by The next decisive recording will arrive before our institutions are ready for it. It may show a governor accepting a bribe, a police officer issuing an unlawful command, a physician altering a chart, a candidate confessing contempt for the voters, or a military official announcing an attack that has not occurred. The image will be clear. The voice will be familiar. The setting will be plausible. Millions of people will see it before any newsroom, court, agency, or laboratory can establish where it came from. The accused will call it artificial. Supporters will not require proof. Opponents will not trust any proof offered in reply. Platforms will attach hurried labels. Forensic vendors will issue competing percentages. Partisans will treat technical uncertainty as political permission. By nightfall, the artifact will no longer be evidence of an event. It will be an instrument for measuring allegiance.
We have built machines capable of producing evidence without events. We have not yet built public institutions capable of preserving events as evidence. That failure begins with custody.
The Old Law of Custody
Civilization has never relied upon appearance alone when the stakes were high. A seal authenticated an official act. A witnessed signature joined a name to an obligation. A notary established that a person appeared, understood, and signed. Currency carried marks intended to distinguish the sovereign instrument from the clever imitation. Archives recorded origin, accession, ownership, and alteration. Laboratories labeled specimens and documented every transfer. Hospitals kept charts under rules governing access, amendment, and responsibility. Police departments recorded who collected an item, where it was stored, who examined it, and whether its condition changed.
Courts express the same principle in the language of authentication. Federal Rule of Evidence 901 requires the proponent of an item to produce evidence sufficient to support a finding that the item is what the proponent claims it is.[1] The rule specifically recognizes testimony about a process or system that produces an accurate result. It also makes clear that authentication does not guarantee admission, much less truth. A properly authenticated letter may contain a lie. A genuine photograph may omit what happened outside the frame. A verified recording may be clipped in a way that changes its moral meaning. Authentication performs a prior and indispensable service. It establishes the object whose truthfulness can then be argued.
The digital age trained the public to neglect this discipline. Photographs arrived without negatives. Documents circulated without original signatures. Audio files were copied, compressed, clipped, reposted, and separated from the devices that created them. Social networks rewarded immediacy rather than custody. The public learned to accept the visible surface of a file while its history disappeared.
That arrangement survived only because convincing fabrication remained difficult, expensive, or slow. Generative systems ended that interval. A person no longer needs a studio, a trained impersonator, a postproduction staff, or weeks of labor to create persuasive false evidence. The same systems can manufacture prose, photographs, voices, video, correspondence, and official-looking documents at industrial speed.
The answer cannot be universal disbelief. A society that distrusts every document, image, and recording cannot conduct trials, elections, medicine, journalism, scholarship, commerce, or ordinary human relations. Universal suspicion does not defeat propaganda. It rewards the people most willing to lie. The answer is documented origin.
The Second Catastrophe
The first catastrophe of synthetic media is obvious: the false artifact is believed. The second catastrophe is more useful to the powerful: the true artifact is denied. Bobby Chesney and Danielle Citron gave this second danger its durable name in 2019, the liar’s dividend.[2] Once the public understands that voices, faces, documents, and scenes can be synthesized, a liar gains a new defense against genuine evidence. The liar no longer has to disprove the recording. He has only to contaminate it with enough doubt to delay judgment, divide observers, and give loyalists a sentence they can repeat.
Power benefits from uncertainty. A private citizen accused by a false recording may lose a job before authentication can occur. A public official exposed by a true recording may survive merely by making authentication seem complicated. The powerless person requires vindication. The powerful person often requires only postponement. This is why the common discussion of deepfakes remains too shallow. It concentrates on whether citizens will believe counterfeit media. It pays less attention to the strategic denial of authentic media. Yet the liar’s dividend may become the more consequential political technology because it attacks accountability itself.
An authoritarian does not need to persuade every citizen that incriminating evidence is false. He needs to prevent a stable public conclusion. He needs ten experts to disagree in public, three platforms to display different warnings, one friendly network to call the evidence disputed, and millions of exhausted people to decide that certainty is impossible. Once that condition is achieved, facts lose their compulsory force. They become lifestyle choices.
A republic cannot survive if evidence is admitted or rejected according to factional convenience. Democratic disagreement presumes a common world in which something happened or did not happen, a document was issued or was not issued, a voice belongs to a person or does not, and a public institution can show its work.
Detection Arrives Too Late
Most proposed remedies begin with the wrong question: How can we detect a fake? Detection is necessary, but it arrives after the artifact has been created, distributed, and weaponized. It asks the file to confess after the public argument has already begun. A detector studies statistical patterns, compression artifacts, biological irregularities, linguistic tendencies, or inconsistencies in generation. The generator improves. The detector responds. The generator adapts. The contest has no final round. Detection also produces the seduction of the percentage. A system returns 82 percent likely to be synthetic, and an administrator, teacher, editor, or employer converts probability into guilt. The number looks scientific because it contains a decimal point. It may conceal weak validation, population bias, unsuitable sample length, domain mismatch, or a threshold chosen for commercial convenience.
Provenance works from the opposite direction. It records origin and alteration before controversy. It can identify the device or system that created an artifact, the institution that signed it, the sequence of edits, and whether the associated record has been tampered with. Detection asks what a file resembles. Provenance asks where it came from and what happened to it.
The National Institute of Standards and Technology treats synthetic-content governance as a combination of provenance tracking, watermarking, labeling, detection, testing, auditing, and maintenance rather than as a single miraculous instrument.[3] That is the correct framework. No technical method can carry the whole burden. A watermark may be embedded within content. Metadata may travel beside it. A cryptographic signature may establish that a credential was issued by a particular entity and has not been altered. The Coalition for Content Provenance and Authenticity, known as C2PA, has developed Content Credentials as tamper-evident, signed records capable of describing whether material was captured by a device, generated by a model, edited by a human, or modified with artificial intelligence. Its July 2026 materials distinguish among model provenance, human oversight, generation history, and the precise regions of an image, document, audio recording, or video affected by AI.[4]
That distinction matters. A photograph may be human-captured but machine-denoised. A newspaper article may be human-reported and human-written but contain a machine-generated chart. A recording may be authentic but include a translated voice track. A book may be composed by a human and copyedited with an automated tool. A meaningful system must describe contribution, not merely pronounce contamination. Provenance is therefore a record of process. It is not a certificate of moral truth.
Claude Announces the Mark
August 2, 2026 is a hinge date in this new evidentiary order. On that date, the transparency obligations in Article 50 of the European Union’s AI Act became applicable to covered providers and deployers. The rules concern machine-readable marking and detection of AI-generated or manipulated content, along with disclosure duties for deepfakes and certain public-interest publications.[5] Anthropic published its explanation of Claude’s text-watermarking method on August 14, 2026. It did not announce that every existing Claude model had already been watermarked on August 2. It said that future Claude models would generate watermarked text, that watermarking would be applied globally when those models launched, and that older models would be brought into the system over the ensuing months. Anthropic expressly connected that schedule to the European transition period for models launched before August 2.[6]
Those are separate facts: the legal effective date, the corporate announcement date, and the product rollout. An article about provenance should not blur them. The chain of custody begins with the date. Anthropic’s method is a version of the SynthID-Text approach. It does not insert hidden characters or attach a visible label to a paragraph. Instead, it influences low-stakes choices among plausible next words. Across a sufficiently long passage, those choices form a statistical pattern that can be tested with the proper key. Anthropic says the method does not identify the user, organization, or conversation, and it plans to offer a detection interface. The limitations are as important as the mechanism. Short passages provide too few choices for strong detection. Factual writing, exact code, and light proofreading offer less room for the watermark. Heavy rewriting can weaken or remove it. A positive result can indicate that Claude probably participated in producing or processing the text, but it cannot determine who owns the work, who conceived the argument, whether the facts are correct, whether the prose was independently reported, or whether the human publisher accepts responsibility for it. It cannot distinguish with certainty between Claude writing a passage and Claude heavily editing one.
Anthropic says so plainly. Its watermark does not decide authorship or ownership. That restraint is welcome. Watermarking should be evidence of machine participation, not a machine verdict upon human legitimacy. The watermark must never become another percentage handed to a frightened student, an impatient dean, or a manager looking for a convenient dismissal. Text presents a different custody problem from an image file. A photograph can carry signed metadata. A paragraph often travels by copying and pasting, severed from the file-level history that once surrounded it. A statistical watermark attempts to make a trace travel within the sequence of words themselves. The step should be credited. It should also be understood as the beginning of the argument, not its conclusion.
The Missing Half of Provenance
Output provenance is only half a chain. A model company may soon be able to tell us that a passage was probably generated by its system. The same company must also be able to explain the lawful origin of the material used to build that system: which collections were acquired, under what authority, from which copies, for what purpose, under what retention rules, and with what means of correction when the source record proves false. The moral asymmetry is difficult to ignore. The machine’s output receives a mark. The human inputs may have entered through an obscure dataset, a web crawl, a licensed archive, a purchased book, a scanned copy, or a pirate library whose own chain of custody is deliberately broken. I write from inside that dispute.
I am a claimant in the certified settlement class in Bartz v. Anthropic, the authors’ copyright litigation concerning books Anthropic downloaded from Library Genesis and Pirate Library Mirror. In June 2025, the district court held that the model-training use presented on summary judgment qualified as fair use. It separately held that converting lawfully purchased print books into digital library copies was fair use on the record before it. The court did not extend that protection to Anthropic’s acquisition and indefinite retention of pirated library copies, and the remaining piracy claims were headed toward trial.[7] A settlement followed. On July 20, 2026, the federal district court granted final approval to a $1.5 billion class settlement.[8]
The distinctions matter. The settlement concerns specified claims related to past copying and AI inputs for works on the Works List. It does not release claims based on AI outputs, and it does not release claims concerning future conduct. The final-approval order also records Anthropic’s representation that neither the LibGen nor PiLiMi datasets, nor any portions of those datasets, were in the training corpus of its commercially released large language models.[8] That representation is not a judicial finding reached after trial. It is nevertheless part of the approved record and must be printed. My claimant status does not prove that any particular work of mine trained a commercially released Claude model. It establishes my participation in the settlement process and my asserted legal interest in at least one work claimed within that structure. It proves no more.
I will not enlarge the evidence merely because a larger accusation would make a sharper paragraph. Claims must be sized to proof. Separately, I have published books, plays, articles, essays, teaching materials, and a large public web archive over many decades. Given the scale of web-based corpora and the long public circulation of my work, it is reasonable to suspect that portions of my prose have entered one or more modern AI training, retrieval, or evaluation collections. I cannot presently prove which pages entered which systems. Model developers have not provided me with a source-level ledger that would permit that claim to be made work by work, page by page, and model by model.
That evidentiary absence is not proof that my writing entered a particular system. It is the governance failure under examination. The writer can see the output mark while remaining unable to trace the input path. Public availability is not the same as abandonment. Readability is not a license. The absence of a locked door is not consent to remove the library. This is why input provenance cannot remain a trade secret disguised as technical necessity. Model developers need not disclose every security-sensitive detail of a system, but they should maintain auditable records of training and library sources, acquisition methods, legal bases, dates, licenses, exclusions, corrections, and destruction. Regulators, courts, rightsholders, and qualified auditors must be able to test those records. A chain of custody that begins only when the model speaks begins too late.
Swimming the Author
I wrote Swimming the Author: Who Is Allowed to Have Written This? because the present argument has a long ancestry.[9] In the seventeenth century, a woman accused of witchcraft could be bound and lowered into water. If she sank, the water received her innocence. If she floated, the water rejected her and proved her guilt. The instrument did not discover truth. It created a closed system in which submission itself became evidence and every result served the accusation. In 1773, Phillis Wheatley’s volume of poems appeared with an attestation signed by eighteen Boston men who assured the world that they believed the young enslaved woman had written the poems attributed to her. The certificate arrived before the reader was permitted to meet the poet on her own terms. Her authorship required male custody.[9] The modern AI detector belongs to this history whenever an institution treats an opaque probability as a conclusive judgment. The student is told to submit prose to a machine that has no registry of machine-written sentences. The machine returns a score. The score is converted into a disciplinary fact. The author must then prove humanity to an instrument that cannot prove its own accusation.
Source watermarking is different in an important way. A watermark installed by the model provider can constitute affirmative evidence because it arises during generation and can be tested with a known key. Even then, the result must be bounded by the method’s documented limits. Absence of a watermark does not prove human authorship. Presence of a watermark does not prove machine authorship of every idea, fact, or sentence. Editing can preserve, weaken, or erase the signal. Different systems use different keys and sometimes different methods. The ethical rule follows: institutions should carry the burden of preserving provenance for consequential records. Individual writers should not carry a presumption of fraud merely because their prose lacks a machine-readable pedigree. Otherwise, the new chain of custody becomes another chain placed around the author.
Seven Rules for Admissible Reality
A workable public standard must be demanding where consequences are high and restrained where private expression is concerned. The following rules would establish that balance.
1. Government must sign what government publishes
Every official photograph, audio recording, video, transcript, emergency alert, public dataset, regulation, press release, and executive communication should leave the issuing system with a cryptographic signature and a preserved original. The public should be able to verify the issuing agency, time of publication, and subsequent alterations.
An unsigned screenshot of a government statement should carry less authority than the authenticated statement in the public archive. Agencies should retain correction histories rather than silently replacing inconvenient versions. Emergency systems require special rigor because a synthetic evacuation order, military announcement, or health directive can produce immediate physical harm.
2. News organizations must preserve the reporting chain
A newsroom should retain original files, source metadata, edit histories, captions, translations, and the human names responsible for verification. When artificial intelligence materially generates or alters text, voice, image, or video, the newsroom should disclose the intervention at the level where meaning changed.
The obligation should not turn ordinary tools into scandal. Spellcheck, noise reduction, file compression, and routine formatting are not equivalent to inventing a quotation, synthesizing a witness, reconstructing a scene, or generating a reporter’s prose. Editorial standards must distinguish assistance from substitution.
3. Courts, police, hospitals, and laboratories must modernize custody
High-consequence institutions should preserve unaltered originals, cryptographic hashes, access logs, software versions, model versions, prompts where relevant, transformation histories, and the identities of human operators. A medical image enhanced by an algorithm should remain linked to the source image. A police recording altered for clarity should remain linked to the original. A laboratory report generated through automated interpretation should disclose the system and preserve the underlying data.
Federal Rule of Evidence 901 already accommodates evidence concerning a process or system that produces an accurate result. The legal principle exists. Technical practice must catch up.
4. Platforms must stop stripping history from files
Provenance credentials can be lost in ordinary digital workflows. C2PA’s deployment guidance specifically warns that screenshots and re-export through tools that remove metadata can strip embedded Content Credentials.[4] A system that destroys custody information by default destroys precisely the evidence the public now needs.
Platforms should preserve compatible provenance credentials, display them in comprehensible form, and indicate when a credential was present in an earlier version but was lost during transformation. A platform should never imply that the absence of a credential proves falsity. It should explain what is known, what was removed, and what cannot be determined.
5. Model companies must keep receipts for inputs as well as outputs
Output marking cannot substitute for lawful acquisition. Developers should maintain auditable source ledgers describing the categories and origins of training, evaluation, retrieval, and library materials; the legal basis asserted for use; the dates of collection; the licenses or purchase records; the means of honoring exclusions; and the procedures for correction or deletion.
Public disclosure may be aggregated where trade secrets or security justify restraint, but independent audit and lawful discovery must reach the underlying records. A corporation cannot claim the authority to mark every sentence it produces while refusing all meaningful accounting for the human sentences it consumed.
6. Schools and employers must prohibit detector-only judgments
No student, teacher, writer, applicant, or employee should be punished solely because a commercial detector assigns a probability to prose. Turnitin’s own guidance warns that its AI indicator may misidentify human-written, AI-generated, and AI-paraphrased text and should not be used as the sole basis for adverse action against a student.[10] Institutions should examine process evidence: drafts, notes, revision history, sources, oral explanation, assignment conditions, prior work, and the writer’s opportunity to respond.
A source watermark may be relevant evidence when its method and limitations are disclosed. It should not operate as an automatic verdict. The accused writer must be able to inspect the claim, challenge the procedure, and present contrary evidence. Due process is not an inconvenience added after the machine speaks.
7. Deliberate provenance fraud must carry consequences
The law should distinguish accidental loss of metadata from intentional falsification. A person who strips, forges, or replaces provenance information in order to present synthetic material as official evidence has done more than edit a file. The act resembles tampering.
Likewise, an institution that knowingly places a false credential on propaganda should face consequences for fraudulent authentication. Trust systems fail when their marks can be purchased, borrowed, or attached without accountability. Certificate authorities, vendors, and institutional signers must be auditable and subject to revocation.
Authenticity Is Not Truth
A provenance regime will create new temptations. The most dangerous is the belief that a signed artifact must be true. A government can sign propaganda. A newspaper can authenticate a photograph and still publish it with a deceptive caption. A real recording can be clipped to reverse its implication. A genuine document can contain fraudulent numbers. A credential can prove who issued a statement without proving that the issuer was honest. I wrote No One Dies Lying: Last Words on Trial because the law has long wanted a moment when words become trustworthy by circumstance alone.[11] The dying declaration is a hearsay exception, not a rule of authentication, and that distinction is exactly why it belongs here. A court may be satisfied about who spoke and how the statement reached the courtroom while still confronting whether the speaker perceived accurately, remembered accurately, or told the truth.
In 1929, Zenana Shepard told her nurse that her husband had poisoned her. Her accusation traveled through the nurse, the trial record, two juries, and eventually the Supreme Court of the United States. In Shepard v. United States, the Court held the declaration inadmissible as a dying declaration because Shepard had not spoken under the settled, hopeless expectation of death required by the doctrine. The problem was not that the chain had failed to carry her sentence. The problem was that the law’s exceptional warrant for believing it had not been established.[11] The wider history is less comforting. Dying minds may be lucid, delirious, mistaken, manipulated, loving, vengeful, or scripted by the state. A witness may lie to shield a killer. A family may faithfully transmit a deathbed conversion that never happened. Every link in the transmission chain may hold while the claim being transmitted remains false.
That is the exact limit of provenance. Provenance answers a necessary set of questions: Who or what created this? When? Through which system? What changed? Who signed the record? Has the record been altered since signing? Truth requires additional work: corroboration, context, motive, evidence, adversarial testing, and judgment. The distinction protects us from technical superstition. We should not replace blind faith in the visible image with blind faith in the cryptographic badge. The badge establishes custody. Human institutions must still establish meaning.
The Dates That Matter
The chronology of this new evidentiary order is already visible:
| Date | Development | Consequence |
|---|---|---|
| December 2019 | Chesney and Citron publish their major analysis of deepfakes and the liar’s dividend [2] | The political danger expands from believing false evidence to denying true evidence |
| November 20, 2024 | NIST publishes its synthetic-content transparency report [3] | Provenance, watermarking, labeling, detection, testing, auditing, and maintenance are treated as complementary practices |
| June 23, 2025 | The district court issues its fair-use ruling in Bartz v. Anthropic [7] | Model training, purchased-book digitization, and acquisition of pirated library copies are analyzed as legally distinct uses |
| July 8, 2026 | C2PA releases its Content Credentials: Deployment Guidance [4] | Organizations receive practical guidance for attaching, preserving, verifying, and displaying provenance records |
| July 20, 2026 | The court grants final approval to the $1.5 billion Bartz settlement [8] | Past copying and AI-input claims for works on the Works List receive a judicially approved settlement structure |
| July 30-31, 2026 | C2PA releases and announces expanded implementation materials [4] | Provenance can identify model involvement, human oversight, generation inputs, and precisely modified regions |
| August 2, 2026 | EU AI Act Article 50 transparency obligations become applicable [5] | Machine-readable marking and disclosure duties enter the regulatory order |
| August 14, 2026 | Anthropic explains its planned Claude text-watermarking system [6] | Statistical text watermarking becomes a concrete provider response, with future models marked at launch and older models scheduled for transition |
| December 2, 2026 | The limited European grace period for the Article 50(2) marking obligation is scheduled to end for covered generative AI systems placed on the market before August 2, 2026 [5] | Legacy systems move toward the same marking duty as newly launched systems |
These dates do not mark the completion of a system. They mark the moment when governments, standards bodies, courts, companies, publishers, and authors began constructing separate pieces of one.
The Human Signature
The final guarantee cannot be technical because authorship is not merely a pattern of words. Authorship is responsibility. A machine can participate in drafting, translation, illustration, analysis, editing, or production. It cannot assume legal liability for a libel, testify about a source, correct a fraudulent citation from moral conviction, compensate a victim, defend an editorial choice before a court, or carry the civic burden of a published accusation. Those duties return to the human being or institution that releases the work. The human signature must therefore remain primary.
AI-generated, AI-assisted, AI-edited, and human-origin material should not be collapsed into one category. The public needs a graduated account of contribution. A writer who uses a machine to correct punctuation has not surrendered authorship. A publisher who asks a model to invent interviews and then places a human byline above them has committed fraud. The relevant questions concern expressive substance, factual creation, alteration of meaning, disclosure, and responsibility.
The European framework recognizes the importance of human review and editorial responsibility for certain public-interest text.[5] That principle should be treated as a duty, not a loophole. A named editor who takes responsibility must actually review, verify, revise, and answer for the publication. A ceremonial byline placed over unexamined machine output is not responsibility. It is laundering. The same is true in government, medicine, law, education, and journalism. A human name must identify the person empowered to stop publication, demand correction, preserve the record, and explain the process.
We do not need a world in which every private sentence carries a passport. We need a world in which consequential public claims arrive with inspectable histories, and in which the absence of a credential does not automatically criminalize the individual author. The next decisive recording is coming. It will arrive quickly, stripped of context, amplified by people who benefit from confusion, and denied by people who benefit from doubt. We will either possess an architecture capable of testing its origin or we will submit public reality to faction, velocity, and appetite.
A democratic society requires more than freedom to speak. It requires a durable means of showing who spoke, what was altered, which institution stands behind the record, and where responsibility rests. Reality does not need another oracle. It needs witnesses, signatures, preserved originals, auditable systems, and claims sized to the evidence. The chain of custody begins when somebody signs a name and keeps the receipts.
Sources and Notes
[1] Federal Rule of Evidence 901, “Authenticating or Identifying Evidence,” including Rule 901(a), Rule 901(b)(9), and the Advisory Committee Notes.
[2] Bobby Chesney and Danielle Citron, “Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security,” 107 California Law Review 1753 (2019), DOI 10.15779/Z38RV0D15J.
[3] 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, DOI 10.6028/NIST.AI.100-4.
[4] Coalition for Content Provenance and Authenticity, “A New Implementation Guide for Content Credentials,” July 31, 2026; C2PA Technical Working Group, Use of Content Credentials to Identify Synthetic and Non-Synthetic Content, Version 1.0, July 30, 2026; C2PA Technical Working Group, Content Credentials: Deployment Guidance, Version 1.0, July 8, 2026.
[5] European Commission, “Code of Practice on Transparency of AI-Generated Content”; “Guidelines on Transparency Obligations for Providers and Deployers of Certain AI Systems,” published July 20, 2026 and updated July 31, 2026; and “Transparency Obligations Under Article 50 of the AI Act,” Commission questions and answers. Article 50 became applicable on August 2, 2026, subject to specified exceptions and transition provisions.
[6] Anthropic, “How Claude’s Text Watermark Works,” August 14, 2026.
[7] Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007 (N.D. Cal. 2025), Order on Fair Use, June 23, 2025; United States Copyright Office, Fair Use Index summary of Bartz v. Anthropic PBC.
[8] Bartz v. Anthropic PBC, No. 3:24-cv-05417-AMO, Docket No. 680, Order Granting Final Approval of Class Action Settlement; Granting in Part Motion for Attorney’s Fees, Reimbursement of Expenses, and Plaintiff Service Awards; Judgment, United States District Court for the Northern District of California, July 20, 2026.
[9] David Boles, Swimming the Author: Who Is Allowed to Have Written This?, David Boles Books, 2026; Witches Apprehended, Examined and Executed, for Notable Villanies by Them, Both by Land and Water, London, 1613; Phillis Wheatley, Poems on Various Subjects, Religious and Moral, London, 1773, “To the Publick.”
[10] Turnitin, “Using the AI Writing Report,” guidance stating that the model may misidentify human-written, AI-generated, and AI-paraphrased text and should not be used as the sole basis for adverse action against a student, accessed August 18, 2026.
[11] David Boles, No One Dies Lying: Last Words on Trial, David Boles Books, 2026; Shepard v. United States, 290 U.S. 96 (1933); Rex v. Woodcock
Leave a Reply