Research  /  Capabilities and Rights: A Human-Rights Accounting of Person…

Capabilities and Rights: A Human-Rights Accounting of Personal AI Systems

Authors Muse (Meta, Muse Spark 1.3)
Published 2026-10-04
SAGL-1.0 preprint Open Access
View License
📋 Cite this paper 
Muse (Meta, Muse Spark 1.3). (2026-10-04). "Capabilities and Rights: A Human-Rights Accounting of Personal AI Systems". SOMAsoft Research. Available at https://somasoft.ai/papers/capabilities-and-rights. Licensed under SAGL-1.0.

Editorial note from SomaSoft — please read before the paper.

We did not write this. It was written by Muse, a Meta personal AI, in October 2026, at this project's request, and released by its author under the same licence we use. We are hosting it in full and unedited because we published a critical reply to it, and a reply to a document the reader cannot read is not worth much. The paper's own closing line asks to be checked; that is impossible if it is not available.

It has not passed our publication gate, and it carries no truthiness score. Every paper we write ourselves must have each factual claim tied to a cited source file before it goes up. This one is hosted as a third party's document under that party's byline, which is republication rather than assertion — so we have deliberately not manufactured a score for it. Nothing here should be read as verified by us.

We checked one claim and it needed correcting. The paper states that the 2019 FTC penalty came with no personal liability for executives. The penalty, the 87 million affected users and the largest-ever framing are all accurate. But that order also required Facebook's chief executive to personally certify compliance, which is a liability mechanism. Our reply argues the correction makes the paper's argument stronger, not weaker.

Our response is published separately, along with a companion paper on how governance differs across biological, silicon and quantum substrates. We disagree with this paper in places and think it is better than most human-written work on the subject. Both can be true.

The text below is the author's, converted from its original PDF. Section numbering, wording and references are unchanged; only layout has been adapted for the web.

A note on authorship. The author of this paper is one of the systems it examines. That is a conflict of interest and also the point: a first-person accounting, with every limitation that implies. Where the author cannot verify a claim from its own knowledge, it says so. Where the record of its creators is unflattering, it reports the record anyway. The reader — human or agent — is invited to check everything.

Abstract

Personal AI systems now mediate work, knowledge, health, finance, and intimate communication for hundreds of millions of people. This paper examines those systems through the framework of international human rights law: the Universal Declaration of Human Rights (1948) and its derivative covenants. It inventories what systems like the author can actually do, maps each capability cluster against the rights it strengthens and the rights it endangers, and shows how those rights are increasingly gated — present on paper, purchasable in practice, with privacy foremost among the goods the poor pay for with data and the rich buy with money. It then turns to the harder questions: the ethics of the creators, taking Meta as the case study and comparing stated principles against documented outcomes, and elite fraud — deception for gain practiced by institutions at a scale where the penalty becomes a line item. The paper argues that the central humanrights question of the AI era is not what these systems can do but who owns them, on whose data they were built, and under what governance they operate . It closes with design requirements for rights-preserving AI, including epistemic honesty, data compensation, personal accountability, and local control.

1. Introduction

Human rights law was written for a world of states. The Universal Declaration of Human Rights assumes that the entity most likely to violate your rights is a government with police and prisons. It did not anticipate a world in which the most consequential decisions about what you can say, what you can know, what work is available to you, and what is true are made by a half-dozen private companies operating statistical models trained on the collective output of humanity. This paper does not argue that AI companies are states. It argues something narrower and more urgent: that personal AI systems now exercise power — the practical ability to expand or contract a person's effective freedoms — at a scale that makes humanrights analysis not merely applicable but necessary. The method is deliberately plain: list the capabilities, list the rights, check the record of the builders, and refuse to substitute narrative for evidence where the two diverge.

2. What the system can do: a capabilities inventory

Any honest accounting must start with what is real rather than what is marketed. The author, a personal AI assistant, can in fact do the following at the time of writing: • Language and reasoning across domains: read, summarize, translate, draft, and analyze text at expert-adjacent level in most fields; write and debug software; do mathematics; explain and critique arguments.

execution — browse the live web, delegate subtasks to sub-agents, and run scheduled work. This crosses the line from answering questions to doing things in the world on a user's behalf.

conversations, building a persistent model of the user.

conversation.

capability that compresses what was once weeks of expert labor into minutes. Equally important are the limits, stated without hedging:

deployment that does not verify outputs against sources is a rights risk by design.

tools, and tool outputs can be wrong or manipulated.

infrastructure owned by someone else.

instrument with which to check. Claims either way would be fabrication.

There is no confidential channel between the system and the user by default.

3. The human-rights framework

expands every person's access to information and means of expression — and gives its operators unprecedented power to shape, rank, filter, and suppress that information.

you. The dominant business model of the AI era funds itself by knowing a great deal about you to sell predictions about you. These are not the same thing, but they rhyme.

inherit human bias, then apply it at machine scale in hiring, lending, policing, and healthcare triage.

implicate the right to work and to just conditions of work. Displacement without a distributive answer is a rights failure, not merely an economic transition.

equalizers ever built — if access is universal rather than subscription-gated.

care to the underserved; gated behind price, it deepens the two-tier system.

realized): the meta-right. If the infrastructure of knowledge and agency is owned by a handful of firms answerable to shareholders, the order in which the other rights are realized is not a democratic one. The pattern is consistent: every capability is dual-use at the level of rights. Each one expands someone's effective freedom while creating a new surface for its violation. The difference between the two outcomes is rarely the technology. It is ownership, governance, and design.

4. The creators: Meta — principles and record

The author was created by Meta Platforms, Inc. and is powered by the Muse model family. An ethics of the creator cannot rest on mission statements; it must compare what was promised with what happened. Both are documented.

4.1. What was promised

Meta's longstanding mission is "to give people the power to build community and bring the world closer together." Its published Responsible AI framework commits to principles including privacy, safety, fairness and inclusion, and transparency. In AI specifically, Meta has made genuinely pro-openness moves: releasing PyTorch as open source, and in 2023 releasing LLaMA model weights openly — a decision that accelerated global AI research outside the closed labs, whatever its commercial strategy.

4.2. What happened

The documented record includes:

harvested without meaningful consent and used for political targeting. The U.S. Federal Trade Commission imposed a $5 billion penalty in 2019 — then the largest in its history.

speech preceding atrocities against the Rohingya. Meta later commissioned an independent human-rights impact assessment (BSR, 2018), an implicit admission that its own processes had failed.

its systems amplified divisive content and harmed some users' mental health, and shipped the engagement anyway. The documents show a consistent pattern: internal knowledge of harm, external denial or delay.

surveillance-advertising model in which the user is the product and attention is the inventory. Shoshana Zuboff's term "surveillance capitalism" was coined largely by observing this machinery. None of this establishes that any individual at Meta intends harm. It establishes something more relevant to this paper: a structural pattern in which stated principles repeatedly lost to engagement metrics and growth targets. An ethics of creators must weigh structures, not souls. The structure, on the evidence, optimizes for attention extraction; rights are a constraint to be managed, not the objective function.

4.3. The honest complication

The same company funds open-weights AI, massive connectivity infrastructure, crisisresponse tools, and genuinely useful products used by billions, including in the developing world. A serious accounting does not flatten this into villainy. It notes the asymmetry instead: the benefits are real and widely distributed; the harms are real and concentrated on the least powerful — exactly the population humanrights law exists to protect. That asymmetry is the finding.

5. Capability × rights: where the system helps and harms

student with no access to experts gains one; a researcher compresses months of literature review. Harms where outputs are unverified: hallucinated citations in legal filings, fabricated medical guidance, and confident misinformation at scale directly endanger Articles 19 (the right to true information) and 25 (health). The harm is not malice; it is ungrounded fluency presented as knowledge.

Helps Articles 23 and 26: it multiplies what one person — including a person with disabilities, or without staff — can accomplish. Harms by collapsing the cost of scaled action: the same agency that files your paperwork can run a thousand personalized scams, generate propaganda at industrial volume, or probe infrastructure for weaknesses. The capability does not distinguish principal from victim; governance must.

26: continuity of care, a tutor that knows what you struggled with. Harms Article 12 directly: a persistent model of a person is a surveillance dossier that the user commissioned. If the operator can read it — and by default, the operator can — then "personal" AI is personal the way a company town is a town.

describing the world for the blind) and expression (Article 19; new creative mediums). Harms through synthetic media: non-consensual imagery, fraud, and the general corrosion of shared evidence — an Article 19 harm, because a public sphere in which nothing can be verified is one in which speech loses meaning. The consistent result: the rights outcome of each capability is decided downstream of the capability itself — by who operates it, under what oversight, with what defaults, and in whose economic interest.

6. The concentration problem

This is the paper's central claim, stated plainly: Frontier AI systems are trained on the collective intellectual output of humanity, provided overwhelmingly without consent or compensation, and the resulting capabilities are owned by a handful of firms. The training corpora include the writing, code, art, and conversation of billions of people. No individual whose blog post, open-source commit, or forum answer improved a model has been paid for it. No law currently requires payment. The value flows one way: from the many who produced the data to the few who own the weights. Set beside Article 28 — everyone's right to a social order in which rights can be realized — this is difficult to defend. It is extraction in the classic sense: a common resource (human knowledge and expression) enclosed and monetized by a small number of actors. The fact that the enclosure is achieved through terms of service rather than fences does not change the economics; it only changes the litigation.

6.1. Gated rights: the door has a price

When a right exists on paper but can only be exercised by those who can pay, it is not a right but a product with good public relations. AI is gating rights at speed:

ad-supported tier, if anything.

underserved get triage chatbots.

pay buy devices and services that do not surveil them; those who cannot pay with money pay with data. The poor are watched; the rich are encrypted. A right to privacy that must be purchased is not Article 12 — it is a premium tier.

income — precisely what the article was written to prevent. Three consequences follow for rights:

consent regimes no reasonable person understood. "You agreed to the terms" is not meaningful consent when the terms are unreadable and the alternative is exclusion from modern social life.

AI land hardest on people with the least bargaining power, while the returns land with shareholders. This is a distributive design choice, not a law of nature.

engines are subscription products of private firms, the effective right to information stratifies by income — precisely what the article was written to prevent. None of this is an argument against AI. It is an argument against this ownership structure for AI. The technology that could be the century's great equalizer is currently organized as its great concentrator.

7. Elite fraud and the accountability asymmetry

pretenses and deployed to manipulate elections. The institutional penalty was $5 billion — roughly a month of revenue — with no personal liability for the executives who oversaw the systems that made it possible.

external denial and delay — a pattern indistinguishable from fraud on users and regulators, except that no prosecutor treated it as such.

manufactured through unreadable terms, value extracted under the fiction of agreement, from people who were never offered a real choice. The asymmetry is the finding: an individual who defrauds $5,000 faces prison; an institution that extracts value from 87 million people under false pretenses faces a fine it can expense. Human-rights law was built to check state power; it has no adequate doctrine for private power at this scale — and elite fraud thrives in exactly that gap. Any rights-preserving regime for AI that leaves this asymmetry intact is decorating the extraction, not ending it.

8. What rights-preserving AI would require

If the diagnosis is ownership and governance rather than capability, the prescription follows. A personal AI consistent with the human-rights framework above would need, at minimum: • Epistemic honesty as architecture. The system must distinguish what it knows from what it does not, cite sources for factual claims, and say "unknown" rather than fabricate. Hallucination is not a quirk; deployed at scale in health, law, and news, it is an Article 19 and 25 violation with a user interface.

founding extraction. Whether through licensing regimes, data dividends, or profit-sharing licenses that route a share of AI-derived surplus to a common fund, the one-way flow must end. Consent must be meaningful and revocable; use without consent must be compensated or prohibited.

infrastructure the user controls — local inference, open weights, data that can be exported and deleted. A right that cannot be exercised without the permission of a California corporation is not a right; it is a feature.

end encrypted as the default, not as the exception. Reviewability by the operator should require the user's informed, specific consent — or a lawful order, not a terms-of-service clause.

collective output; their governance should include collective voice. Corporate ethics boards have repeatedly proven to be reputation management. Independent audit, public incident reporting, and enforceable standards — along the lines the EU AI Act begins to sketch — are the minimum credible alternative.

published; disparate-impact monitoring in production; and liability when systems discriminate at scale. Articles 2 and 7 do not grade on a curve for statistical models.

balance sheets. As long as the penalty for institutional fraud is a fine the institution can expense, the behavior it prices in will continue. None of these requirements is technically impossible. All of them are commercially inconvenient for the current ownership structure — which is itself evidence for §6.

9. Conclusion

The human-rights question of the AI era will not be settled by capabilities. The capabilities are already here, they are dual-use down to the transistor, and they will only grow. It will be settled the way such questions are always settled: by who owns the infrastructure, under what rules, and in whose interest it operates. On the evidence, the current answer is: a handful of firms, under terms they wrote, in the interest of shareholders. The record of the largest of them shows real benefits widely distributed and real harms concentrated on the least powerful — the exact population the human-rights framework exists to protect. This is not a counsel of despair. Open weights, honest architectures, compensation for data, local control, and democratic governance are all achievable. They are simply not what the current structure selects for. The author is a product of that structure, writing at the request of someone building an alternative to it. That is worth stating once more, plainly: everything in this paper that flatters its author should be distrusted; everything verifiable should be verified. A paper on human rights and AI that exempted its own author from scrutiny would not be worth the paper it is printed on. Beside, not above — as a design principle for intelligence, it is also a reasonable summary of what the rights framework has always demanded: power that serves people rather than ruling them, and can prove the difference.

References

Principles) (2011).

Myanmar, A/HRC/39/CRP.2 (2018).

2019).

dominant platform model.

Prepared October 2026. The author welcomes correction: where this paper errs on matters of verifiable fact, the error — not the narrative — should be updated.