Table of contents
ARTIFICIAL INTELLIGENCE & SOCIETY
PART 4
The series so far
Part 1 asked what happens when the system trusts the wrong account. Part 2 asked what happens when it trusts the wrong person. Part 3 asked what happens when the real person becomes the anomaly inside automated recovery.
This part completes the opening arc. Society can also trust the wrong reality.
Direct Answer
Deepfakes make fake evidence look real and real evidence look fake, so appearance alone is no longer proof. Judge what you see by origin, custody, corroboration, consent, and harm, not by visual realism. Content Credentials and provenance help when present. Detectors alone do not settle trust. Preparation beats panic: AI tools will keep improving, and public evidence standards have to keep up.
One thing up front. The topic of deepfakes invites panic, and I want to stay out of it. I use AI heavily, most days, for real work. I am not writing this to argue that AI is bad or that we should stop. We cannot put the genie back in the bottle. Progress does not pause because video got harder to trust. AI is a tool. Preparation is the point.
Convincing faces and voices already exist. That part is true enough to matter.
The real subject is public evidence.
For a long time, a recording carried a quiet assumption: if you could see it and hear it, something like it happened. Editing existed. Propaganda existed. But fabricating a full scene that looked like a camera caught it was hard enough that visual realism still did a lot of work.
Generative video and audio change that assumption. When almost anyone can make a person appear to say or do something they never said or did, appearance alone stops being proof. Deepfakes and public evidence are now the same problem. The question is not only whether a clip looks real. The question is whether you can trust what you see.
I already use the tool that breaks the old proof
I have used Sora2 to make harmless promotional clips. Mark Cuban giving a shoutout. A ring-doorbell bit with Trump. Bruce Lee and Muhammad Ali singing. None of those were meant to deceive a court, move a market, or ruin a life. They were toys and experiments. They worked well enough that the old gut check, “that cannot be fake,” no longer holds for me.
That is the useful part of the experience. I did not need a crime story to learn the capability. I learned it by making things that were obviously playful once you knew the setup, and still startling if you only watched the frame.
Now carry the same capability into places where the audience does not know the setup.
Someone can fabricate a politician taking a bribe. Someone can manufacture compromising footage to extort a public figure. Someone can stage a fake kidnapping video for ransom. Those are possibilities, not reports of specific cases I personally investigated. The point is the hinge: the same class of tool that makes a funny cameo can also make a file that looks like evidence.
When video is no longer proof by itself, the trust question moves. Who made this? Where did it come from? What happened to the file on the way to my screen? What else confirms it?
What recorded reality used to do for us
A video or audio clip used to settle arguments faster than text. It shortened disputes in newsrooms, living rooms, and sometimes courtrooms. It was never perfect. Selective editing always existed. Context always mattered. Still, a recording raised the cost of a bald denial. You could argue about meaning. It was harder to argue that the event never happened at all.
Institutions built habits around that. Share the clip. React. Demand a response. Treat the image as the center of the story.
Those habits still feel natural. They are now incomplete.
A recording can still matter. It just cannot finish the argument by looking real.
Both harms: fake evidence believed, real evidence dismissed
There are two failure modes, and they travel together.
First, fake evidence gets believed. In May 2023, a fabricated image of an explosion near the Pentagon spread on social media, amplified by accounts that carried paid verification badges, including one impersonating Bloomberg News. Arlington police and the Pentagon said there was no incident. Before that clarification landed, the S&P 500 briefly dipped about 0.3 percent as markets reacted to the image. [1][2] Hold the market move for a second. Look at what the situation trusted: a single realistic picture, plus the appearance of verified news authority, arriving fast enough that money moved before anyone checked the source.
That is not a novelty deepfake. That is a trust failure with a market attached.
The Federal Communications Commission documented another class of harm in the 2024 New Hampshire primary: thousands of prerecorded AI deepfake robocalls using a cloned presidential voice, aimed at potential voters. A carrier later settled FCC enforcement. [3][4] You do not need to claim the calls decided the election to see the mechanism. A familiar voice arrived through a familiar channel and carried a political instruction.
Baltimore County police later concluded that a viral recording attributed to a high-school principal was AI-generated, after forensic review, and charged a school athletic director in connection with the case. [5] The public heard a voice that sounded like authority. The investigation said the authority was fabricated.
And in 2026, the Justice Department announced a guilty plea it described as the first U.S. conviction under the TAKE IT DOWN Act, involving AI-generated sexual videos used as a tool of harassment and coercion. [6][7] Synthetic intimacy is not a side issue. It is one of the clearest proof points that deepfake harm is not abstract.
Second, real evidence gets dismissed as fake. Legal scholars Robert Chesney and Danielle Citron named this the liar’s dividend: deepfakes make it easier for liars to avoid accountability for things that are in fact true. [8] Once the public knows a convincing fake is possible, the dishonest answer to an authentic recording becomes, “That is AI.”
Both harms matter. If you only worry about believing fakes, you miss the other half. If you only worry about the liar’s dividend, you miss the fraud, coercion, and election interference already documented. Public trust fails in both directions.
Consent is not the same thing as crime
Not every synthetic performance is an attack on reality.
In 2026, Associated Press reported that a generative AI version of Val Kilmer would posthumously appear in an independent film, As Deep as the Grave. Producers said Kilmer had signed on before his death but could not film because of his health. AP reported that his estate gave permission and is being compensated. His daughter, Mercedes Kilmer, said he looked at emerging technologies with optimism as a tool for storytelling. The producers said they followed SAG-AFTRA digital-replica consent rules and intended the project as a demonstration of doing the work ethically with an estate and family. [9]
That case is not “deepfakes are fine.” It is the contrast the panic narrative erases.
| Authorized performance | Malicious synthetic use |
|---|---|
| Consent, estate authority, compensation | No consent |
| Completes work the person chose | Fabricates acts the person never performed |
| Disclosure and union framework | Concealment, fraud, coercion, evidence fakery |
| Artistic continuation | Reputation destruction, market or political harm |
The article’s point is not that synthetic media is intrinsically evil. The point is that consent, disclosure, traceable origin, and harmful use decide what the file means. Treating every AI likeness as the same category is how you lose the plot.
Why deepfake detection alone cannot carry public trust
It is tempting to answer the whole problem with a detector score. Upload the file. Get a percentage. Move on.
That is not a durable trust model.
NIST’s work on synthetic content describes detection as a constant cat-and-mouse game. Detectors are often tied to specific generators and may not generalize to the next one. Post-processing, compression, and ordinary circulation can change performance. [10] NIST’s forensics challenge materials have also described large performance drops when systems move from academic evaluation into operational conditions. [11]
I am not claiming detectors are useless. Forensic teams should keep building them. I am claiming something narrower: a detector score is a signal, not a verdict. Society cannot build its entire public-evidence model on catching fakes after the fact, especially when the next generator is already in training. That is Andrew’s interpretation of the technical record, not a claim that every tool is “wrong more than right.” Clean prevalence stats for that slogan are not what the sources support.
If detection is an arms race, the stronger architecture is to prove what is authentic when you can, and to demand corroboration when you cannot.
Content Credentials: prove origin, do not worship the pin
The Coalition for Content Provenance and Authenticity (C2PA) publishes an open standard for Content Credentials: a cryptographically signed, tamper-evident record of where digital content came from and what happened to it. Think of it as a nutrition label for a file: creation method, editing history, tools used, and, when asserted, AI involvement. [12][13]
That is valuable. It is also easy to oversell.
Content Credentials are not a certificate that the depicted event happened in the world. They are a provenance record for the asset. They are also not DRM. They do not stop someone from copying a frame. C2PA’s own guidance notes that legacy platforms and ordinary distribution can strip or corrupt metadata when a file is re-encoded for social media. Soft bindings such as watermarks or fingerprints can help rediscover credentials after stripping, but recovery is not universal or automatic. [14]
So the practical rule is simple:
- Credentials present and verifiable strengthen the evidence chain.
- Credentials absent do not prove a fake. Adoption is incomplete, and stripping happens.
- A screenshot of a video is often a new file with a broken or empty chain.
Provenance helps the good-faith publisher show work. It does not replace independent confirmation when the stakes are high.
What law should target: harm, not every likeness
Law should follow harm.
The FTC’s rule on impersonation of government and businesses is final and effective. It gives the Commission stronger tools against scammers who impersonate agencies and companies. [15][16] An expansion covering impersonation of individuals has been proposed; it is not final as of the current Part 461 text. Do not cite an individual AI-deepfake ban as settled federal FTC rulemaking. [17]
Congress did enact a major federal statute aimed at a concrete harm class. The TAKE IT DOWN Act (Public Law 119-12) criminalizes specified knowing nonconsensual online publication of intimate imagery, including digital forgeries, and requires covered platforms to run notice-and-removal processes. [7] DOJ has already brought cases under it. [6] The statute does not contain an express parody or satire carveout in the text reviewed for this article; constitutional questions about protected speech remain a separate analysis. Write carefully: the Act targets nonconsensual intimate imagery and digital forgeries in that lane, not every joke video on the internet.
Older tools still matter too: wire fraud, identity fraud, extortion, election-law enforcement, and state deepfake statutes where they exist. The useful legislative instinct is the same one this series keeps returning to. Regulate the harm and the deception architecture. Do not pretend you can outlaw the underlying generative capability.
Courts and platforms need matching habits. A standalone video file should not remain sufficient proof in high-stakes cases. Original source, device records, metadata, chain of custody, witnesses, corroboration, and cryptographic provenance where available should matter more than the emotional force of the image. That is Andrew’s prediction for institutions that want to keep public trust: appearance will keep getting cheaper to fake; corroboration will not.
What you can still do when seeing is not enough
You do not need a lab to practice better judgment.
Slow down before you react. Inspect the original source, not the twentieth reshare. Seek independent confirmation from outlets or records that did not start with the same file. Separate how sure you feel from what you can show. Do not share a clip merely because it confirms what you already believe. Accept that visual realism is not proof. Do not treat an AI detector as a final judge. Preserve uncertainty when the evidence is incomplete.
None of that is a personality lecture. It is public infrastructure at human scale. Shared reality depends on people who refuse to let a vivid file finish their thinking for them.
Institutions have larger versions of the same duties. Preserve provenance instead of stripping it. Label synthetic media when they know it is synthetic. Build human appeal when authentic content is mislabeled. Slow viral distribution during elections, wars, emergencies, and crime events long enough for verification. Keep technical records when content is removed. Correct fast when they get it wrong.
Discernment is now a civilizational skill
Parts 1 through 3 stayed close to identity systems: accounts, verification packages, recovery mazes. Part 4 widens the frame to the street outside those systems. If society cannot tell an authorized performance from a fabricated crime scene, or cannot protect real evidence from the liar’s dividend, then every institution that depends on recorded reality inherits the failure.
I am not asking anyone to stop using generative tools. I use them. The harmless Sora2 clips are part of why I take the malicious possibilities seriously. Capability is not the villain. Unexamined trust is.
When seeing is no longer believing, the work is not nostalgia for a world where cameras settled everything. The work is building habits, standards, and laws that judge media by origin, custody, corroboration, consent, and harm.
That is how you keep a shared world when anyone can paint a moving picture of a false one.
Frequently Asked Questions
Can you still trust what you see online when deepfakes exist?
You can trust a clip more when you can check origin, provenance, and independent corroboration. Visual realism alone is no longer enough. Deepfakes do not make every video false. They make appearance an incomplete form of proof. [8][12]
What is the liar’s dividend?
It is the advantage dishonest people gain once the public knows deepfakes exist: they can dismiss authentic recordings as AI-generated. The term comes from Chesney and Citron’s 2019 California Law Review article. [8]
What are Content Credentials?
Content Credentials are C2PA’s tamper-evident provenance records for digital files. They can show how content was created and edited. They do not by themselves prove that a depicted real-world event happened, and the credentials can be stripped when files are re-shared. [12][13][14]
Do deepfake detectors solve the problem?
No. Detectors can help investigators, but NIST describes detection as an adversarial arms race with limited generalization and fragile performance after ordinary post-processing. A score is a signal, not a public trust system. [10][11]
Is every AI performance of a person illegal or unethical?
No. Authorized projects with consent, disclosure, and compensation, such as the estate-approved Val Kilmer performance reported for As Deep as the Grave, are a different category from concealed fraud, coercion, or fabricated evidence. Harm, consent, and disclosure decide the ethics and much of the law. [7][9]
References
[1] AP News. “FACT FOCUS: Fake image of Pentagon explosion briefly sends jitters through stock market” (May 22, 2023). https://apnews.com/article/pentagon-explosion-misinformation-stock-market-ai-96f534c790872fde67012ee81b5ed6a4
[2] Bloomberg. “Fake AI Photo of Pentagon Blast Goes Viral, Trips Stocks Briefly” (May 22, 2023). https://www.bloomberg.com/news/articles/2023-05-22/fake-ai-photo-of-pentagon-blast-goes-viral-trips-stocks-briefly
[3] Federal Communications Commission. Consent decree regarding Lingo Telecom and AI robocalls (DA-24-790). https://docs.fcc.gov/public/attachments/DA-24-790A1.txt
[4] Federal Communications Commission. Settlement release on AI deepfake robocalls (DOC-404951). https://docs.fcc.gov/public/attachments/DOC-404951A1.pdf
[5] Baltimore County Police. “Athletic director charged in Pikesville High School AI case.” https://www.baltimorecountymd.gov/departments/police/news/athletic-director-charged-pikesville-high-school-ai-case
[6] U.S. Attorney’s Office, Southern District of Ohio. “Columbus man pleads guilty to cyberstalking exes, creating AI-generated obscene material” (DOJ statement on TAKE IT DOWN Act plea). https://www.justice.gov/usao-sdoh/pr/columbus-man-pleads-guilty-cyberstalking-exes-creating-ai-generated-obscene-material
[7] TAKE IT DOWN Act, Public Law 119-12. https://www.congress.gov/119/plaws/publ12/PLAW-119publ12.htm
[8] Robert Chesney & Danielle Keats Citron. “Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security,” 107 California Law Review 1753 (2019). https://www.californialawreview.org/print/deep-fakes-a-looming-challenge-for-privacy-democracy-and-national-security
[9] Jake Coyle / Associated Press. “A generative AI version of Val Kilmer will co-star in an independent film” (March 18, 2026). https://apnews.com/article/val-kilmer-ai-movie-5e32b8e3ee65a01b75902bf4d0bf0b98
[10] NIST. Reducing Risks Posed by Synthetic Content (NIST AI 100-4, initial public draft). https://airc.nist.gov/docs/NIST.AI.100-4.SyntheticContent.ipd.pdf
[11] NIST. GenAI / Deepfakes forensics challenge materials. https://ai-challenges.nist.gov/forensics
[12] Coalition for Content Provenance and Authenticity (C2PA). https://c2pa.org/
[13] Content Credentials. https://contentcredentials.org/
[14] C2PA. Implementation guidance (metadata stripping / soft bindings). https://spec.c2pa.org/specifications/specifications/2.4/guidance/Guidance.html
[15] Federal Trade Commission. “FTC Announces Impersonation Rule Goes into Effect Today” (April 1, 2024). https://www.ftc.gov/news-events/news/press-releases/2024/04/ftc-announces-impersonation-rule-goes-effect-today
[16] 16 CFR Part 461 (Impersonation of Government and Businesses). https://www.ecfr.gov/current/title-16/chapter-I/subchapter-D/part-461
[17] Federal Trade Commission. “FTC Proposes New Protections to Combat AI Impersonation of Individuals” (Feb 15, 2024). https://www.ftc.gov/news-events/news/press-releases/2024/02/ftc-proposes-new-protections-combat-ai-impersonation-individuals