Mastery checklist: pause, verify, respond

A museum uses a clearly labeled generated narrator with a fictional face to guide visitors through an exhibit. No real person’s likeness is used. Is this a deepfake?

Answer: It is synthetic media, but “deepfake” is not the most useful label. It does not make a real person appear to say or do something they did not. The museum should still make the presentation understandable and accessible, but the impersonation and consent risks are lower than in a simulated real speaker.

The lesson is to treat synthetic media as a broad production category and deepfakes as a narrower impersonation concern. Neither word alone determines whether a use is acceptable.

An automated tool says an audio clip is “91% AI-generated.” Can an employer discipline the supposed sender based on that result?

Answer: Not responsibly. The employer needs to know how the detector was evaluated, whether the recording conditions match its tests, and what error rates apply. It should preserve the original, gather account and device evidence through authorized procedures, interview relevant people, and provide review and appeal.

A detector can prioritize investigation. It should not silently become the investigator, witness, and judge.

A video caller who looks like the CEO requests an emergency payment. The face appears natural. What is the best next step?

Answer: Follow the established payment process: end or pause the request, contact the CEO or responsible approver through a known independent channel, verify supplier details, and require the normal approvals. Do not treat natural appearance as authentication.

The control works against deepfakes, compromised accounts, coercion, mistakes, and ordinary fraud. Robust processes are better than asking every employee to make a forensic judgment under pressure.

An advertisement prominently says that its voice is AI-generated. The voice imitates a real athlete who never agreed to the campaign. Is the label sufficient?

Answer: No. The label informs viewers that the audio is generated, but it does not provide the athlete’s consent or make the apparent endorsement fair. The organization should not use the identity without an appropriate legal and ethical basis.

Consent and disclosure serve different people and purposes. Responsible use often requires both.

A viral clip appears to show a mayor announcing a curfew. You cannot determine whether it is generated. What can you do?

Answer: Check the city’s known official website, alert service, public-safety department, and published contact information. Seek independent accountable reporting. Do not use links or phone numbers supplied only by the suspicious post. If no confirmation exists, avoid sharing it as fact and report the rumor to the responsible authority.

You can make a safe decision without proving the clip’s production history.

A public figure says a damaging recording “must be a deepfake,” but offers no evidence. Should the recording be ignored?

Answer: No. Possibility is not proof. Preserve and examine the original where authorized, investigate provenance and context, seek corroboration, and follow fair evidence procedures. The recording should not be accepted uncritically, but neither should it be dismissed automatically.

Healthy skepticism applies to both the recording and the denial.

Before believing, sharing, or acting on consequential media:

• I can state the exact claim separately from the caption’s emotional framing.
• I know what action the item encourages and what could happen if it is wrong.
• I checked the earliest or fullest available source.
• I used a trusted channel independent of contact details inside the item.
• I looked for corroboration that does not merely repeat the same post.
• I treated artifacts and detector scores as clues, not verdicts.
• I checked date, location, editing context, and whether an older item was relabeled.
• I used words such as confirmed, contradicted, misleadingly edited, or unverified accurately.
• I avoided forwarding harmful material unnecessarily.
• I escalated urgent or high-impact cases to the responsible organization.

If several boxes remain unchecked, “do not act or amplify yet” is a valid result.

Before generating or publishing realistic media:

Identity and consent

• A real likeness or voice is necessary for the purpose.
• Consent is informed, specific, voluntary, documented, and limited in duration and use.
• New purposes require new review rather than presumed permission.
• Children, employees, and other people affected by power differences receive stronger safeguards.

Audience and meaning

• The content does not falsely claim a real event, endorsement, or customer experience.
• Disclosure is prominent, plain, accessible, and timed before misunderstanding.
• The notice says what was simulated when a broad “AI used” label would be unclear.
• Likely clips, screenshots, and reposts have been considered.

Operations and remedy

• Source recordings, likeness assets, and generated outputs have restricted access and deletion schedules.
• High-consequence publication receives independent review.
• Affected people can report, appeal, correct, or request removal through a visible route.
• Incident owners can pause distribution and preserve necessary evidence safely.
• Measurements cover response time, consistency, false positives, user understanding, and outcomes for targeted people.

Reject “I can always tell by the eyes.” Artifacts change and authentic media can look strange after compression. Reject “a detector proved it.” A score depends on test conditions and must be combined with other evidence. Reject “it has a label, so consent is solved.” Audience notice does not grant identity rights. Reject “everything can be fake, so nothing can be known.” Documents, trusted channels, provenance, witnesses, and independent records still build confidence.

Also reject public detective work that exposes a targeted person to more harm. Verification should reduce risk, not turn uncertainty into harassment.

Synthetic-media literacy leads directly to content safety. A product needs rules for what it permits, labels, limits, reviews, and removes; reporting routes for people at risk; and tests that cover realistic abuse without teaching harmful creation. Safety is an operational system, not a single classifier.

It also leads to AI ethics. Consent, dignity, fairness, autonomy, transparency, and accountability shape whether a technically possible use should exist. A platform can enforce its written rule and still overlook unequal burdens on people with fewer resources to defend their identity. Ethical analysis asks who benefits, who bears risk, who can refuse, and who can obtain remedy.

Carry one principle forward: trust should rest on accountable processes, not merely realistic appearance. That principle helps whether you are evaluating a video, designing a product, writing policy, or supporting someone targeted by impersonation.