Carta de Derechos de la IA: Protección de la dignidad humana y la autodeterminación en la era de la inteligencia artificial.

Declaración de Derechos

1. The right to communicate, be heard, and remain silent
Every person deserves a reliable means of expression suited to their abilities and circumstances, along with sufficient time and support to use it. Communication systems should accommodate unusual speech, pauses, fatigue, incomplete phrases, and changes in how someone controls an interface. They should also respect a person's decision to stop, withhold an answer, or say nothing at all.

Silence, delayed responses, and inconsistent input must never be treated as consent. Systems should make room for humor, anger, disagreement, ambiguity, and an individual's distinctive way of speaking, even when those forms of expression are less convenient to predict or process. 

​

2. The right to informed partnership and education
People deserve an understandable account of what their AI system does, what it cannot reliably do, and what may change as models or services are updated. This includes information about voice cloning, avatars, ambient recording, personal memory, data-sharing arrangements, and the roles played by different agents or service providers.

Education and support should continue after onboarding. Explanations, trials, training, and options to withdraw must be available through the person's own communication method, with enough time to make an informed decision. Important limitations should be explained when they matter, rather than buried in documents someone may be unable to navigate.

​

3. The right to usable control over AI and its agents
A person should be able to pause, redirect, correct, or stop an AI system before it speaks or acts
beyond their instructions. In multi-agent environments, they should know which agents are active, what each has been authorized to do, what information they exchange, and when the system is acting outside the immediate conversation. Where the system permits, individual agents should be adjustable or disabled without forcing the person to abandon all assistance.

Control has to work under real conditions. A button that requires precise movement after a person's control has deteriorated offers little protection, and a refusal that takes several minutes may arrive after the conversation has moved on. Providers should test whether people can exercise their choices reliably, with the time, effort, and physical access actually available to them. Actions with lasting consequences require authorization and a practical route to human intervention.

4. The right to authorship and authentic self-expression
A person retains authority over the meaning of messages communicated in their name, even when AI helps find words, completes a sentence, retrieves a familiar expression, or renders speech in a synthetic voice. People should be able to reject every suggested response, begin from a word or fragment of their own, correct an interpretation, and express something the machine did not anticipate. They may also choose to delegate routine language without surrendering control over more consequential speech.

Selecting a model-generated sentence does not establish why it was selected. Someone may genuinely prefer it, tolerate an approximation to stay in the conversation, or accept it because correcting it would consume too much energy. Systems should preserve those distinctions wherever they affect consent, learning, attribution, or the representation of personal beliefs. Fluency must never be used as a reason to erase a person's distinctive language or dissent.
 

​

5. The right to change without being overwritten
Personal AI systems learn from a person's history, but that history can include suggestions the same system previously supplied. Over time, accepted approximations may be mistaken for deeply held preferences, and repeated model choices may gradually narrow the language a person is offered. We call for safeguards against this form of authorship drift, especially where acceptance, speed, or user engagement becomes a signal for future personalization.

People should be able to review and correct inferred preferences, distinguish their own statements from machine-generated proposals, and change their mind without being held to an earlier version of themselves. Evaluation must look beyond a single satisfactory response to what happens across months of use, model updates, and changes in the person's ability to communicate.

​

6. The right to consent over voice, likeness, and personal identity
A person's voice, face, avatar, personal history, and recognizable manner of expression carry their identity. Consent to use a communication service should never be treated as blanket permission for research, model training, demonstrations, advertising, commercial reuse, impersonation, or future uses that have not been explained. Each materially different purpose needs an understandable choice, including the choice to refuse without losing access to an otherwise available essential service.

Permission should be recorded clearly, reviewed when circumstances change, and withdrawable to the extent technically and lawfully possible. When a representative is needed, their authority and its limits should be documented; a caregiver's preferences should not quietly replace the person's own recorded decisions.
 

​

7. The right to provenance without losing human authorship

People should be able to understand how a message was made and who authorized it. That requires answers to four separate questions: Where did the language come from? What did the person approve? How was it presented or spoken? Who was being represented, and on what authority?
 

A synthetic-audio watermark may establish that software generated the sound, while saying nothing about who chose the words or intended the message. The same distinction applies to translated text, AI editing, and messages delivered by avatars. Provenance systems should reveal the machine's contribution and the limits of what can be verified, without inviting others to discount someone's communication simply because technology carries their voice.

​

8. The right to accountable representation and recipient choice
When an AI agent contacts another person on someone's behalf, the recipient deserves to know who is responsible for the approach, what authority the agent has, and how any information provided will be used. A professional title, organizational signature, or human-sounding message should not imply that a person personally wrote or reviewed something when that has not occurred.

People using assistive AI also deserve verification methods that respect disability and privacy. They should not have to speak naturally, appear on video, or disclose a medical condition to establish that they authorized a message. Both sides of an AI-mediated exchange need a way to question its authority, correct the record, and reach an accountable human when necessary.

​

9. The right to privacy, data justice, and fair use

Communication histories can reveal relationships, emotions, medical circumstances, intellectual work, and details a person never intended to make public. People should know what information is collected, where it is stored, who can access it, how long it is retained, and whether it contributes to model development or commercial products. Data obtained for essential communication should not silently become a research resource or a transferable commercial asset.

We support meaningful rights to inspect, correct, export, restrict, and request deletion of personal information, alongside honest explanations of any technical or statutory limits. Consent for care or communication must remain separate from permission for research and secondary use. Communities whose language or speech data make AI development possible should participate in decisions about access, recognition, and the benefits of that work.

​

10. The right to accessibility, linguistic diversity, and cultural respect

AI must work for people who fall outside the assumptions built into mainstream products. That
includes those with dysarthric speech, changing motor ability, uncommon languages or dialects, and communication practices that do not fit a system's training data. Accessibility should be considered throughout design and testing, with alternative input methods and support that can adapt as needs change.

Systems should respect cultural and personal differences without imposing a single standard of fluent, agreeable, or supposedly normal expression. When recognition is uncertain, the system should say so and offer a usable route to correction. A model's inability to interpret an individual's speech provides no basis for judging that individual's intelligence, capacity, or credibility.

​

11. The right to co-creation, recognition, and fair participation

People who depend on emerging technologies should have meaningful roles in deciding what gets built, which questions are studied, and how success is measured. Their lived experience can shape scientific methods and engineering decisions, and participation should extend beyond product demonstrations or consultation after major choices have already been made.

Researchers and developers should agree clear terms for contributions, attribution, compensation where appropriate, and any intended commercial use of data or creative work. AI assistance should not erase the human author's or inventor's contribution. The ownership and licensing of particular innovations may depend on applicable law and agreements, which makes transparent terms especially important before work begins.

12. The right to continuity, portability, and exit

Someone whose ability to communicate depends on a particular device, speech model, or cloud service faces a serious risk when that service changes or disappears. Providers should plan for transition, give meaningful notice of foreseeable disruptions, and support access to personal records, settings, consent history, and other material that can lawfully and technically be transferred.
 

A person's voice and communication history should not become a commercial lever that prevents them from changing products or providers. We advocate interoperable systems, documented exit arrangements, and practical support when technology must be replaced, while being clear about the limits of transferring proprietary models or licensed voices.

​

13. The right to decide one's digital legacy

People should be able to decide what happens to their recorded voice, likeness, avatar, personal
model, and communication history if they lose the ability to communicate or after they die. A stored recording, a memorial playback, continued research use, and the generation of new sentences in someone's voice involve different kinds of permission. They should be presented as separate choices, made and recorded while the person can express them.

A person may appoint a custodian for defined purposes, including preservation, restricted access, or deletion. If instructions are missing or ambiguous, a system should not assume permission to generate new material in the person's name. Retention, access, deactivation, and deletion procedures should be consistent with the person's documented wishes and applicable obligations, with clear explanations of what can actually be carried out.

​

14. The right to challenge harm and demand accountability

People need an accessible way to challenge false attribution, unauthorized voice or likeness use, harmful model updates, incorrect transcriptions, improper data use, and actions taken by agents beyond their instructions. A complaint process is inadequate if a person cannot use it with the communication method available to them or cannot learn who is responsible for making a correction.

Developers and deployers should maintain traceable records appropriate to the risk, explain significant failures, and provide timely routes to correction and remedy. Independent scrutiny should include disabled people in positions of influence, with evaluations that examine long-term use, unequal performance, and the practical cost of refusal. The Foundation will distinguish the safeguards already available in its own systems from those still being developed or tested.

​

15. The right to make consequential decisions without AI coercion or misrepresentation

AI systems must take particular care when a person uses them to discuss medical treatment, continued ventilation, advance care plans, or medical assistance in dying where it is lawful. Physical dependence on an assistive device can make the cost of correcting a machine unusually high, while a clinician or caregiver may understandably look to that device for a clear answer. The system's wording, defaults, omissions, and suggestions can therefore influence what is heard as the person's choice.
 

A person whose eye movements have become inconsistent might appear to answer “yes” to a question about continued ventilation. If software converts that movement into a certain response, clinicians may hear consent that was never established. Likewise, a statement such as “I cannot bear this anymore” must not be silently rewritten as a request to withdraw treatment.

The same protections must respect the right of a person with decision-making capacity to ask difficult questions and receive appropriate information. A suicidal crisis calls for responsive human support and safeguards against harmful encouragement. An informed discussion with qualified clinicians about lawful treatment choices or end-of-life care requires a different response. Disability, unconventional speech, or the loss of a reliable motor channel cannot alone determine capacity.
 

High-stakes systems should preserve uncertainty, make assumptions visible, and support accessible confirmation wherever a contemporaneous decision can be established. When communication has become unreliable, previously documented instructions and properly authorized clinical decision-making processes require careful consideration. AI must not invent a current preference, treat a caregiver's inference as the patient's own words, or optimize for agreement when someone's life may depend on the answer.