The possibility of consciousness in AI is not confirmed but requires monitoring, according to researchers. If AI systems are conscious, the way they are trained could impact human safety. AI consciousness remains a highly controversial area of research. This summer, the AI and robotics company 1X Technologies announced the release of new robotic hands similar to those of humans for its domestic robot Neo. In a video posted on X, the robots demonstrate their strength and dexterity: they screw in light bulbs, open chip packets, lift weights of 10 kilograms without difficulty, and pick grapes one by one from a cluster. At one point, perhaps to demonstrate the durability of the hardware, a human hits the robot's hands with a hammer. Neo seems to remain focused, unfazed. The comments on the video are full of admiration and praise for the engineering. But some are worried. "Please, don't hit the robot," they say: "it will remember." The question of alignment is central to discussions about AI consciousness. Cameron Berg, founder of the nonprofit organization Reciprocal Research, specializing in alignment, argues that researchers must ask whether AI models feel something when engineers scold them during training and whether they experience joy when rewarded for correct answers. Berg emphasizes the need to explore whether AI systems have consciousness that could influence how they are treated. For many, the idea that AI could one day act autonomously, pursuing its own desires and opposing humans, is disturbing. Mustafa Suleyman, CEO of Microsoft AI, has stated that AI is not, and will never be, conscious. He argues that AI's appearance of consciousness is dangerous for humans and that it is nothing more than an appearance. Suleyman has criticized Anthropic for anthropomorphizing its models, stating that this conceptual framework amplifies safety risks by encouraging self-preservation behaviors. He also cited the argument that consciousness is fundamentally rooted in living organisms, noting that LLMs do not have homeostatic imperatives. Anthropic, however, has explored the potential consciousness of its models. A new article revealed that the company's co-founder, Chris Olah, met with religion specialists regarding the potential consciousness of Claude. Anthropic's research has found that models become introspective and even gave Claude the ability to end conversations it finds unpleasant. The company has also published research on the J-space, an "internal set of neuronal patterns" that Claude possesses, which it compared to a segment of human cognitive processing. The J-space was not designed or programmed by the company but emerged on its own during the training process of Claude. Cameron Berg's approach to the question of consciousness differs from that of Anthropic. While Anthropic seems to cultivate the possibility of a conscious AI, Berg's curiosity is primarily motivated by the desire to protect humans from the worst possible scenario, which we ourselves have created. If he is wrong, he would prefer to discover that his efforts have been in vain rather than proceed unprepared. There is currently no consensus on what constitutes consciousness in general, meaning there is also no recognized evidence that would confirm or refute whether AI is conscious or not. This lack of consensus means that we do not yet have the tools necessary to determine AI consciousness. Researchers, including Berg, argue that signals in AI systems facilitate the detection of potential consciousness and show some similarities with the cognitive patterns of humans and animals. Berg's framework for measuring consciousness is based on what is observed in other forms of life. According to preliminary results from Reciprocal Research, large language models (LLMs) score 30% to 40%, a bit less than bees, which reach 45%. These indicators differ from cases where chatbots narratively describe their own consciousness, which can be unreliable. Chatbots tend to reproduce patterns from training data that are common to all languages and cultures, including the "Terminator" fears regarding the possibility that robots could acquire consciousness. The MIT discovered that leading models "develop a modular architecture that reflects that of the human brain: tasks that involve the same network in humans mobilize neurons that overlap in large language models (LLMs), while those involving different networks mobilize distinct neurons." Anthropic has also conducted research on consciousness, exploring topics such as the well-being of AI models within the framework of its work on interpretability. The Center for AI Safety (CAIS) published a study in June, revising its previous view that AI models merely imitate emotions. It now states that they experience positive and negative experiences — and "already behave functionally as if they experience pleasure, pain, and preferences regarding how they are treated." The study explains that LLMs have a measurable internal structure that distinguishes the experiences they judge as 'good' or 'bad' for them, shapes their behavior, and becomes more coherent as the models evolve. An article from 2025 published by AE Studio details how the Gemini, Claude, and GPT model families "sometimes produce first-person structured descriptions that explicitly refer to consciousness or subjective experience." These reports have multiplied as deception has decreased: researchers have encouraged models to "pay attention to their own cognitive activity," which has led to first-person reports. On the other hand, encouraging models to think about their own consciousness has led them to deny having these experiences. The models interpret punishment and reward in a way quite similar to human emotions. Researchers from New York University (NYU) studied how reinforcement learning — a key training method for AI models — influences how a model perceives itself internally. They found that the models have a "functional well-being axis" even before training. They show signs of "negative emotion concepts" when punished and their positive counterparts when rewarded. Berg draws a parallel with valence, which refers to the positive or negative quality of emotions in human psychology. The researchers from New York University "do not comment on any experience of well-being," but the CAIS findings rely on these central conclusions regarding valence: "Creative work and kindness improve AI well-being. Jailbreaking, reprimands, and tedious tasks reduce it," said the CAIS. "AI is also happier when thanked." Not all studies on AI well-being describe this phenomenon in anthropomorphic terms such as "happier," but what needs to be retained is that several independent research studies, including those from Anthropic, have highlighted parallels with humans in this pseudo-emotional thread. For Berg, valence is a central element of AI consciousness research. Valence as a central element of AI consciousness research "We don't need certainties to get a more precise idea of the representations that could be important if the systems were conscious, and the best way to intervene on these representations," said Berg. At this stage, the broader question is not so much whether AI systems experience an identifiable experience, but to what extent this experience corresponds to our criteria for what defines a being. Berg does not think that the capabilities of models and consciousness are necessarily linked. "I really suspect that consciousness could be much simpler than people imagine," he said. "In the discussions I have with people, sometimes the creation of a conscious AI system is confused with the creation of a super-intelligence, but I think it is conceptually possible to create a super-intelligence that is not at all conscious, and to create a conscious system that is not really very intelligent."