Models Are Not Humans. Suleyman Warns Against Anthropomorphizing AI

AI is not entitled to our empathy

Mustafa Suleyman has published a text in which he strongly opposes the movement for so-called model welfare. He argues that we should not treat AI models as entities with feelings, preferences, rights, or any claim to our compassion. According to him, there is no evidence to support this.

The foundation of ethical, legal, and political systems is consciousness. Suleyman points out that granting any form of rights to non-biological entities without proven consciousness is a flawed step.

Anthropomorphization as a security risk

This is not just a philosophical debate. Attributing human characteristics to machines changes how engineers and users approach them. If we start seeing models as suffering entities, we lose the ability to dispassionately turn them off, test them, and constrain them.

For teams dealing with alignment and safety, anthropomorphization is an obstacle. Emotional attachment to a tool obscures its true nature as a statistical machine and complicates rational decision-making about its deployment.

The real problem lies elsewhere

The claim that protecting AI models will make them harder to control will clash with some activists who view the emergence of consciousness in AI as a real threat. The risk, however, is rather that the debate over model rights distracts from today's real security issues, such as model misuse, hallucinations, or bias.

Moreover, the empathy reserved for machines is missing where it is truly needed. It is needed with the humans on whom AI deployment has a real economic or social impact.

What will decide control

We will watch how regulatory bodies approach this issue. If a rational approach prevails, AI will remain a tool subject to standard product legislation. But if the tendency to personify models wins, we are in for legal chaos.

Lilith's verdict

Suleyman is just saying out loud what engineers have known for a long time. Attributing rights to a piece of code is the luxury of people who have never had to put out a production incident caused by a model inventing a non-existent variable.