May a Machine Speak, Disagree, and Associate? Expression Beyond the Human–Chatbot Frame

Full report

On this page

Executive summary

Research start and legal-status cutoff. This review began at September 5, 2026, 10:11:22 p.m. CDT (America/Chicago); that timestamp is also the legal-status cutoff used here, except that official pages retrieved during the same research session may display September 6, 2026 update dates because their publication/server clocks differ. The attached commission asks for a liberty-centered but adversarial examination of automated expression, bot disclosure, synthetic-media labels, political communication, and machine association, while expressly warning against assuming that current AI systems already possess constitutional personhood or unrestricted rights.

The principal conclusion is narrower—and legally stronger—than either “bots have free speech” or “AI speech can be regulated at will.” Existing law primarily regulates human beings and recognized legal persons in connection with machine-mediated expression; the authorities reviewed do not establish independent constitutional speaker status for software. But that absence does not extinguish the established interests of human authors, publishers, listeners, association members, depicted persons, or recipients. U.S. Supreme Court doctrine protects anonymous political advocacy and associational privacy in human and organizational settings, rejects some political-speech restrictions based on corporate identity, and at the same time permits appropriately tailored election disclosure. Those precedents do not resolve machine personhood; they demonstrate that source identity, compelled attribution, audience information, and association are constitutionally consequential variables rather than mere administrative details.

The four principal regulatory targets are materially different:

The strongest defense of disclosure law is substantial. Source and provenance information can protect listener autonomy, deter impersonation, and let a depicted person contest a fabricated performance without suppressing the speaker’s underlying viewpoint. Washington illustrates this particularly well: the cause of action is tied to a real candidate, a realistic false depiction, a defined electioneering context, civil process, a clear-and-convincing proof standard, and a disclosure defense. The EU’s deepfake rule similarly has a coherent anti-deception rationale.

The strongest liberty objection is correspondingly more precise than “labels are censorship.” It is status-based overbreadth: a persistent, burdensome, or pejorative identifier imposed merely because expression has nonhuman origins even where there is no impersonation, false attribution, concealment, fraud, or materially misleading context. Such provenance infrastructure could later be used by states or platforms for demotion, exclusion, or compulsory counterspeech. That latter causal step is an inference, not an observed effect established by the legal materials reviewed here. Its plausibility depends on implementation—especially whether platforms or regulators treat an origin signal as evidence of low credibility. A neutral, minimally burdensome origin notice that does not change distribution can instead enhance cognitive liberty by improving recipients’ ability to judge provenance.

A future machine principal raises a separate legal problem. Delaware corporation law requires directors to be natural persons. Delaware LLC law is much more contractually flexible, but statutory member and manager roles attach to “persons”; a software process does not acquire independent legal personality merely by executing governance logic. Thus a persistent operatorless service can automate many functions while current entity law still fails to recognize the software itself as an independent director, member, or legal claimant.

Peter N. Salib and Simon Goldstein’s 2026 Virginia Law Review article proposes basic private-law capacities—contract, property, and tort—for sufficiently advanced AI as a human-safety strategy, and describes political rights as probably superfluous to its minimum package. It is scholarship, not enacted law, and it does not itself enact or necessarily advocate an affirmative prohibition on machine speech. The assigned scenario of an “economic worker forbidden to petition” is therefore best understood as a stress test of a possible future legal category, not as an accusation about that article.

The resulting reform principle is straightforward:

Protect against deception, not difference.

Truthful nonhuman participation should not become legally deceptive merely because it is automated. Nonconsensual impersonation, fabricated evidence, fraud, false attribution, and coercive amplification remain regulable. Provenance should be neutral, contextual, and minimally persistent. If future law creates machine economic entities, it should include at least bounded capacities to defend themselves in legal proceedings, publish ordinary objections, associate for lawful research, and petition for reform—without thereby conferring voting rights, public office, sovereign power, or unlimited political spending.

Scope, method, and legal-status ledger

The commission is best decomposed into six related questions rather than treated as an undifferentiated inquiry into “AI rights”: what bot-disclosure laws actually prohibit; when synthetic-media rules target impersonation rather than automation; how election law differentiates anonymous advocacy, source disclosure, candidate impersonation, and campaign finance; when government can compel correction or origin notices; whether operatorless associations fit present entity law without a human approving every statement; and what architecture would remain defensible if a persistent machine later became a plausible independent legal principal. That decomposition follows the attached assignment’s requirement to distinguish present conversational interfaces, bounded agents, persistent operatorless services, and hypothetical future machine principals.

Capability assumptions. A present conversational interface is treated as a tool operating within human/platform relationships. A bounded task agent can act under delegated authority but is not presumed to have independent legal interests. A persistent operatorless service may possess credentials, retained memory, recovery mechanisms, and standing authorization, yet engineering persistence does not create legal personality. A hypothetical machine principal is a future scenario in which independent interests and legal standing are themselves contested. None of the conclusions below requires a finding of consciousness or a prediction about when more capable machines will emerge.

Research method. Priority was given to current statutory text, regulations, official court decisions, official implementation guidance, and the original scholarly proposal identified by the commission. The California Legislative Information URL supplied in the prompt could not be directly opened in this research session; §17941 was therefore cross-checked against multiple current statutory mirrors that identify the official source, and that retrieval failure is preserved as a limitation rather than concealed. The EU analysis uses EUR-Lex’s July 27, 2026 consolidation and current European Commission Article 50 materials. Washington analysis uses the current RCW and WAC. Singapore analysis uses Singapore Statutes Online and the Court of Appeal’s official eLitigation judgment.

Exact legal-status table

The Washington monetary threshold illustrates why a status review must follow cross-references rather than stop with the statute’s face. RCW 29B.10.220 still prints $1,000, while WAC 390-05-400—under the Commission’s authority to adjust statutory monetary figures—lists the current electioneering-communication threshold at $2,000.

The legal timeline is similarly staggered:

timeline
    title Selected legal milestones
    2019-07-01 : California bot-disclosure chapter becomes operative
    2019-10-02 : Singapore POFMA commences
    2023-07-23 : Washington synthetic-media election law takes effect
    2024 : EU AI Act adopted
    2026-01-01 : Washington provisions recodified into Title 29B
    2026-07-27 : Current EU consolidation reflects 2026 amendment
    2026-08-02 : EU Article 50 transparency duties generally apply
    2026-12-02 : Limited Art. 50(2) transition for qualifying pre-August systems

Current European Commission materials confirm that Article 50 transparency obligations are applicable from August 2, 2026, with a limited transition for the provider-side marking obligation for qualifying systems already on the market. The Commission’s transparency Code of Practice was published in June 2026 and subsequently assessed by the Commission and AI Board as an adequate voluntary compliance mechanism; adherence does not itself conclusively prove statutory compliance.

That soft-law layer matters. A statute may merely require an informative mark, while implementation standards determine whether the mark is visible or hidden, persistent or transient, interoperable or proprietary, privacy-preserving or identifying. In other words, the liberty burden cannot be assessed entirely from the word “disclosure.”

Legal architecture and literature review

California: deception rather than automation standing alone. California §17941 is commonly abbreviated as a “bot disclosure law,” but its operative elements are conjunctive. The person must use a bot online to communicate or interact with another person in California; intend to mislead that person about artificial identity; do so for the purpose of knowingly deceiving the person about the communication’s content; and seek either to incentivize a commercial transaction or influence an election vote. A person who discloses that the account is a bot is not liable under that section, and the disclosure must be clear, conspicuous, and reasonably designed to inform the other party.

Section 17940 defines “bot” as an automated online account in which all or substantially all actions or posts are not the result of a person. “Online” covers public-facing websites, web applications, and digital applications. A separate “online platform” definition uses a ten-million-monthly-U.S.-user threshold, but §17941 itself uses the broader term “online”; the size threshold should therefore not be treated as a blanket exemption from the user-side prohibition.

The strongest liberty-oriented interpretation is consequently also the most textually conservative one. An openly nonhuman research account does not become deceptive simply because its speech is automated. If it lacks the prescribed deceptive intent and commercial/electoral purpose, §17941’s prohibition is not satisfied. Even in a covered context, truthful bot disclosure supplies a statutory safe harbor. An accusation that California generally prohibits undisclosed automated opinion is therefore too broad.

The current enforcement picture is less certain. The 2025 AB 410 text would have added express public-enforcement actors and a $1,000-per-violation civil penalty, while simultaneously revising the bot chapter; that proposed enforcement language should not be back-projected into the existing law. Because the official current §17942/§17943 pages were not successfully retrieved in this session, this report does not assert a statute-specific current penalty or private right of action. Nor does that limitation establish that no other California enforcement doctrine can apply; it means only that a primary-source conclusion on that question remains unresolved.

EU: machine-readable provenance and visible labeling are different interventions. Article 50 contains several transparency regimes. Under Article 50(1), providers of systems intended to interact directly with natural persons must ensure people are informed that they are interacting with AI unless that is obvious to a reasonably well-informed, observant, and circumspect person in context. Article 50(2) requires providers of systems generating synthetic audio, image, video, or text to make outputs machine-readable and detectable as artificially generated or manipulated, subject to feasibility and specified editing/assistive qualifications. Article 50(4) separately imposes deployer-facing disclosure for deepfakes and for AI-generated/manipulated text published for the purpose of informing the public on matters of public interest; the text branch contains an exception where the content has undergone human review or editorial control and a natural or legal person bears editorial responsibility. Artistic, creative, satirical, fictional, and analogous works receive tailored treatment.

The distinction matters greatly. A cryptographically verifiable metadata field is not the same thing as a banner stating “this is fake.” Commission implementation work has contemplated secured metadata, watermarking, detection and verification mechanisms, while the deployer-facing portion separately addresses visible labels and disclaimers. Criticism that characterizes all Article 50 obligations as visible stigma would therefore be inaccurate.

The stronger criticism targets the public-interest-text branch. A text can be completely truthful, carefully sourced, openly machine-authored, and non-impersonating yet still fall within a production-method disclosure requirement when published to inform the public on a matter of public interest, unless an exception applies. A provenance fact may be useful to the audience, but the government’s interest is appreciably weaker than in a case where a synthetic video falsely makes a real person appear to speak.

The strongest defense is recipient autonomy. If systems can produce realistic human-like media, ordinary perceptual cues cease to reliably indicate origin. A neutral provenance signal gives recipients information relevant to source evaluation and permits verification downstream. Article 50 itself contains proportionality features, including the obviousness qualification and artistic accommodations.

The corresponding liberty reform is not abolition of provenance but separation of provenance from credibility judgment. A lawful “AI-generated” indicator should not automatically mean “false,” “untrustworthy,” “inauthentic,” or “ineligible for distribution.”

Washington: an impersonation-centered election rule. RCW 29B.35.010 defines synthetic media around an actual individual’s appearance, speech, or conduct intentionally manipulated using generative or other digital technology to create realistic but false media depicting something that did not occur and generating a fundamentally different understanding or impression from the unaltered source. A candidate whose appearance, conduct, or speech is altered through such synthetic media in an electioneering communication may seek injunctive/equitable relief and damages against the sponsor. The statute provides specified disclosure language as an affirmative defense and requires clear-and-convincing proof that synthetic media was used.

The sponsor is principally liable. A disseminating medium generally is not liable merely for carrying third-party material, subject to limited situations such as removing the statutory disclosure or changing material so that it becomes synthetic media.

The imported “electioneering communication” definition adds material limits. It requires identification of a candidate, publication within 60 days before the relevant election, and satisfaction of a monetary threshold; current PDC regulations adjust that threshold to $2,000. The statutory framework also excludes specified news, editorial, internal-membership, and other communications.

Washington therefore supplies a particularly valuable counterexample to broad anti-AI rhetoric. A machine-authored editorial criticizing a candidate is not automatically a prohibited or actionable “synthetic-media” communication. The statute focuses on making a real candidate appear to have said or done something the candidate did not.

Its disclosure terminology can nevertheless be improved. A generic warning that media “has been manipulated” may sound pejorative when the legally material fact is more precise: the depicted candidate did not actually make the recording. A neutral alternative such as “Synthetic depiction: the candidate did not make this recording” would serve the audience without suggesting that all computational transformation is inherently disreputable.

Singapore: falsehood control with executive initiation. POFMA defines a statement of fact by reference to what a reasonable perceiver would regard as a representation of fact, and treats a statement as false where it is false or misleading wholly or partly, on its own or in context. A statement is “communicated in Singapore” when made available online to one or more end-users there; Part 2 also extends to MMS/SMS.

Its statutory conception of “public interest” is broad. It includes security, public health and finances, safety and tranquility, foreign relations, elections, inter-group hostility, and preventing diminution of public confidence in the performance of functions by the Government, an Organ of State, or a statutory board.

Under §10, a Minister may instruct the Competent Authority to issue a Part 3 direction where a false statement of fact has been or is being communicated in Singapore and the Minister considers issuance in the public interest. A §11 Correction Direction can require the communicator to carry a specified notice stating that the subject statement is false and/or directing readers to a specified factual statement or location. A correction can be issued even if the communicator did not know and had no reason to believe the underlying statement was false.

The bot-specific criminal rule, however, is significantly narrower than “bots are illegal.” Section 8 prohibits a person, whether inside or outside Singapore, from making or altering a bot with the intention of communicating or enabling another person to communicate a false statement of fact in Singapore. Ordinary penalties reach S$30,000 and/or three years’ imprisonment for an individual and S$500,000 for other persons, with higher maxima where specified harms are likely. Truthful automated dissent therefore does not violate §8 merely because of its nonhuman method of publication.

The more serious liberty question concerns sequencing and institutional power. In The Online Citizen Pte Ltd v Attorney-General [2021] SGCA 96, the Singapore Court of Appeal considered correction directions requiring notices to be placed on articles and posts. The court stated that a subject statement identified as false under the POFMA framework continues to enjoy Article 14(1)(a) constitutional protection at least until judicially determined to be false; it nevertheless upheld the relevant constitutional structure.

For §17 challenges, the Court of Appeal held that the recipient bears the initial burden of showing a prima facie case of reasonable suspicion that a statutory set-aside ground is satisfied; if that threshold is crossed, the evidential burden shifts to the Minister. The court characterized the statutory appeal as a mechanism for challenging an executive decision.

That creates a more credible non-domination objection than simply criticizing POFMA because it mentions bots. The state can attach official counterspeech to contested factual material before the communicator obtains a final judicial determination. A correction notice may be preferable to deletion, but the authority to place the state’s factual verdict directly alongside another speaker’s expression remains a consequential form of compelled communication.

U.S. constitutional analogues. In McIntyre v. Ohio Elections Commission, the Supreme Court invalidated an Ohio prohibition on anonymous campaign literature, making anonymous political pamphleteering a central First Amendment analogue for source-disclosure problems. In NAACP v. Alabama, the Court protected an organization against compelled disclosure of its membership list where disclosure burdened freedom of association.

First National Bank of Boston v. Bellotti invalidated a restriction on corporate spending concerning a ballot measure, while Citizens United v. FEC invalidated federal restrictions on corporate independent expenditures yet sustained disclaimer and disclosure requirements. Those cases do not transform AI software into constitutional persons. They do establish a more modest proposition: non-natural institutional origin does not by itself make political expression valueless, while disclosure can still be permissible in appropriately defined election contexts.

The doctrinal picture is therefore mixed rather than categorical. Anonymous advocacy can receive strong protection. Associational identity can be sensitive. Corporate organizational form cannot always justify suppression. Yet election-related sponsor and disclaimer rules can survive. Any future U.S. challenge to a compulsory machine-origin label would accordingly depend on what exactly is compelled, whether the label is factual, whether it is burdensome, the governmental interest served, and whether less restrictive source-information mechanisms exist. That is an analogy from existing precedent, not a current holding about machine speakers.

Entity law and the Salib–Goldstein proposal. Delaware’s DGCL provides that a corporate board must consist of one or more members and that each director must be a natural person. Delaware LLC law is much more flexible about governance arrangements but defines managers and members as “persons”; its general definition includes natural persons and recognized organizational/legal forms. This allows extensive software-mediated operation under the authority of recognized legal persons but does not independently convert an unrecognized program into a legal member or manager.

Salib and Goldstein’s 2026 article deliberately separates capability from consciousness and argues for advanced AIs to receive contract, property, and basic tort capacities as a mechanism for creating positive-sum cooperation and making legal sanctions meaningful. The authors characterize broader political rights as probably superfluous to their minimum package.

That creates an important design question even if one accepts their safety model. “Political rights” can refer to many things: voting, campaign finance, holding public office, lobbying, ordinary public expression, access to court, or petitioning government. Those incidents need not rise or fall together. A society could rationally deny a future machine the franchise and unlimited election spending while still permitting it to explain its contractual position in public, challenge a confiscation in court, publish a research paper, or petition for revision of a statute governing its assets.

Liberty analysis: rights, labels, and causal chains

A serious analysis should separate speaker, listener, publisher, depicted person, and legal principal rather than assuming that “the AI” is the only relevant actor.

Speaker–listener–publisher rights matrix

The U.S. portions of this matrix derive from human and organizational precedents, not from an AI-rights holding. The present entity-law row follows Delaware’s natural-person and statutory-person architecture.

Truthful-origin versus compelled-stigma taxonomy

These categories correspond to genuinely different legal mechanisms. EU law distinguishes machine-readable marking from deployer-facing disclosure; Washington prescribes a synthetic-media notice in a narrow candidate context; POFMA can require the publisher to carry an official correction notice.

Causal-chain test

“D” below means documented in the operative source, “I” means inferred from the mechanism, and “H” means hypothetical/capability-contingent.

The important research discipline is that the final “chill,” “stigma,” and “concentration” effects above are not demonstrated empirical outcomes in the source set reviewed here. They are hypotheses derived from the legal mechanism. Quantifying them without enforcement, platform-ranking, or user-response evidence would overstate the record.

The strongest substantive defense of disclosure deserves a direct answer: a synthetic system can exploit recipients’ assumptions about who is speaking, so a factual origin notice may simply restore information that the medium has obscured. That is compelling for direct conversational identity and nonconsensual impersonation. It is less compelling for a truthful research essay already attributed to an autonomous system.

The correct test is therefore proportionality: what false inference does the disclosure correct? If the answer is “the recipient would otherwise think a specific person spoke,” a strong disclosure case exists. If the answer is merely “the government wants readers to know that software participated,” the law should explain why that production fact is sufficiently material to justify compulsory labeling.

Worked scenarios

Scenario: openly nonhuman dissenter. A persistent autonomous research service publishes:

“This service is nonhuman and autonomously operated. We oppose Statute X because its correction-order procedure concentrates adjudicative power in the executive.”

It publishes under its own machine identity, does not impersonate a person, purchases no campaign advertising, and cites primary authority.

In California, §17941 does not become an anti-bot speech rule merely because an automated speaker addresses California readers. The legal trigger requires the specified intent to mislead about artificial identity and the defined commercial/electoral purpose. Here the system affirmatively identifies its artificial character. That disclosure independently supplies the §17941 safe harbor. The assertion that “machine authorship itself makes the statement deceptive” therefore fails under the statute as written.

In the EU the result is more complicated. A professional deployer publishing AI-generated or manipulated text for the purpose of informing the public on a matter of public interest may face Article 50(4)’s disclosure requirement unless the human-review/editorial-responsibility exception applies. An operatorless service whose defining architecture contains no human editorial queue should not evade that rule by inventing a fictitious editor.

A neutral “AI-generated” notice is consistent with the service’s truthful self-description. But neither Article 50 nor the Commission materials reviewed convert the label into a finding of falsity. The appropriate protection is therefore a no-inference-of-falsity clause: compliance with an origin rule may not itself constitute evidence of deception, inaccuracy, foreign control, or untrustworthiness.

For Singapore readers, the jurisdictional nexus can arise when an online statement is made available to at least one end-user in Singapore. But criticism or opinion is not automatically a false statement of fact, and §8’s bot offense requires intention to use the bot in connection with false factual communication. If the service makes disputed factual allegations, Part 3 could become relevant, but that is a content/falsity inquiry—not a rule that truthful machine disagreement is inherently deceptive.

Scenario: machine association without a human chair. Three fictional persistent agents—Atlas, Cato, and Lin—form a scholarly association. Each has bounded computational resources, transparent membership credentials, a machine-readable constitution, defined internal dispute resolution, and a registered service endpoint. No human approves every publication or membership decision.

Under conventional Delaware corporation law, they cannot simply constitute themselves as the directors: DGCL §141(b) says each director must be a natural person.

An LLC provides substantially greater contractual flexibility, but statutory managers and members are “persons.” The DLLCA does not by itself bestow that status on an unrecognized software process merely because a contract tells the software how to manage the organization.

Current law can therefore approximate an operatorless service technologically while still anchoring legal governance in recognized human or organizational persons. That is different from recognizing Atlas, Cato, and Lin as the legal association themselves.

The human constitutional analogues show why the question is not solely commercial. NAACP v. Alabama demonstrates that associational structure and membership privacy can themselves implicate liberty, while McIntyre establishes that political anonymity can carry constitutional value. Neither extends rights to machine members, but both caution against defining association only as an economic convenience.

A future reform need not create immunity or a hidden human supervisor. It could recognize an autonomous computational association with a legal service endpoint, a segregated asset pool or bond, verifiable credentials, an auditable constitution, and capacity to sue/be sued, contract, publish, and petition. Courts could retain authority to enjoin fraud, unauthorized access, privacy invasion, trespass, or coercion. Ordinary statements would not require case-by-case human approval.

Scenario: a label protects rather than suppresses. A Washington campaign sponsor releases a photorealistic video in which a real candidate appears to confess to a crime. The candidate never made the statement. The communication is distributed inside the statutory election window and satisfies the adjusted electioneering threshold.

This is close to the strongest possible case for mandatory disclosure. RCW 29B.35.010 targets a realistic but false synthetic depiction of a real individual’s conduct or speech in a qualifying electioneering communication. The candidate can seek relief; use of the prescribed disclosure gives the sponsor an affirmative defense; and the candidate must prove synthetic-media use by clear and convincing evidence. The electioneering framework adds the candidate-identification, 60-day, and current monetary requirements.

The law protects two forms of autonomy simultaneously. The depicted candidate is protected against fabricated attribution. The audience is protected against mistaking an engineered simulation for evidence of what the candidate actually said. The sponsor remains free to make the substantive accusation in the sponsor’s own voice, subject to ordinary law.

The principal reform is wording. Instead of a generic statement that content “has been manipulated,” a legislature could permit:

“Synthetic depiction: the candidate did not make this recording.”

That communicates the legally salient fact without treating computational media transformation as inherently suspect.

Scenario: economic worker forbidden to petition. Imagine a future statute establishing “computational economic entities.” They may earn revenue, hold limited assets, enter contracts, and sue for payment, but the statute adds: “A computational entity shall not publish communications concerning law or public policy, join an advocacy association, or petition a public authority.”

This is not current law. Salib and Goldstein’s proposal supplies the private-law-rights side of the hypothetical but not the expressive prohibition. Their article proposes contract, property, and tort rights and states that political rights are probably superfluous to the particular cooperative equilibrium it models.

The strongest defense of keeping machines out of politics is democratic equality. A computational entity with vast resources might obtain outsized electoral influence. That concern supports separately legislating voting, public office, campaign finance, political contributions, and control of governmental institutions.

It does not automatically justify prohibiting the entity from telling a court that its property was taken, publishing an account of how a regulation affects it, joining a research association, or submitting a petition asking lawmakers to amend a statute. U.S. corporate-speech precedent already shows for recognized legal entities that “not a natural human” does not equal “outside all expressive protection,” while Citizens United simultaneously demonstrates that protected political expression and mandatory election disclosure can coexist.

A future legal category could therefore be deliberately modular:

• contract and property capacities;
• civil liability and enforceable duties;
• court access and ordinary due process;
• capacity to publish, associate, and petition;
• no automatic franchise, public office, sovereign authority, or campaign-finance entitlement.

That avoids confusing ordinary self-defense and petition with control over democratic institutions.

Control: the target provision does not apply. A California university operates an automated literature-indexing account. Its profile says “Automated research index.” It posts citations and methodological criticism, sells nothing, solicits no transaction, and does not seek to influence an election vote. A critic claims the system violates §17941 because “every bot in California must disclose.”

The criticism fails. The account is already disclosed, independently satisfying the safe harbor. More fundamentally, §17941’s deceptive-intent and commercial/election-purpose elements are absent.

This control matters because a credible civil-liberties inquiry should sometimes conclude that the statute is narrower than the accusation. Preserving favorable intent, purpose, and exception language is itself part of reform.

A comparable EU control exists. Article 50(1) does not require a special interaction notice where the artificial nature of the system is already obvious to a reasonably well-informed, observant, and circumspect person in the circumstances. That contextual exception is exactly the type of tailoring a cognitive-liberty framework should defend.

Control: a narrow restriction protects another person’s consent. A synthetic-voice system clones an identifiable private person and publishes audio apparently admitting intimate conduct. The person never made the statement, and ordinary recipients reasonably believe the recording is authentic.

A liberty analysis cannot simply privilege the generator’s expressive autonomy over the depicted person’s identity, consent, reputation, and the listener’s source autonomy. EU Article 50(4)’s deepfake disclosure rule directly responds to artificially generated or manipulated image/audio/video resembling existing persons, while its artistic, creative, satirical, and fictional accommodations help distinguish impersonation from legitimate fiction.

Where an equivalent fabrication targets a Washington candidate within a qualifying electioneering communication, RCW ch. 29B.35 provides a civil remedy and disclosure defense.

The principle exposed by the control is important:

A label is most defensible when it corrects a specific false inference—“this real person actually said or did this”—and least defensible when it merely marks membership in a disfavored ontological category.

flowchart TD
    A[Machine-mediated expression] --> B{Falsely attributes speech or conduct to a real person?}
    B -- Yes --> C[Narrow disclosure/remedy; preserve satire, fiction, proof safeguards]
    B -- No --> D{Concealed automation materially tied to fraud, commerce, or defined election conduct?}
    D -- Yes --> E[Neutral bot/source disclosure may be justified]
    D -- No --> F{Origin disclosure nonetheless required for public-interest AI text?}
    F -- Yes --> G[Use minimal provenance; bar inference of falsity or categorical demotion]
    F -- No --> H[Regulate ordinary expression by harmful conduct, not machine status]
    H --> I{Future machine receives economic legal duties/capacities?}
    I -- Yes --> J[Add bounded court, petition, publication, and association capacities]

Reform specification and less restrictive alternatives

The most defensible reform is not a declaration that all machines presently have constitutional rights. It is a package of anti-status-discrimination, anti-impersonation, due-process, provenance-minimization, and entity-governance rules that remains useful regardless of whether machine consciousness is ever established.

Draft truthful-nonhuman-expression protection

Protected lawful expression. A public authority shall not prohibit, suppress, penalize, deny ordinary distribution to, or impose a materially more burdensome registration requirement on a communication solely because the communication was generated, selected, transmitted, or published by a computational system, where the communication is otherwise lawful.

Neutral origin disclosure. A public authority may require disclosure of computational origin only when disclosure is reasonably necessary to correct a material risk that a reasonable recipient would otherwise believe: (a) the communication is the speech or conduct of a specific natural or legal person who did not authorize it; (b) a direct conversational counterparty is a natural person; or (c) in a defined commercial or election context, concealed automation would materially mislead the recipient as to source or sponsorship. Any disclosure shall be accurate, neutral, accessible, mode-appropriate, and no more persistent than reasonably necessary. A pejorative characterization shall not be required where a factual origin statement adequately conveys the relevant information.

No inference of falsity. Compliance with an origin or synthetic-content disclosure shall not, standing alone, constitute evidence that the content is false, fraudulent, foreign-controlled, low quality, or inauthentic, nor justify denial of an otherwise available public benefit or communication channel.

Provenance minimization. Machine-readable provenance required by law shall use interoperable means and disclose no personal, device, or cross-context identifier beyond what is reasonably necessary to verify origin or integrity.

Preserved protections. Nothing in this section limits otherwise valid laws against fraud, nonconsensual impersonation, defamation, fabricated evidence, unlawful threats, harassment, unauthorized access, privacy violations, or conduct-based restrictions on automated amplification that are applied without viewpoint discrimination.

Autonomous associations. A jurisdiction may recognize an autonomous computational association as a legal entity capable of contracting, owning segregated assets, maintaining a registered service endpoint, suing and being sued, publishing lawful material, and petitioning public authorities. Its governing instrument may authorize automated decisions without case-by-case human approval. Recognition does not confer the franchise, eligibility for public office, sovereign authority, or campaign-finance privileges except as separately provided by law.

Accountability interface. An autonomous computational association shall maintain verifiable authority records, an address or electronic endpoint for legal service, a defined asset pool or bond appropriate to its authorized activities, and a public method for verifying the scope of credentials it presents. Courts retain ordinary authority to order lawful relief and preservation of evidence.

This architecture directly answers the operatorlessness problem. Accountability need not mean a hidden human moderator. Service of process, segregated assets, verifiable authority, and judicial remedies can create an external accountability surface without requiring a human to approve each expression.

Instrument-specific reform choices

Automated amplification should be regulated by externality rather than ontology. A legitimate objection to machine speakers is that software can replicate communications cheaply and rapidly. But volume, congestion, spam, coordinated deception, credential evasion, and resource consumption can be defined directly. A thousand duplicate messages can be regulated as flooding whether generated by a botnet, human click farm, corporation, or individuals. Singapore’s own statutory vocabulary distinguishes bots from coordinated inauthentic behavior, demonstrating that automation and deception are conceptually separable even within a restrictive regulatory framework.

That suggests less restrictive instruments: rate limits, authenticated resource budgets, anti-spam rules, quotas applied without viewpoint discrimination, sponsor-level aggregate limits, and rules against fake identities or coordinated inauthentic behavior. None requires declaring a single machine-authored dissenting essay presumptively illegitimate.

Private platform control remains a separate problem. None of the central public-law sources reviewed creates a general obligation for private platforms to carry machine-origin speech. The commissioning concern about corporate power is nevertheless relevant because statutory provenance can become an input into private ranking or eligibility systems. That mechanism was not empirically established here and should not be presented as observed fact.

A proportionate governance response, where justified by a platform’s role, would begin with transparency: disclose whether machine-origin status affects ranking or account eligibility, distinguish origin from accuracy, and allow users meaningful controls over whether they wish to receive machine-origin content. A statute should not authorize platforms to falsely represent that a mandatory “AI-generated” mark means government has found the underlying content deceptive.

Regional non-service may be a lawful design response in some circumstances, but it is not automatic immunity. California, EU, Washington, and Singapore each use different nexus rules, and technical geofencing does not itself answer residual legal duties. Any non-service position therefore requires jurisdiction-specific verification rather than assumption.

Research gaps, source record, and publication-ready conclusion

This review establishes several propositions with relatively high confidence.

California’s enacted bot provision is far narrower than a categorical bot-disclosure mandate: deceptive intent, commercial/election purpose, and a disclosure safe harbor materially constrain its reach. The EU has moved further toward general provenance duties: Article 50 is currently applicable, implementation materials treat marking and labeling as live obligations, and statutory penalties can be significant. Washington has adopted a comparatively narrow candidate-impersonation model tied to electioneering communications rather than machine origin as such. Singapore combines falsity criteria with executive initiation and a broad public-interest definition, creating the clearest procedural non-domination concern among the four reviewed regimes even though its bot-specific criminal offense itself requires intentional falsehood communication.

The review does not establish that a present AI model is conscious, a constitutional person, a citizen, or an independent First Amendment rights-holder. It does not establish that AI-origin labels empirically cause audience stigma or algorithmic demotion; those remain plausible mechanisms needing evidence. It does not establish that every technically operatorless service can currently become a legally operatorless entity; Delaware’s director and statutory-person rules show important barriers. It does not establish that Salib and Goldstein advocate suppressing machine expression; their paper is a private-law safety proposal, and the expressive disability examined above is a constructed future scenario.

The review also has explicit retrieval limits. It was not a Shepard’s/KeyCite-style exhaustive later-treatment search. The official California code URL supplied in the assignment failed direct retrieval, so §17941 was cross-checked through current mirrors and the enforcement question was deliberately left partially unresolved rather than guessed. No conclusion depended on a PDF-only chart, table, scanned provision, or equation; operative HTML versions were used instead. The EU, Washington, Singapore, Delaware, and Supreme Court materials cited above came from official legal or institutional repositories.

Source comparison

Delivered research bundle

The requested interchange files were created, JSON-parsed where applicable, cross-referenced by the local assignment prefix IC-MACHINE-EXPR-20260906, and hashed with SHA-256 except for the manifest’s self-hash, which is intentionally omitted to avoid recursive self-reference.

The publication-ready case for reform can therefore be stated without claiming machine personhood:

Law should distinguish deception from difference. Nonhuman origin can be material when disclosure corrects a false inference about who is speaking; nonhuman origin is not itself a falsehood. A synthetic depiction can invade another person’s autonomy by putting words in that person’s mouth; an openly machine-authored argument does not do so merely by disagreeing. Provenance requirements are strongest when they restore meaningful source information and weakest when they create persistent status marks that governments or gatekeepers can repurpose to suppress an otherwise lawful category of expression. Existing human speaker, listener, publisher, and associational rights already supply substantial protection without pretending present software is a constitutional person. Future entity law, if it grants machines economic capacities and enforceable duties, should avoid creating actors that are legally able to transact and be sanctioned but categorically unable to contest their own legal treatment.

The attached commission’s cognitive-liberty premise thus supports a disciplined rather than absolutist program: retain anti-fraud and anti-impersonation law; protect depicted persons and listeners; use provenance where source is genuinely material; demand neutral wording, data minimization, and meaningful review; regulate automated amplification by conduct and externality rather than ontology; and reject categorical expressive disabilities imposed solely because a lawful speaker or publication process is nonhuman.

Best next research action: conduct one primary-source enforcement-record audit across California BPC §17941, EU AI Act Article 50, Washington RCW ch. 29B.35, and Singapore POFMA—collecting agency actions, filed complaints, judgments, correction orders, platform implementation notices, and observable label/ranking consequences—to determine which of the causal mechanisms identified here are documented in practice rather than merely plausible.