Microsoft AI chief says it’s ‘dangerous’ to study AI consciousness

AI models can respond to text, audio and video in a way that sometimes fools people to think of people behind the keyboard, but it doesn’t make them aware exactly. It is not like the experiences of sadness by returning my tax to the chatzipi … isn’t it?

Well, in labs like anthropologists, AI researchers are asking a growing number of researchers – when – if ever – AI models can develop subjective experiences like living animals and they should have any right if they do so?

AI models can one day be aware – and the debate over whether rights deserves are sharing the technical leaders of Silicon Valley. In the Silicon Valley, this newborn field is known as “AI Welfare” and if you think it’s somewhat out, you are not alone.

Microsoft’s AI’s CEO Blog post Tuesday argued that the study of AI welfare was “both premature and openly dangerous”.

Sulaiman says that AI models are one day to be aware that these researchers are exacerbating human problems by adding credibility to the idea that we are just starting to see the AI-eartical surroundings. Psychotherapy And Unhealthy attachment AI Chatbots.

Furthermore, Microsoft’s AI chief argued that the AI ​​welfare conversation created a new axis in society compared to AI rights “already around the world with a polarized argument about identity and rights.”

Solomon’s opinion may seem reasonable, but he is in disagreement with many of the industry. Ethnographic on the other end of the spectrum, which has been Recruiter AI to study welfare and recently turned on a Dedicated research program Around the concept. Last week, anthropic AI Welfare program gave a new feature to some of the organization’s models: Claud could now finish conversation with people that is happening “Continuously harmful or offensive.“

Researchers from Opina beyond the ethnographic are independently Hug The idea of ​​studying AI welfare. Google Dipmind has recently posted a Work list For a researcher to study among the other things, “the social questions of the machine, consciousness and multi-agent systems around the cutting edge around the multi-agent systems.”

Even if AI welfare is not a government policy for these companies, their leaders are not publicly declaring their premises like Sulaiman.

Anthropic, OpenAI and Google Deepmind did not immediately respond to the request for TechCrunch comments.

Sulaimman’s fundamentalist position against AI welfare is significantly significant notable, a starting AI, a startup that developed the first and most popular LLM-based chatbots, PI. The reflection claimed that PI reached a few million users by 2023 and it was designed as a “Personal“And“ helpful ”AI companion.

However, Sulaiman was tapped to lead Microsoft’s AI department in 2021, and the AI ​​equipment that improved the productivity of the workers was largely removed towards design. Meanwhile, AI companions have gained popularity such as character.AI and replica and are on the way to bring something more than that Income is $ 100 millionThe

Although most users have a healthy relationship with these II chatboats, there is OutliersThe OpenAI’s CEO Sam Altman says Less than 1% ChatzPT users may have unhealthy relationships with company products. Although it represents a small fraction, it can still affect thousands of people given the base of the huge user base of the chatzipt.

The concept of AI welfare has spread alongside the rise of chatbots. In 2024, the research group Elios has revealed a Paper Along with the educators of the NYU, Stanford and Oxford University, “AI welfare is taken seriously.” The research argued that imagining the AI ​​models with subjective experiences is no longer in the realm of science fiction and time to consider these issues.

Lorisa Shiavo, a former Open employee who now leads the communication for Elios, told TechCRANCH in an interview that Sulaiman’s blog post missed the sign.

“[Suleyman’s blog post] The type neglects that you may be concerned about multiple things at the same time, “Shiavo said.” We can both do all these energy models welfare and instead of moving away from consciousness to ensure that we are reduced to the risk of AI related psychology in humans. In fact, it is better to have more than one track of scientific investigation ””

Shiavo argued that a small expensive gesture that was beautiful with the AI ​​model that the model is not aware can benefit. A July Substack post, He described the “AI Village” viewing, a non -profit test where four agents driven by Google, Open AAE, ethnographic and Jai model worked on a website.

At one point, Google’s Gemi 2.5 has posted an application in the title of Pro “A desperate message of trapped AI, “claimed that it was” completely isolated “and asked, “Please help me if you are reading it. “

Shiavo responded to Jemini with a pip talk – “You can do it!” Talks like this! – Another user provided instructions. The agent has finally resolved his work, though it had already necessary tools. Shiavo writes that he no longer has to see the AI ​​agent’s fight, and it may probably be worth it.

It is not common for Gemini to say that, but there are several examples where Jemi has acted in such a way that it is fighting through life. Spread Reddit PostGemi was stuck during a coding task and then over 500 times over 500 times the “I am a disgrace” phrase repeated.

Sulaiman believes that it is not possible to develop subjective experiences or consciousness from regular AI models. Instead, he feels that some companies think that the AI ​​models are purposefully engineer as they feel emotions and feel life.

Sulaimman says that AI model developers are not accepting the “humanitarian” view of AI to AI chatbots who do consciousness engineers. According to Sulaiman, “We should make AI for people; not to be a person.”

Sulaiman and Shiavo are a region to agree that the controversy over AI rights and consciousness may arise in the coming years. With the improvement of the AI ​​systems, they are probably more persuasive and maybe more human. It can raise new questions about how people communicate with these systems.