, employee records, private files, credentials, and external publication until a manager, IT, legal, compliance, or policy owner gives guidance.
Sources and review notes
Sources were accessed on the dates shown. Links open the original organization’s page.
- SRC-01Artificial Intelligence Risk Management FrameworkNational Institute of Standards and Technology · Published 2023-01-26 · Accessed 2026-06-20
- SRC-02Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence ProfileNational Institute of Standards and Technology · Published 2024-07-26 · Accessed 2026-06-20
- SRC-03Privacy FrameworkNational Institute of Standards and Technology · Accessed 2026-06-20
- SRC-04Artificial IntelligenceCybersecurity and Infrastructure Security Agency · Accessed 2026-06-20
- SRC-05Privacy and SecurityFederal Trade Commission · Accessed 2026-06-20
- SRC-06Fighting back against harmful voice cloningFederal Trade Commission · Published 2024-04-08 · Accessed 2026-06-20
- SRC-07FCC Makes AI-Generated Voices in Robocalls IllegalFederal Communications Commission · Published 2024-02-08 · Accessed 2026-06-20
- SRC-08Artificial intelligence in educationUNESCO · Accessed 2026-06-20
- SRC-09AI PrinciplesOrganisation for Economic Co-operation and Development · Accessed 2026-06-20
- SRC-10Copyright and Artificial IntelligenceU.S. Copyright Office · Accessed 2026-06-20
- SRC-11WCAG 2 OverviewWorld Wide Web Consortium · Accessed 2026-06-20
- SRC-12Avoiding the Discriminatory Use of Artificial IntelligenceU.S. Department of Education Office for Civil Rights · Accessed 2026-06-20
- SRC-13Prompt engineering techniquesMicrosoft Learn · Published 2026-05-13 · Accessed 2026-06-20
- SRC-14Prompt engineeringOpenAI · Accessed 2026-06-20