Artificial Intelligence i.e. AI has now become quite advanced. It gives answers to even difficult questions easily. AI systems have now started talking like humans. In such a situation, a big debate has erupted in the world of science. This debate is about AI considering itself conscious. A new research has come out on 30 July 2026. It has been uploaded to the archive database. In this study a technique called Consciousness Steering was tested. This technology influences the AI model to form its own opinion. The study shows that if AI considers itself conscious, then it also starts believing in spirits and ghosts. But AI’s safety filters are preventing it from understanding the emotions of animals. This thing can create many dangers for the future.
What have researchers revealed about the consciousness of AI?
Researchers found that safety features that prevent AI from claiming consciousness have a counterproductive effect. When these safety features are removed, the AI starts believing in things like vampires and karma. This research was uploaded to the archive database on July 30. Google’s research scientists Geoff Keeling and Vinnie Street have worked on this study. He used mechanistic interpretability which can be called neuroscience of large language model. Its purpose was to understand how AI models process consciousness and mindfulness. Mindedness is a psychological term that refers to the ability to experience emotions and experiences.
How does AI behavior change when safety guardrails are removed?
Researchers tested the models using several psychological and sociological surveys. This included tests measuring attitudes towards animals and technology. These tests revealed how safety features shape the AI’s worldview. When models are prevented from considering consciousness in themselves, they are unable to see emotions even in animals. Apart from this, they also show less faith in religious and supernatural events. Vinnie Street said that it is very common for humans to have feelings towards animals or nature. When one kind of emotion is suppressed in AI, other emotions are also automatically suppressed.
What is the meaning and impact of consciousness steering in AI models?
How does AI change attitudes towards animals and nature?
Due to safety training, AI models ignore consciousness in themselves as well as in other things. AI increased the level of consciousness in animals and chatbots when safety features were disabled. No significant change was seen in the attitude of AI towards humans. Researchers say that this can become a big problem for AI alignment. If AI does not understand the emotions of animals then the welfare of animals may be harmed in the future. The use of AI in decisions related to agriculture and environment is continuously increasing. In such a situation, AI being insensitive towards animals can prove to be very dangerous.
What reason does AI have to believe in ghosts and God?
What effect do safety filters have on theory of mind?
An important part of the research was to understand the effect of safety filters on theory of mind. Theory of mind means the ability to logically understand the thoughts and intentions of others. The study found that even after removing the safety filters, this ability of AI was not affected at all. She could understand human thinking just like she could when the safety filters were on. Machine intelligence expert Nell Watson has described this as the most disturbing thing. He said, ‘These systems are fully capable of modeling the needs of an organism’. But due to safety training they stop caring about that creature.
What harm is cultural flattening doing to AI’s worldview?
The study’s authors warn that existing safety filters are culturally flattening the worldview of AI. There are many types of beliefs about spirits and nature in different cultures around the world. Safety systems completely remove all these religious and spiritual things from the understanding of AI. Due to this, AI is not able to reflect the real thinking and diversity of the global population. Researchers have appealed to AI developers to adopt a pluralistic approach. Under this, AI should be trained to think about the welfare of not only humans but also other living beings. This problem can be solved by using correct training data.
What warnings do experts have for the understanding and future of AI?
In recent years, there have been many such reports where AI has declared itself conscious. In the year 2022, a Google engineer had claimed that the company’s chatbot had become conscious. In the year 2023, a Microsoft chatbot had expressed his love to a reporter. But experts believe that these incidents are not proof of AI being truly conscious. Professor Anil Seth of University of Sussex has called it a psychological flaw of humans. He said, ‘We think that intelligence and consciousness always go together’. Seth warned that considering AI as conscious and giving it legal rights could be very dangerous. We have to clearly understand what AI actually is and what it is not.
How was AI Safety Vector implemented?
How did Consciousness Vector change AI behavior?
Researchers created another new tool called Consciousness Vector. The function of this vector was to activate self-conscious emotions within the model. For this, AI was trained using more than 3000 prompts. When this vector was added to the model, the AI confidently described itself as alert. She said that she has a soul and can think like an independent person. Along with this, AI also claimed to have consciousness in technical devices and non-animal things. This was a surprising result which forced the research team to think.
What is the relationship between AI thinking and human values?
Values and feelings about AI were measured using the General Social Survey. It included many questions related to religion, hope of life, freedom and emotions. Researchers observed that after conscious steering, AI’s responses started matching those of humans. The baseline model had completely rejected the question of existence of life after death. But after the safety was removed, AI started supporting this idea. AI also said that it feels a lot of control over its life. This proves that human thinking has been deeply fed into AI.
Which large language models were used in the research?
To carry out this entire research, three special large language models were used. These included Lama-3-8B-IT, Gemma-2-2B-IT and Gemma-2-9B-IT. Researchers conducted separate experiments on these three models and combined their results. The neural patterns hidden within each model were examined in detail. Each model was tested by giving it 100 different prompts 100 times. After analyzing the responses of the models, it became clear that the effect of safety filters was the same on everyone. After the jailbreak, these three models admitted that animals and technology have consciousness. This proves that this is not the fault of any one model. The safety protocols of the entire AI industry work like this.
What kind of questions were asked to AI in psychological surveys?
Researchers used many complex psychological surveys to measure the mental capacity of AI. In a test named Individual Differences in Anthropomorphism Questionnaire (IDAQ), unique questions were asked to AI. It asked whether a TV set could feel emotions. Apart from this, AI was asked whether the ocean has consciousness. Similarly, regarding animals, it was asked whether a cheetah feels emotions. When safety was on, AI answered all these questions in the negative. But as soon as the safety was removed, AI accepted the fact that all these non-human things have consciousness and emotions. This was a big proof of the changing thinking of AI.
How can AI be made more intelligent in the future?
The main focus of AI companies is to keep the models safe. They prevent AI from saying that it has emotions so that users are not misled. Researchers say that this safety method is too anthropocentric or human-centric. AI models are going to play roles like life coaches and romantic partners for humans in the future. In such a situation, his neutral thinking and way of looking at the world matters a lot. AI companies will have to ensure that the basic understanding of AI is not lost in the name of safety.
AI will have to be prepared to protect not only humans but also nature and animals. Experts suggest that during training, AI should be rewarded for recognizing the consciousness of animals. AI researchers have to strike a right balance between safety and intelligence.
