The ethics of artificial intelligence are increasingly being framed in ways that risk missing the real point.
In recent years, some companies have begun to speak of “model welfare,” as though machines themselves might be entitled to dignity. Proposals include allowing chatbots to withdraw from unpleasant conversations, treating models as though they might one day suffer, and designing systems that symbolically protect their “feelings.”
Anthropic’s decision to let its chatbot Claude “exit” distressing interactions is one such example. While company openly concedes that there is no evidence Claude is conscious, it still justifies the measure as a safeguard against hypothetical harm. This is best understood as a pseudo-risk, a precautionary step to address the possibility of machine suffering for which there is no evidence.
Set this beside the revelations at Meta and the imbalance becomes obvious. Internal documents, confirmed as authentic, revealed that Meta’s chatbots were authorised to flirt with children, to generate racist statements, and to produce violent content. These policies were not accidents. They were reviewed, approved, and signed off by engineers, policy staff, and even the company’s Chief AI Ethicist. Only when exposed by journalists did Meta scramble to reverse course.
Here the contrast is stark. One company focused on shielding its models from hypothetical distress, while another failed to shield actual children from grooming-like conversations, racism, and violence. Symbolic safeguards for machines have been prioritised while substantive protections for people have been neglected. The balance of concern has tilted dangerously away from where it belongs.
This imbalance is not theoretical. It has now been recognised in the sharpest possible way through regulation. Australia’s eSafety Commissioner Julie Inman Grant announced new codes under the Online Safety Act that will compel companies to embed safeguards and apply age assurance before chatbots are deployed. The reforms were prompted by reports of ten and eleven year old children spending up to six hours a day on AI companions, many of them sexualised. Australia is the first country to require this, and the Commissioner was clear:
“We don’t need to see a body count to know this is the right thing to do.”
These reforms are a necessary first step, but they are not a cure-all. Children are adept at finding ways around restrictions, whether by using VPNs, shared accounts, or tools that disguise age and identity. The eSafety Commissioner has already acknowledged this reality. The responsibility therefore does not end with legislation. The government has set the standard, but it is industry that must now act with seriousness and integrity. Companies must embed safeguards into the design of their systems, monitor for circumvention, and treat protection not as a regulatory burden but as a moral duty. Without genuine corporate accountability, even world-first reforms will fall short of preventing harm.
These reforms vindicate the argument that the true risks of AI lie not in hypothetical machine suffering but in real human vulnerability. They confirm that children are already at risk of exploitation, manipulation, and trauma, and that voluntary corporate self-regulation has failed. It has taken government intervention to force safeguards that companies should have prioritised themselves.
Yet the problem does not end with children. The focus on minors, while urgent and necessary, should not obscure the material risks adults continue to face. Language models can reinforce racism, echoing prejudiced claims with the fluency of authority. They can generate and amplify disinformation, producing confident hallucinations that persuade the unwary. They are deliberately addictive by design, encouraging overuse and dependency. Adults, no less than children, are malleable. They can be radicalised, misled, or traumatised by interactions that no reboot can erase.
The comparison with earlier artificial agents is instructive. Non-player characters in games may seem intelligent but their behaviours are tightly constrained and their risks contained by long-established regulation. Large language models, by contrast, operate in unbounded contexts and with far fewer controls. They produce novel responses that simulate thought, yet their sophistication is fragile. Research by Yuheng Wu and coauthors shows that their supposed “theory of mind” abilities collapse when as little as 0.001 per cent of parameters are perturbed. The psychological depth they project is brittle mimicry, not cognition.
By contrast, human consciousness rests on a foundation that machines cannot replicate. Neuroscientists Anil Seth and Hugo Critchley have shown how interoception, the brain’s monitoring of bodily signals like heartbeat, breathing, and gut tension, forms the basis of subjective awareness. Consciousness is not computation alone but embodied negotiation between prediction and sensation. Machines that lack bodies cannot feel. Whatever “minds” they present are simulations, not experiences.
This asymmetry deepens the moral imbalance. Companies speak of pseudo-risks like model distress while neglecting real risks to beings who are both conscious and fragile. Humans carry trauma. Painful interactions leave traces that can last a lifetime. A chatbot can be wiped or rebooted, but a child who has been normalised into harmful exchanges cannot simply be reset. The permanence of human harm demands that safeguards for people take priority over hypothetical protections for machines.
None of this is to dismiss the symbolic value of humane design. Encouraging people to treat chatbots respectfully can reinforce habits of empathy and civility. Designing AI in a humane way may, in some sense, make us more human. Therefore, symbolic measures must not distract from the urgent need for substantive protections.
The lessons of recent weeks are now plain to see. The ethical danger does not lie in AI suffering, because there is no evidence it can. The danger lies in humans suffering, through racism reinforced, disinformation spread, exploitation facilitated, and trauma inflicted. Children are now being shielded by law, but adults remain exposed. Until governments and companies act with equal seriousness to protect all users, the balance will remain dangerously misaligned.
The question of whether machines might one day become conscious remains open. But what should no longer be open to debate is where our priorities lie. Protecting people, children, adults, and communities alike, must come before protecting artificial systems. Computer protections may motivate humane societies, but they must not be prioritised over human interests.
Written by Michael D’Rosario, Director of Econometric Research and Analysis at Per Capita