The Crisis of Knowing

For most of human history, seeing carried unusual authority. A photograph, a recording, a signed document, or a voice on the telephone could end an argument because the evidence itself was costly to fabricate.

Generative AI changes that bargain. A convincing image can be created without a camera. A familiar voice can speak words the person never said. A video meeting can contain people who were never present. A legal filing can cite witnesses who never existed.

The crisis is larger than false content. Once fabrication becomes cheap enough, every genuine artifact acquires a new burden: prove that you are real.

The verification premium has left the workplace. It now applies to reality.

Evidence Used to Arrive with Friction

Traditional evidence was never infallible. Photographs could be staged, witnesses could lie, documents could be forged, and recordings could be edited. The important difference was friction. High-quality fabrication usually required access, equipment, skill, time, or coordination.

Generative systems compress those costs. They let an ordinary user produce plausible text, imagery, audio, and video at a speed once reserved for professional studios or organized fraud. The result is synthetic abundance.

Synthetic abundance creates authenticity scarcity. There can be plenty of convincing files and very little dependable information about which ones deserve belief. Making another image or recording becomes easy. Establishing where it came from and what it proves can still take work.

That shift mirrors Chapter Eighteen. When production becomes cheap, verification becomes relatively expensive. In public life, the cost is paid by journalists, courts, companies, schools, families, platforms, investigators, and ordinary people who now have to ask whether the evidence in front of them came from the world it claims to depict.

Seeing No Longer Closes the Question

A 2024 fraud case in Hong Kong shows how quickly the old visual shortcut can fail. Police reported that an employee received a phishing email from someone impersonating the chief financial officer of the employee’s company and joined what appeared to be a group video conference. The employee ultimately authorized transfers totaling about HK$200 million.

Investigators said the conference had been fabricated from publicly available video clips and voices of the impersonated executive. The meeting was prerecorded rather than genuinely interactive. The fraud worked because the familiar verification cues were present at the same time: face, voice, title, meeting context, and apparent colleagues.

The practical lesson is uncomfortable. Verification must go beyond sensory plausibility. A high-stakes instruction arriving through video, audio, or text now needs an independent channel that the artifact itself does not control.

The cheapest fake can impose an expensive investigation.

When a Conversation Imitates Credibility

A forged image is only one way to create false confidence. A conversation can do the same work, one plausible answer at a time. A professional described on LinkedIn an approach from a supposed recruiter about a senior position. After receiving her résumé, the contact returned a detailed assessment of her fit and named a company. The exchange seemed to understand her experience. Requests for an executive profile, leadership philosophy, and 30-60-90-day plan made her suspicious. She asked whether the contact was a person or an AI system.

This is her account, not a verified investigation. The available screenshot does not establish who operated the account, whether AI was used, or what the contact intended. It illustrates a vulnerability: personalized, polished replies can make a claim of identity feel more credible without independently proving it.

The Federal Trade Commission has warned that scammers impersonate well-known companies and recruit for fake jobs on LinkedIn and other job platforms. Its practical advice is to contact the company through a number or website you know is legitimate, rather than one supplied by the supposed recruiter.

AI changes fraud from mass deception to personalized deception. A system can read a résumé, mirror its language, sustain a long exchange, and answer detailed questions. Spelling, tone, and apparent familiarity become weaker checks when a deceptive actor can produce them cheaply. The same problem applies to a vendor seeking access, a caller requesting payment, or a message purporting to speak for an institution.

When AI can imitate credibility, credibility is no longer verification.

The safer question is not whether the exchange sounds human. It is whether the identity and authority behind a consequential request have been checked through a separate trusted channel. Before sending sensitive information, money, credentials, or work that could be misused, find the organization’s contact details independently and confirm both the person and the request. Legitimate recruiting can involve AI assistance; that alone does not make it fraudulent. The boundary is independent verification before the relationship gains authority.

This extends Chapter Eighteen’s verification premium. The cost of making a persuasive approach falls, while the recipient must spend effort checking it. Principle 3 calls for a boundary before information or authority moves. Principles 4 and 5 require the organization and platform to make meaningful checking possible, rather than leaving an individual to judge authenticity from polish alone. Delegation requires verification.

The Cost Is Asymmetric

Synthetic media creates a cost asymmetry that matters even when the fake fails. The creator may spend minutes producing an audio clip, document, image, or video. The recipient may need hours of calls, source checks, metadata review, interviews, legal analysis, or forensic work to establish what happened.

That asymmetry turns the verification premium into a public tax. A fraudster can impose it on a company. A fabricated image can impose it on a newsroom. A false document can impose it on a court. A synthetic accusation can force a real person to spend time and money proving an event never occurred.

The attacker therefore does not need universal belief. Creating enough uncertainty to delay action, trigger an investigation, damage a reputation, or consume institutional attention can be valuable by itself. The same economics explains why verification has to be designed before a crisis rather than assembled from scratch after the artifact goes viral. Verification policy is cheaper before the emergency. High-consequence systems should decide in advance which claims require a second channel, an original file, a signed source, independent corroboration, or human confirmation.

Detection Is a Weak Foundation

One response to synthetic media is to become better at spotting it. That helps. It is a poor foundation for an information system.

A 2026 study in Computers in Human Behavior Reports tested 38 participants across real faces and three forms of synthetic or manipulated faces. Participants identified Stable-Diffusion-generated faces and real faces with relatively high accuracy, but StyleGAN-generated faces were usually misclassified as real and face-swapped images were also difficult to detect. More viewing time did not solve every category.

The result is less a scorecard of human gullibility than a warning about the task. Detectability depends on the generation method, the manipulation, the viewing conditions, and the detector. A person can learn yesterday’s artifact and still miss tomorrow’s model.

Automated detectors have the same structural problem. NIST’s synthetic-content work treats detection as one layer among several, alongside provenance, watermarking, authentication, testing, and auditing. A detector can contribute evidence, but it is too brittle to serve as society’s single machine for deciding what is real.

A workable trust system has to function for ordinary users as well as forensic examiners.

When a Score Becomes a Verdict

Nicole Ramirez reported in an October 2026 LinkedIn post that an AI detector gave her 2018 TEDx talk an “87% AI” result and flagged other writing produced before 2019. The account does not identify the tool or provide enough information to repeat the test. It is an anecdote about a possible false positive, rather than a measurement of detector reliability.

There is stronger evidence that this risk deserves attention. Turnitin acknowledges that its AI writing detector can misidentify human writing and says its report should not be the sole basis for penalizing a student. A 2023 study of seven detectors found frequent misclassification of English essays by non-native writers. That older study documents a risk; it does not measure the performance of every detector used in 2026.

On September 25, 2026, the Académie Goncourt removed Thélyson Orélien’s C’était ça ou mourir from its longlist after allegations of AI use. Reuters reported that Orélien denied using AI and that the academy also considered separate plagiarism allegations concerning earlier work. The removal is established; the detector result alone does not establish AI authorship. Those distinctions matter before making a judgment about the writer.

Creative authorship involves choosing questions, shaping an argument, checking claims, revising the work, and accepting responsibility for it. Publishers and prize organizers can set rules about AI assistance, but enforcing those rules requires evidence about the actual writing process. Drafts, revision histories, and the author’s explanation can help establish that process. A polished style cannot establish it on its own.

The institution still owns the judgment that follows. It must decide what its evidence supports, give the person an opportunity to respond, and remain answerable for any penalty. A detector may justify an investigation. It cannot relieve a human decision-maker of the duty to examine the evidence.

The Liar’s Dividend

Synthetic media creates a second problem that is more corrosive than a successful fake. Real evidence can now be attacked by claiming that it is synthetic.

Legal scholars call this the liar’s dividend. In a 2026 experiment published in Psychology of Popular Media, 182 participants viewed genuine political video and then saw either an apology or a false claim that the video was a deepfake. The false deepfake claim increased the tendency to misidentify the genuine recording as manipulated and improved perceptions of the politician’s leadership ability relative to the apology condition.

The denial produced no direct decline in general media trust, while participants who were more likely to misidentify the real video as fake reported lower media trust. The study concerns a controlled political scenario rather than every kind of evidence. Its central result is still important: awareness of fabrication can become a tool for discrediting authentic material.

The liar’s dividend turns every fake into an alibi for the real.

Teaching people to distrust more can deepen the crisis of knowing. Generalized suspicion can reward the very actor who benefits when authentic evidence becomes contestable.

Provenance Is Not Truth

Chapter Fifteen introduced the creative chain as a rights chain. The same infrastructure now matters for evidence because origin and history can sometimes be verified even when appearance cannot.

C2PA is a shared standard for attaching a digital history record, called Content Credentials, to a file. Compatible tools can record how the file was created or edited. A digital signature helps other compatible tools check who signed the record and detect changes to the signed information. The standard reached version 2.4 in 2026, and adoption has expanded across technology companies, platforms, camera makers, and media organizations.

The distinction C2PA itself makes is essential: provenance is not truth. A valid credential can help establish that a file came from a particular tool or passed through a recorded sequence of edits. Questions about whether the depicted event occurred, whether the caption is accurate, and whether the creator is honest remain outside the credential.

A synthetic image can have impeccable provenance and accurately disclose that it is synthetic. A real image can lose its credential when it is cropped, screenshotted, or passed through software that does not preserve the chain.

A label can tell you how a file was made. It cannot tell you whether the world it depicts existed.

Authentic Media Can Still Mislead

The synthetic-media problem can distract from an older fact: authenticity and truth are different questions. A genuine photograph can be paired with a false date. Real video from one country can be reposted as evidence of an event in another. A complete recording can be clipped so that the omitted minute changes the meaning. A legitimate document can be described as proving something it never says.

A file’s recorded history cannot settle every claim made about it. It may show which camera captured an image and which software edited it, while leaving a false caption unchallenged. Chain of custody means the record of who held or handled the file and what happened to it along the way. That history helps, but someone must still check whether the file supports the story attached to it.

This distinction matters because the easiest misinformation may eventually be a real file placed in a false frame. Synthetic content raises the cost of verification, but it also makes audiences so focused on pixel-level fakery that they can overlook ordinary miscaptioning, selective editing, and misleading context.

The strongest information systems therefore separate three questions: Is this the file it claims to be? What happened to the file on the way here? And does the file actually support the claim attached to it? Provenance can strengthen the first two. Corroboration and accountable interpretation still have to answer the third.

Trust Moves from Appearance to Chain of Custody

The answer is layered verification: several checks that support one another. Chapter Fifteen asked where creative material came from and how it was made. Here those questions help us decide whether to trust a file and the claim attached to it.

For consequential content, origin matters. Who captured, generated, signed, or first published it? History matters. What happened to the file before it reached you? Context matters. What claim is being made about what the artifact proves? Corroboration matters. Is there independent evidence that reaches the same conclusion through another path?

This resembles the way mature institutions already handle valuable evidence. Courts use authentication rules and chains of custody. Newsrooms seek multiple sources. Financial systems use dual controls. Scientific research records methods, data, and provenance. Those systems exist because trust is stronger when it rests on process rather than impression.

The shift ahead is cultural as much as technical. Trust will move from appearance toward provenance, corroboration, and accountable publication.

Chain of custody beats visual confidence.

Verification Needs Institutional Memory

Provenance works best when organizations preserve it. A newsroom that keeps original files, source records, edit histories, and publication logs is easier to defend than one trying to reconstruct them after a challenge. A company that requires payment instructions to be confirmed through a known channel has evidence outside the suspicious message. A research group that keeps signed data and version histories can distinguish an altered derivative from the record it actually produced.

This is the institutional version of the verification premium. An organization spends time keeping records so it can answer questions later. Digital credentials can help reveal changes to signed information. Ordinary habits matter too: preserving originals, recording sources and edits, confirming important instructions independently, and identifying which version is authoritative when copies disagree.

Verification is easier when evidence arrives with history than when history has to be rebuilt after trust has already collapsed.

When Fabrication Enters an Institution

The cost of synthetic evidence rises sharply when an institution gives it authority. That is where a plausible artifact can acquire consequences far beyond the file itself.

Chapter Twelve described the authority jump: the point where machine output stops being information and becomes state action. The New Mexico Supreme Court supplied a stark institutional version in September 2026. It fined and held a defense lawyer in contempt after an appellate brief contained false police testimony and wholly fabricated witnesses produced with ChatGPT. The lawyer told Reuters he had used the system to summarize trial materials and did not understand how far it could invent facts.

The significant failure was verification. Fabricated material crossed from a generative system into a signed court filing, where it acquired the appearance of professional and legal authority. The human signature did not make the underlying evidence real.

The same pattern can occur in journalism, medicine, finance, research, or corporate decision-making. Synthetic content becomes consequential when a trusted institution imports it without independently checking the path by which it arrived.

Verification is the toll charged at the authority jump.

Provenance Is Becoming Product Architecture

The technology industry is beginning to move verification upstream. In 2026 C2PA published new implementation guidance specifically for distinguishing synthetic, AI-modified, and non-synthetic content using tamper-evident Content Credentials.

OpenAI now says supported images and audio generated with its tools can carry provenance signals through C2PA credentials and SynthID watermarking, with a public verification tool and API for checking supported files. OpenAI also states the limitation clearly: a detected signal can indicate origin in its tools, but it does not prove that the content is accurate, unedited, legally owned, or presented in the correct context. OpenAI systems are used in this Living Edition’s production workflow, so that description is a provider claim with the standing disclosure attached.

This is Principle 5 operating at the content layer. If verification depends entirely on the person at the end of the chain noticing a suspicious pixel, the architecture has already failed. Provenance works better when the generating system, editing tools, platform, publisher, and viewer can carry compatible signals forward, beginning the verification system before anyone has reason to suspect a fake.

Transparency Is Becoming Law

Regulation is moving in the same direction by assigning transparency duties before synthetic content reaches the final viewer.

The European Union’s AI Act transparency duties under Article 50 began applying on August 2, 2026. Providers of generative systems must make synthetic outputs machine-readable and detectable as artificially generated or manipulated where technically feasible. Deployers must disclose deepfakes and, in specified circumstances, AI-generated or manipulated text published to inform the public on matters of public interest. The European Commission also finalized a voluntary Code of Practice and implementation guidelines to support compliance.

These rules leave the truth problem intact while changing who must carry part of the verification burden. The provider has obligations at generation, and the professional deployer has obligations at publication.

The United States has taken a narrower approach in one high-harm category. The TAKE IT DOWN Act covers nonconsensual intimate imagery, including AI-created digital forgeries. Since May 19, 2026, covered platforms have been required to provide a removal process and take down covered images and known identical copies within 48 hours of a valid request, with FTC enforcement.

These regimes regulate different harms, but they share an architectural idea: the verification burden should move upstream instead of leaving the victim or viewer to prove every fabrication after it spreads. That is a more realistic allocation of the verification burden.

Current Evidence: Verification Is Becoming Infrastructure

Reviewed September 12, 2026. Text provenance evidence added October 6, 2026. Text-detector case evidence added October 10, 2026. Connected election-abuse testing evidence added October 15, 2026.

NIST’s current synthetic-content framework treats no single technical method as sufficient. Authentication, provenance, watermarking, detection, testing, and auditing are complementary layers. That is the right architecture for a world in which both generation and detection continue to improve.

C2PA’s 2026 specification and implementation guidance show provenance moving from an industry aspiration toward shared infrastructure. The standard’s own caution is equally important: credentials authenticate provenance claims and tamper evidence, not factual truth. Missing provenance can also result from ordinary editing or unsupported tools.

The European Union has now turned part of that architecture into a legal obligation. Article 50 transparency requirements began applying in August, backed by Commission guidelines and a voluntary Code of Practice that had attracted roughly 190 company and organizational signatories by the end of July.

U.S. enforcement is narrower but already operational. The FTC began enforcing the TAKE IT DOWN Act’s platform duties in May 2026 and sent warning letters to 12 websites offering so-called nudify tools, reminding them that covered platforms must provide a removal process for nonconsensual intimate imagery, including digital forgeries. A warning letter is not a finding of liability, but it shows the verification problem becoming an operational duty: platforms must be able to receive a claim, identify covered material, act on known identical copies, and document a response within a fixed period. The law therefore converts one category of synthetic-media harm from a user’s private evidentiary struggle into a platform process.

OpenAI’s July expansion of its provenance program added supported audio, SynthID watermarking, and API access to its public verification system. The company explicitly warns that origin signals do not establish accuracy or context. That limitation is the chapter’s central distinction written into a provider’s own verification product.

On October 5, 2026, OpenAI announced optional text watermarking for selected API models and plans to add it to eligible ChatGPT and Codex output in the European Union over the following weeks. Detector access initially goes to approved researchers and expert organizations. The company warns that detection can produce false positives and miss edited or short passages. A detected watermark does not establish human contribution, ownership, responsibility, or truth; an absent watermark does not prove human authorship.

The accountability boundary belongs around the decision that follows. A school, employer, or publisher should not turn one detector result into an accusation. Seek corroborating evidence, allow the person to respond, and identify who is responsible for the judgment. A signal can assist review without replacing it.

Human detection remains uneven. The 2026 face study found that some synthetic images were easy for participants to identify while others were overwhelmingly mistaken for real. Longer scrutiny did not create a universal solution. Training people to look harder is therefore useful but structurally insufficient.

The liar’s-dividend experiment demonstrates the opposite failure. Genuine video became easier to dismiss after a false deepfake claim. Synthetic media can therefore create error in both directions: people can accept a fake, and people can reject the real.

Institutional failures make the verification premium concrete. The New Mexico court case shows fabricated AI output entering a signed legal filing. Hong Kong’s deepfake fraud case shows visual and auditory plausibility defeating ordinary corporate verification. In both, the expensive part arrived after the cheap synthetic artifact.

The recruiter anecdote adds a conversational form of the same failure mode. It is not proof of an AI-run scam, but it shows why a coherent, personalized exchange cannot authenticate the speaker. FTC warnings about impersonated employers support the independent-channel rule.

On October 14, 2026, Anthropic reported that outside testers could get Claude to assist a deceptive election campaign by dividing it into smaller requests that appeared harmless separately. The company said its safeguards rejected direct harmful requests but could miss the combined operation, and that it had added detection measures. This is the provider’s account of testing and mitigation, not independent proof that the weakness has been removed.

The boundary must cover the purpose of the whole assignment. Researching an audience, drafting a page and planning distribution can be ordinary tasks. Combined, they can support deception. Providers and organizations using these systems should test connected sequences, decide when activity must be escalated, and identify who can stop the operation.

The evidence points toward a new baseline: important claims need provenance and corroboration before they receive authority.

Accountability Analysis

Principles 3, 4, 5, and 6 carry most of this chapter. Principle 3 requires written boundaries around what evidence can trigger consequential action and what independent verification is required before money, liberty, reputation, or institutional authority moves.

Principle 4 makes review real. A person who merely looks at convincing media is not reviewing it. Meaningful review requires access to provenance, source context, corroborating evidence, and enough time and authority to reject the artifact even when it appears authentic.

Principle 5 places accountability upstream. Generators, editing tools, platforms, publishers, and institutions can preserve provenance, attach labels, maintain logs, design confirmation channels, and prevent unverified synthetic material from acquiring institutional authority.

Principle 6 asks who bears the verification cost. The person targeted by a deepfake should not carry the entire burden of proving a negative. The company receiving a high-value payment instruction should own a verification process. The publisher amplifying an extraordinary claim should bear some cost of checking it. The institution that grants authority to evidence must own the conditions for accepting it.

The crisis of knowing becomes manageable when verification is treated as infrastructure rather than personal skepticism.

Check the file and the claim

Choose an image or recording attached to an important claim. Look for the original source, its date, and the context in which it was made. Then seek independent evidence for the claim itself. A second account repeating the same material may not be an independent check. If the evidence is incomplete, leave the conclusion open instead of declaring the material real or fake from appearance alone.

Accountability Verdict

A blanket policy of believing nothing hands the liar’s dividend to anyone who benefits from confusion. The answer is disciplined verification rather than generalized cynicism.

Respond by raising the evidentiary standard in proportion to consequence. Low-stakes content can remain easy. High-stakes claims should require an independent channel, provenance when available, source accountability, and corroboration that does not depend on the same artifact.

The practical rule is simple: appearance is evidence, not proof.

When anything can be fabricated, reality needs receipts.

Chapter Twenty turns from verification to rulemaking. Governments now have to regulate systems whose capabilities, products, and failure modes can change faster than the law that governs them.

Sources and Notes

1. Hong Kong Government, “LCQ9: Combating frauds involving deepfake,” June 26, 2024. https://www.info.gov.hk/gia/general/202406/26/P2024062600192.htm. Accessed September 12, 2026. Hong Kong Police reported a January 2024 fraud involving a fabricated prerecorded video conference in which an employee authorized transfers totaling about HK$200 million after fraudsters impersonated a chief financial officer using public video and voice material. Police said the conference itself was not genuinely interactive. The chapter uses the case to show that face, voice, title, and meeting context can fail together as authentication cues. Source record: DA-CH19-S1. Archive reference: pending.

2. Zihao Zhao, Shanghong Li, Mingzhi Lyu, Ian McLoughlin, Adams Wai Kin Kong, and Hong Xu, “We are good at detecting Stable-Diffusion-generated deepfakes but not StyleGAN-generated or face-swapped deepfakes,” Computers in Human Behavior Reports 22 (May 2026): 101035. https://doi.org/10.1016/j.chbr.2026.101035. Accessed September 12, 2026. The study included 38 participants overall; the eye-movement analysis used 37 participants because one recording was unusable. Participants judged real, Stable-Diffusion-generated, StyleGAN-generated, and face-swapped faces across viewing durations from 17 to 1,000 milliseconds. Detection varied sharply by image type: real and Stable-Diffusion faces were identified with relatively high accuracy, while StyleGAN and face-swapped images were often misclassified as real. The study is small and concerns facial images rather than all synthetic media, so it supports the claim that unaided human detection is uneven rather than a universal estimate of deepfake detectability. Source record: DA-CH19-S2. Archive reference: pending.

3. Lara Grohmann, Franziska A. Halle, and Markus Appel, “Deepfake! A Liar’s Dividend for Audiovisual Material,” Psychology of Popular Media, advance publication, 2026, DOI 10.1037/ppm0000665. https://doi.org/10.1037/ppm0000665. Accessed September 12, 2026. In a preregistered experiment with a final sample of 182 participants, genuine political video followed by a false deepfake denial produced greater misidentification of the authentic video as manipulated and higher perceived leadership ability than an apology condition. The denial did not directly reduce general media trust, although deepfake misidentification was negatively associated with media trust. The chapter uses this study as evidence that synthetic-media awareness can undermine authentic evidence, not as a general claim about all political communication. Source record: DA-CH19-S3. Archive reference: pending.

4. Bilva Chandra, Jesse Dunietz, Kathleen Roberts, Yooyoung Lee, Peter Fontana, and George Awad, “Reducing Risks Posed by Synthetic Content: An Overview of Technical Approaches to Digital Content Transparency,” NIST AI 100-4, November 20, 2024, updated April 8, 2026. https://doi.org/10.6028/NIST.AI.100-4. Accessed September 12, 2026. NIST surveys authentication and provenance, watermarking, synthetic-content detection, testing, and auditing as complementary approaches and identifies limitations and research needs across them. The chapter uses the report for the layered-verification argument rather than to endorse one technology. Source record: DA-CH19-S4. Archive reference: pending.

5. Coalition for Content Provenance and Authenticity, “C2PA Specifications 2.4,” 2026; and “A New Implementation Guide for Content Credentials,” July 31, 2026. https://spec.c2pa.org/ and https://c2pa.org/a-new-implementation-guide-for-content-credentials/. Accessed September 12, 2026. C2PA defines an open provenance standard using tamper-evident, cryptographically signed manifests and Content Credentials. Its explainer expressly distinguishes provenance from factual truth: verified provenance can establish aspects of origin and history but cannot by itself establish that depicted events or accompanying claims are true. The standard also recognizes incomplete provenance when content passes through unsupported tools or workflows. Source record: DA-CH19-S5. Archive reference: pending.

6. OpenAI, “Advancing content provenance for a safer, more transparent AI ecosystem,” May 19, 2026, updated July 31, 2026; and OpenAI Help Center, “C2PA and SynthID in OpenAI-generated images,” current edition. https://openai.com/index/advancing-content-provenance/ and https://help.openai.com/en/articles/8912793-c2pa-and-synthid-in-openai-generated-images. Accessed September 12, 2026. OpenAI says supported generated images and audio can carry C2PA Content Credentials and SynthID signals and that its public verification system and API can detect supported provenance signals. The company states that detection can indicate origin in OpenAI tools but does not establish accuracy, legal ownership, editing history in every case, or contextual truth. These are provider claims. OpenAI systems are used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH19-S6. Archive reference: pending.

7. European Union, Regulation (EU) 2024/1689, Article 50; European Commission, “Code of Practice on Transparency of AI-generated Content,” final code and implementation materials, updated July 31, 2026. https://eur-lex.europa.eu/eli/reg/2024/1689/2026-07-27/eng and https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content. Accessed September 12, 2026. The July 27, 2026 consolidated EUR-Lex text is verified. Article 50 transparency obligations applicable from August 2, 2026 require providers of covered generative AI systems to mark synthetic outputs in machine-readable form as technically feasible and require deployers to disclose deepfakes and certain AI-generated or manipulated public-interest text. The Commission’s voluntary Code of Practice and guidelines support implementation; the Commission reported roughly 190 company and organizational signatories by the end of July. Source record: DA-CH19-S7. Archive reference: pending.

8. Federal Trade Commission, “FTC Begins Enforcing the TAKE IT DOWN Act,” May 19, 2026; “Complying With the Take It Down Act,” 2026; and “FTC Sends Warning Letters to Companies About Compliance with the TAKE IT DOWN Act,” May 2026. https://www.ftc.gov/news-events/news/press-releases/2026/05/ftc-begins-enforcing-take-it-down-act, https://www.ftc.gov/business-guidance/resources/complying-take-it-down-act, and https://www.ftc.gov/news-events/news/press-releases/2026/05/ftc-sends-warning-letters-companies-about-compliance-take-it-down-act. Accessed September 12, 2026. Section 3 of the federal law requires covered platforms to provide a notice-and-removal process for nonconsensual intimate images, including AI-created digital forgeries, and remove covered content and known identical copies within 48 hours of a valid request. FTC enforcement of the platform obligations began May 19, 2026. The agency also sent warning letters to 12 websites offering so-called nudify tools; the letters are compliance notices, not findings that the recipients violated the Act. The chapter uses the law as a narrow example of shifting verification and response duties upstream, not as a general synthetic-media regime. Source record: DA-CH19-S8. Archive reference: pending.

9. Mike Scarcella, Reuters, “ChatGPT invented fake police testimony in murder appeal, New Mexico high court says,” September 11, 2026. https://www.reuters.com/legal/government/chatgpt-invented-fake-police-testimony-murder-appeal-new-mexico-high-court-says-2026-09-11/. Accessed September 12, 2026. The New Mexico Supreme Court fined attorney Stephen Aarons $5,000 and held him in contempt after an appellate filing contained false testimony and wholly fabricated witnesses. Aarons told Reuters he had used ChatGPT to summarize trial proceedings and had not understood the degree to which it could fabricate facts. The chapter uses the case as an institutional extension of the authority-jump logic from Chapter Twelve, not as a redefinition of the authority jump’s core meaning as the point where machine output becomes state action. OpenAI systems are used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH19-S9. Archive reference: pending.

10. Hanna Gryncwajg, LinkedIn post, screenshot supplied by George L. Taylor, September 25, 2026 (post displayed as one week old and edited). The screenshot shows only the beginning of her account and does not establish the operator, use of AI, or outcome. Used as an anecdote, not independently verified case evidence. Federal Trade Commission, “Scammers impersonate well-known companies, recruit for fake jobs on LinkedIn and other job platforms,” August 8, 2023, https://consumer.ftc.gov/consumer-alerts/2023/08/scammers-impersonate-well-known-companies-recruit-fake-jobs-linkedin-other-job-platforms; “How to spot the latest job scams,” December 11, 2023, https://consumer.ftc.gov/consumer-alerts/2023/12/how-spot-latest-job-scams. Accessed September 25, 2026. Source record: DA-CH19-S10. Archive reference: pending.

11. OpenAI, “Our approach to EU text provenance rules,” October 5, 2026. https://openai.com/index/eu-text-provenance/ Accessed October 6, 2026. Provider announcement and self-reported evaluations, not independent validation. API opt-in starts October 5; eligible ChatGPT and Codex EU rollout remains prospective. Detector access is restricted initially. Used for rollout scope, false positives, missed signals, and the limits of inference about authorship, ownership, responsibility and truth. The institutional decision rule is the author’s analysis. OpenAI systems are used in this Living Edition’s production workflow; the standing production disclosure remains applicable. Source record: DA-CH19-S11. Update 069. Archive reference: pending.

12. Nicole Ramirez, LinkedIn post displayed as October 4, screenshot supplied by George L. Taylor, October 10, 2026. The reported “87% AI” result and older-writing tests are self-reported and not independently reproduced; the detector is unidentified. Source record: DA-CH19-S12. Update 084. Archive reference: author-retained screenshot dated October 10, 2026.

13. Turnitin, “Using the AI Writing Report,” current guidance. https://guides.turnitin.com/hc/en-us/articles/22774058814093-Using-the-AI-Writing-Report Accessed October 10, 2026. Provider guidance acknowledges possible misidentification and requires scrutiny and human judgment before adverse action. Source record: DA-CH19-S13. Update 084. Archive reference: pending.

14. Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou, “GPT detectors are biased against non-native English writers,” Patterns 4, 100779 (2023). https://doi.org/10.1016/j.patter.2023.100779 Author manuscript: https://arxiv.org/abs/2304.02819 Accessed October 10, 2026. Seven detectors were tested on 91 TOEFL essays and 88 U.S. eighth-grade essays. Findings concern those tools and samples in 2023, not all contemporary detectors or the Goncourt novel. Source record: DA-CH19-S14. Update 084. Archive reference: pending.

15. Reuters, “France’s prestigious Goncourt prize drops novel after author accused of using AI,” September 25, 2026. https://www.reuters.com/business/media-telecom/frances-prestigious-goncourt-prize-drops-novel-after-author-accused-using-ai-2026-09-25/ Accessed October 10, 2026. Records removal, denial and separate allegations; no authorship finding adopted here. Source record: DA-CH19-S15. Update 084. Archive reference: pending.

16. Anthropic, “Preparing for the 2026 US midterms,” October 14, 2026. https://www.anthropic.com/news/2026-us-midterms. Accessed October 15, 2026. Provider account of testing with the Institute for Strategic Dialogue and McCain Institute and subsequent classifiers; no separate testing report or independent validation of the mitigation was established here. Used for decomposed election-abuse requests and the announced response, not general political-bias percentages or a claim that the weakness is fixed. Anthropic and OpenAI tools are used in this Living Edition’s production; see the transparency statement. The connected-sequence recommendation is author analysis. Source record: DA-CH19-S16. Update 098. Archive reference: pending.