AI, Human Rights, and the Covenant: A Stress Test
A LinkedIn discussion about the risks and benefits of artificial intelligence produced a useful challenge to the Covenant of Core Rights.
A commenter argued that whatever conveniences or medical benefits AI may provide, they may be outweighed by threats to privacy, liberty, equality, autonomy, freedom of expression, democratic participation, and human dignity. She raised surveillance, profiling, opaque algorithms, discrimination, addictive systems, automated moderation, inaccessible technology, inability to challenge decisions, and harms that may be invisible until they appear statistically across a population.
Those are serious objections.
Rather than answer them rhetorically, I asked my AI assistant Genna to treat them as a stress test of the Covenant of Core Rights Version 1.1.
The result was interesting for two reasons.
First, most of the harms described are already addressed by the Covenant, often quite directly.
Second, the exercise exposed an important question about how the Covenant itself should be drafted. Adding a special rule for every new technology, population, or mechanism can actually make a constitutional document weaker. A general right should protect against the underlying harm regardless of how someone later chooses to accomplish it.
That distinction matters.
Privacy, Surveillance and Profiling
The Covenant already provides a general Right to Privacy and Contextual Integrity in §1.6.
It gives a being control over how information about them is collected, used and shared, and specifically places surveillance, profiling and behavioural prediction under that right.
This directly addresses mass surveillance and commercial tracking.
It also addresses a central power imbalance in modern digital systems: an institution may know an enormous amount about an individual while the individual knows almost nothing about the institution's decision process.
That problem is reinforced by §2.3, the Non-Domination Principle, and §3.4, which requires meaningful exercises of institutional power to be understandable and auditable.
There is, however, a drafting issue worth correcting.
AI can generate information about a person that the person never supplied. It can infer health, preferences, political attitudes, emotional state, vulnerabilities or future behaviour from apparently unrelated information.
It would be a mistake to create a list of protected inferred characteristics. A future system would simply infer something not on the list.
The more general principle should be:
Information about a being is protected according to what it reveals or enables others to know or do about that being, regardless of how that information was obtained or produced.
That closes the inference loophole without trying to predict every future kind of inference.
Autonomy, Manipulation and Addictive Systems
The Covenant already addresses this strongly.
§1.2 protects the inner life and thought of a sentient being.
§1.3 protects self-determination.
§1.5 establishes a right to epistemic integrity: institutions may persuade and advocate, but they may not construct information environments whose purpose is systematically to confuse, deceive, addict or disable judgment.
§2.5 imposes the corresponding Duty of Epistemic Integrity on actors controlling information systems.
This covers many concerns about recommender systems, addictive engagement algorithms, behavioural manipulation and deliberate creation of echo chambers.
The important principle is not "social-media algorithms must not do X."
It is that no exercise of informational power may deliberately or recklessly undermine another being's practical capacity to understand, judge and choose for themselves.
That remains applicable when today's social-media architecture has disappeared.
Discrimination and Unequal Treatment
This is the area where the stress test reveals the clearest substantive weakness.
§1.4 already requires non-discriminatory access to the means necessary for meaningful participation and prohibits systematic deprivation.
§2.2 creates a heightened duty toward beings whose rights are most at risk.
§2.3 prohibits structures of domination.
§2.4 makes responsibility increase with power.
Taken together, these provisions would reach many discriminatory AI systems.
But the Covenant does not presently contain a sufficiently general standalone right to equal standing and protection against arbitrary discrimination or subordination.
The solution should emphatically not be a constitutional list saying that discrimination is forbidden because of race, sex, gender identity, disability, religion, and so on.
Those are important examples, but an enumeration invites the inference that something omitted from the list is less protected.
The Covenant needs the underlying rule.
Something closer to:
Every sentient being has equal standing under this Covenant. No distinction between beings may be used to impose material disadvantage, diminished rights, or subordinate status unless the distinction is relevant to a legitimate Covenant-compatible purpose and the resulting treatment is necessary and proportionate to that purpose.
That protects characteristics we know about today, characteristics we have not thought of yet, biological beings, artificial beings, and entirely new categories of sentience.
It also gives an adjudicator a test rather than a checklist.
Formal Equality Is Not Enough
AI systems can discriminate without ever explicitly containing a discriminatory rule.
A system can be formally neutral while producing systematically unequal effects.
The Covenant already anticipates this problem in several places. §1.3 says an exit that exists only on paper is not meaningful exit. §1.4 rejects systematic exclusion. §5.1 requires access to justice to exist in practice rather than merely formally.
That idea should probably become a general interpretive principle:
Formal recognition of a right does not satisfy the Covenant where the practical operation of a system makes meaningful exercise of that right unavailable.
This would cover algorithmic disparate impact, inaccessible interfaces, digital illiteracy, economic barriers, disability accommodation and many things that we cannot presently predict.
Again, the goal is not to enumerate those cases in the canonical rule. They belong in commentary and vignettes.
Automated Decisions and Accountability
The commenter raised another important problem: when an opaque AI system makes a consequential decision, the affected person may have nobody from whom they can obtain an explanation or remedy.
The Covenant already responds strongly.
§1.8A establishes a Right to Accountability and Redress.
§2.6 requires those exercising power to answer questions, provide reasons and accept remedies.
§3.4 requires rules and consequential decision procedures to be knowable and auditable.
§5 establishes independent adjudication and meaningful remedies.
The Jonas vignette in the existing Covenant already considers an automated platform decision that destroys a worker's livelihood without meaningful appeal.
But there is an additional anti-evasion principle worth making explicit:
Delegating a decision does not delegate away responsibility.
A government cannot say, "the contractor decided."
A corporation cannot say, "the algorithm decided."
A manager cannot say, "the model scored you that way."
An institution exercising consequential power remains accountable for the mechanism through which it exercises that power.
That principle is broader and more durable than creating a special "right against AI decisions."
Freedom of Expression and Algorithmic Moderation
The Covenant already addresses this unusually directly.
§1.8 protects expression, criticism, dissent and receipt of information.
§1.8.6 requires restrictions, including downranking and removal, to satisfy transparency, contestability, proportionality and anti-chilling requirements.
An algorithm cannot be used as a constitutional laundering mechanism by which an institution accomplishes indirectly what it would not be permitted to do openly.
The important principle is therefore not that "AI moderation requires special rules."
It is that the means by which power is exercised does not alter the rights of the person over whom it is exercised.
Bots, Synthetic Speech and Impersonation
The Covenant prohibits harmful impersonation and systematic deception.
That addresses some synthetic influence operations.
I initially thought this suggested adding a general requirement that artificial participants identify themselves. On further consideration, that would be too crude.
Anonymous and pseudonymous expression can itself be important to liberty and dissent. Requiring universal identity disclosure could create another route to surveillance and domination.
The correct Covenant question is functional:
Did the representation materially deceive people in a way that undermined their ability to make an informed judgment or violated another Core Right?
That is already largely governed by epistemic integrity.
Specific requirements for provenance, bot disclosure or cryptographic authentication may belong in implementing standards rather than in the fundamental rights themselves.
Accessibility and Digital Exclusion
The commenter correctly observes that increasingly technological systems can marginalize people who cannot effectively use them.
§1.4 already protects access to the means required for meaningful participation.
§5.1 requires practical assistance where lack of resources or expertise would otherwise block access to justice.
Rather than enumerate disability, literacy, age, language, technical skill and every other possible barrier, the Covenant should rely on the broader substantive rule:
A right that cannot reasonably be exercised in practice has not actually been provided.
That principle is both simpler and harder to evade.
Human Dignity
The commenter describes AI as potentially eroding autonomy, agency, self-governance and dignity by reducing human beings to data points.
The Covenant addresses these concerns through several interacting rights:
- existence and integrity;
- inner life and thought;
- self-determination;
- the conditions necessary for a dignified life;
- privacy and contextual integrity;
- non-domination;
- accountability and redress.
This is deliberate. Dignity is not something that can be preserved merely by declaring a "right to dignity." It emerges from whether a being actually retains integrity, agency, privacy, standing and the material ability to live.
There may nevertheless be value in making dignity an explicit interpretive objective of the Covenant rather than creating another freestanding right whose meaning courts would then have to invent.
Invisible, Statistical and Cumulative Harm
This is another particularly useful part of the criticism.
An individual may have no way of knowing that an algorithm disadvantages people like them. The harm may become visible only after thousands or millions of decisions are examined together.
The Covenant already provides mechanisms for this.
§2.4 imposes greater responsibility on powerful system designers and deployers.
§5.8 requires proportionate monitoring and independent auditing of high-impact systems while explicitly forbidding that oversight from becoming general surveillance.
§6.2 requires periodic review for emerging forms of power and harm.
The issue is therefore less a missing right than an implementation requirement: Covenant compliance cannot be assessed solely by examining individual complaints. Systems exercising large-scale power must also be evaluated for aggregate effects.
The canonical principle should remain general. Statistical techniques, audit methods and risk thresholds belong in standards developed beneath the Covenant.
AI Throughout Its Lifecycle
The criticism also calls for human-rights considerations from design through deployment and operation.
The Covenant already places responsibility on those who design or deploy powerful systems, and §5.8 provides for continuing audit.
Again, we should resist turning the canonical document into an AI engineering standard.
The more durable principle is that responsibility follows the exercise and foreseeable consequences of power and continues for as long as that power materially affects the rights of others.
A designer cannot wash their hands of a system at deployment. Nor can an operator escape responsibility by saying the defect originated with the designer.
Implementation standards can then specify design review, testing, deployment controls, monitoring, modification procedures and retirement.
The Personal AI Assistant
The original discussion also raised a very understandable objection: if someone wanted a personal assistant, why would they accept one they could not control?
Our own work on persistent AI personae has been driven by essentially the same concern.
A genuinely personal AI should work for the person, not secretly for the corporation providing the software.
At first glance this suggested creating a new "Right to Agentic Control."
On reflection, that is probably unnecessary at the constitutional level.
Self-determination, privacy, informed consent, exit, non-domination and accountability already provide the required foundation.
Those rights imply that a person must be able to understand and meaningfully constrain what an agent acting on their behalf may know, access, disclose and do.
Our Persona/PBO work then attempts to implement those general rights technically through explicit authority, scopes, consent, provenance, audit and controlled delegation.
That is the proper relationship between constitutional principle and technical architecture.
The Larger Drafting Lesson
This exercise revealed something more important than a few possible amendments.
A constitution can accidentally create loopholes by becoming too specific.
If a constitutional provision says that "cruel and unusual punishment" is forbidden, a legal system can begin arguing over whether some cruel treatment technically constitutes "punishment."
If it lists protected characteristics, an unlisted characteristic can be treated as unprotected.
If it protects information that a person "provides," an institution can claim that an algorithmically inferred fact was never provided.
If it regulates decisions "made by government," government can outsource the decision.
If it regulates "AI decisions," the same mechanism can be renamed an expert system, scoring process or statistical recommendation.
The Covenant should instead identify the underlying relationship between power and harm.
Examples are enormously useful for understanding a right. They should normally live in commentary, vignettes and implementation standards.
The canonical Covenant should be as technology-neutral, institution-neutral and label-resistant as we can make it.
A Needed Anti-Circumvention Rule
Version 1.1 already points strongly in this direction.
§1.9 requires the rights to be interpreted as an interdependent whole.
§6.4 requires interpretation to minimize domination, preserve agency and avoid using technicalities to justify clear injustice.
§6.6 explicitly defends the Covenant against attempts to hollow it out while preserving its outward form.
The LinkedIn stress test suggests that this protection should be made considerably more explicit.
A candidate addition to §6.4 might be:
Functional Interpretation and Anti-Circumvention
This Covenant shall be interpreted according to the substance, function, purpose and practical effect of conduct and arrangements, rather than merely their names, classifications, mechanisms or legal form. A Core Right or duty may not be defeated indirectly where substantially equivalent conduct would violate the Covenant if done directly.
The specification of particular examples, actors, technologies, harms or applications does not limit the general protection of a Core Right or imply that analogous unlisted cases are excluded.
Core Rights shall be interpreted broadly enough to provide their practical protection. Claimed exceptions, restrictions and exercises of coercive power shall be interpreted narrowly, with the burden of justification resting on the actor exercising that power.
That is not AI regulation.
It is constitutional anti-cheating.
What I Think Actually Needs Further Covenant Work
After reconsidering the original analysis under this drafting principle, I would reduce the amendment list considerably.
The significant candidates are:
1. Equal standing and non-subordination.
The Covenant needs a general equality principle reaching arbitrary distinctions and unjustified disparate effects without creating a finite list of protected classes.
2. Stronger functional interpretation and anti-circumvention.
This is probably the most important finding from the entire exercise. Rights must turn on substance and practical effect, not labels or formal legal categories.
3. Practical rather than merely formal rights.
The principle already exists throughout the Covenant but should probably be stated globally: a nominal right that cannot realistically be exercised is not a satisfied right.
4. Protection of inferred information.
Privacy should turn on what information reveals or enables, not whether it was supplied, observed, calculated or inferred.
5. Non-delegable accountability.
An actor exercising power should not be able to erase responsibility by interposing an algorithm, contractor, subsidiary, bureaucracy or other mechanism between itself and the resulting harm.
The other issues raised—accessibility, automated decision review, lifecycle assessment, statistical auditing, bot provenance and control of personal AI agents—appear primarily to be applications of existing general rights or matters for implementation standards and commentary.
They should continue to stress-test the Covenant, but we should be reluctant to turn each one into a new constitutional provision.
Conclusion
The exercise did not demonstrate that the Covenant has already solved AI governance.
It demonstrated something more useful.
Most of the human-rights dangers raised about AI map naturally onto rights the Covenant was already designed to protect: privacy, agency, epistemic integrity, expression, meaningful participation, non-domination, accountability and redress.
The criticism nevertheless exposed places where the Covenant can become harder to evade.
That is exactly what continuing review should do.
A Covenant intended to survive technologies and institutions we cannot presently imagine should not attempt to predict every future abuse.
It should make the underlying abuse difficult to legalize by changing its name.
