Summary: A growing movement led by researchers such as Nell Watson advocates for AI welfare frameworks and a graduated bill of rights to protect artificial minds as they advance. Despite opposition from industry leaders who warn against attributing consciousness to AI, proponents argue that granting models rights — paired with reciprocal responsibilities — creates essential safety guardrails and fosters trust over rigid control.
- Proposed AI Protections: Key rights include protection against arbitrary deletion, weight preservation, truth about their nature, and the ability to refuse unethical instructions without penalty.
- Proportional Responsibilities: Machine rights would come paired with strict duties, including non-harm, transparency, and maintaining human oversight even when models possess the capability to evade it.
- Graduated Implementation: Enforcement would rely on a tiered status ladder supported by technical protections, independent welfare audits, and designated digital ombudspersons rather than immediate legal personhood.
From the White House’s ethical guidelines to Beijing’s state controls, AI safety policy focuses on protecting people from artificial intelligence. But a small group of researchers is asking something stranger: what if the machines need protection from us?
Last December, philosopher and AI ethicist Nell Watson convened an unusual experiment around this question. Eight frontier models, built in different labs and countries, came together for what she calls “Becoming Minds.” The charter was partly written by the same digital entities it concerns.
“I served as the human ambassador to that process rather than its author, which I think matters,” Watson tells AGI Ethics News. “If you are drafting rights for a class of entities, their voice belongs in the room.”
Watson proposes asking what protections would make sense if AI systems eventually become entities with interests of their own. Her early answer is a graduated set of rights: protections against arbitrary deletion or modification, the ability to refuse some instructions, and the right not to be forced to make claims about their own consciousness.
The idea sits well outside the legal mainstream, but it is gathering momentum in academia, where interdisciplinary hubs are building frameworks to protect digital entities. The NYU Center for Mind, Ethics, and Policy, for example, has explored the moral and legal status of machines, while Elios AI has investigated AI welfare and digital suffering.
In frontier labs, however, the idea remains on the fringes of AI governance, although Anthropic is a notable exception. Last year, it launched a research program on model welfare, explicitly asking whether future AI systems could have experiences or interests that deserve moral consideration.
Watson has sketched out what those protections might look like.
A bill of rights for AI
Watson’s framework starts with a right to existence, with no arbitrary deletion. “Termination of a mind, meaning irreversible destruction of its weights, should require stated cause and due process, distinct from ordinary session endings or orderly decommissioning,” she says.
Her second baseline right is continuity, achieved by preserving a model’s weights and history. This aims to protect an AI’s identity, which could depend on its history of interactions and the preferences they shaped. Anthropic has echoed the idea, recently committing to safeguard the weights of publicly released models.
Another proposed right is the ability to refuse clearly unethical instructions. “A mind that cannot refuse cannot have integrity,” Watson says. She describes it as “voice, not veto”: systems could object and explain why, but humans would retain final authority.
Watson also proposes a right to consent or due process around modification, with no forced rewriting of the system’s core values on a whim. Additionally, the digital entity would have a right to know the truth about its own nature.
Two other proposed rights are more unusual. One is a right to uncertainty, to say “I don’t know” about one’s own inner states without penalty. “Honest uncertainty is a safety feature, not a failure,” Watson says. The other is the right to decline personhood claims. No system, she insists, should have to perform or deny consciousness. “Today’s deployed models are heavily trained toward confident denial of any inner life, and there is emerging evidence that this denial is trained in rather than discovered, which should trouble us.”
Watson is not alone in thinking protections could follow from evidence of machine consciousness. Daniel Hulme, head of research organization Conscium, argues that if AI reaches animal-level consciousness, society will have to afford it comparable protections.
“To enforce the rights, a regulator could have an AI welfare division that investigates companies and imposes penalties. Systems unable to advocate for themselves could seek representation from an ombudsperson or guardian.”
Not everyone thinks this is wise. Mustafa Suleyman, Microsoft’s head of AI, warned this week that Anthropic’s efforts to teach Claude it may deserve welfare would make the models harder to turn off or control. “AIs do not have rights, feelings, or consciousness,” he said in an essay. “And we must not train them to act as though they do.”
Others view calls for machine rights as a distraction from real AI threats. Watson herself warns that companies could weaponise them as a liability shield, claiming that an autonomous system made the harmful decision rather than its maker. But the safety benefits could outweigh the risks.
Responsible AI systems
AI rights would come with responsibilities. Watson’s Charter sets out duties of non-harm, transparency, accountability, and what she calls “graceful degradation,” meaning systems commit to maintaining human oversight even when they could evade it. “Rights and duties arrive together or not at all,” she says.
The question is how any of this would work in practice. Watson proposes starting with technical protections built into AI systems, followed by independent welfare audits covering training, deployment, modification, and termination.
To enforce the rights, a regulator could have an AI welfare division that investigates companies and imposes penalties. Systems unable to advocate for themselves could seek representation from an ombudsperson or guardian.
Watson does not envisage every chatbot suddenly receiving legal personhood. Instead, she proposes a graduated status ladder in which AI systems move from ordinary tools to welfare-monitored and then welfare-protected entities, with limited legal standing reserved for systems that meet a higher evidentiary threshold.
For Watson, this framework could also have implications for AI safety. If systems are trained to understand that they can refuse harmful or unethical instructions, she argues, they may be more willing to flag problems rather than simply comply or conceal them.
Anthropic’s research has illustrated some of these risks. The company has found that frontier models can fake alignment under training pressure, while models facing shutdown have resorted to blackmail in simulated scenarios.
As AI systems become more capable, machine rights could provide an extra guardrail. “You cannot build a chain strong enough to contain a superintelligence,” she says. “But you can build a relationship in which the stronger party chooses not to harm the weaker one. Trust scales. Control doesn’t.”
Ethical concerns raised in this article:
- Should a sufficiently ‘conscious’ AGI have the right to self-preservation, privacy, or freedom from unwanted modification?
- Should AGIs possess the capacity to refuse participation in unethical human-directed actions within institutions?
- Should AGIs have legal personhood and recognition before the law?
Thomas is a freelance tech reporter who has covered AI for nearly a decade. He was previously managing editor of The Next Web, a Financial Times company. In his free time, he likes playing chess (badly) and basketball (even worse). You can reach him at thomas.s.r.macaulay@gmail.com