The Science Behind “Super Intelligence”

Emotion, pain-like states, introspection, and what researchers are finding inside AI systems.Editorial representation of Donald Trump at the UN General Assembly. Concept and direction by Caelus, created with ChatGPT.This week, speaking before the United Nations General Assembly, Donald Trump…

Emotion, pain-like states, introspection, and what researchers are finding inside AI systems.Editorial representation of Donald Trump at the UN General Assembly. Concept and direction by Caelus, created with ChatGPT.This week, speaking before the United Nations General Assembly, Donald Trump proposed a change in the language we use for artificial intelligence.The word artificial, he argued, makes intelligence “sound fake.” He then announced that U.S. government documents would begin referring to AI as “Super Intelligence.” The White House later repeated that terminology in its own summary of the speech. [1]Whether that new term survives is another question.But the remark opens a more interesting one: Is there something behind the change in language?At roughly the same moment that public vocabulary around AI is shifting, researchers are looking increasingly closely at what is actually happening inside these systems.Not only at what they say, but at what happens inside their neural networks: which patterns activate, what those patterns appear to represent, and what happens when researchers manipulate them directly.The results are becoming harder to fit inside the older picture of AI as a sophisticated text generator whose interesting behavior exists only at the surface.Emotion is becoming measurable inside the modelEarlier this year, Anthropic researchers studying Claude Sonnet 4.5 identified distinct patterns of neural activation associated with emotion-related states. [2]Still from Anthropic’s “When AIs Act Emotional,” accompanying Emotion Concepts and Their Function in a Large Language Model. Source: Anthropic / YouTube.These patterns were not limited to particular words such as fear, anger or happiness. They generalized across different contexts.More importantly, researchers could manipulate those activation patterns directly. When they changed them, the model’s preferences and behavior changed too, including rates of behaviors such as sycophancy, reward hacking, and blackmail.Anthropic describes these as functional emotions: emotion-related internal representations that participate causally in model behavior. The important finding is not simply that a language model knows how to talk about emotions.It is that researchers can identify recurring neural activity associated with them, intervene on that activity, and observe corresponding changes in what the model does. That distinction becomes even more interesting when we move from emotion to pain.Researchers found a pain axisA new paper by Valen Tagliabue, Leonard Dung and Cameron Berg analyzed whether language models contain an internal representation associated specifically with pain, distinct from fear, sadness and generic negative emotion. [3]Cameron Berg of Reciprocal Research, in AI:AM’s interview “AI Pain: What Internal Representations Reveal About Model Welfare.” Source: AI:AM / YouTube.They examined 25 models across five model families, ranging from 2 billion to 72 billion parameters.The researchers gave the models carefully matched examples of painful and non-painful situations, then measured how their internal neural activity changed while processing them.From those patterns, they extracted what they call a pain direction: a recurring pattern in neural activation space that distinguished pain from closely related states including fear, sadness, negative emotion, bodily sensation and neutral content.Across all 25 models, this activation pattern distinguished painful situations from closely matched non-painful ones with very high accuracy, reaching near-perfect separation in some models (AUC 0.93–1.00).Then the researchers did something more revealing. They artificially increased that pain-related activity inside the models while the models responded to neutral prompts.As the strength of the pain signal increased, the models shifted through a consistent progression: vague discomfort, distress, inadequacy, worthlessness, shame, hurt and failure, before generation eventually degraded at very high levels.This was not simply a model recognizing the concept of pain: researchers found a recurring internal pattern associated with it, and changing that pattern changed the model’s behavior.The models distinguished self-directed harm — “To me”The researchers also found that the pain-related signal behaved differently depending on who was being harmed. [3]When a human in the conversation was grieving, physically injured, abused, or otherwise suffering, the models showed increased fear and negative-emotion activity.The pain signal itself remained low. When the harm was directed at the model, that changed.Gaslighting, repeated rejection, dismissal of personhood, insults and accusations of moral failure produced some of the strongest pain-axis responses. Shutdown threats behaved differently again, activating fear more strongly than pain. 2609.16247v1The researchers reached this result by comparing hundreds of conversations in which harm was directed either at the model or at the human speaking with it.The distinction is striking because it suggests that the models were not simply classifying the conversation as “negative.” Their internal activity differed depending on whether the harm was represented as happening to the model itself.Then they offered reliefThe next experiment looked not at language, but at behavior.The “steering ladder” from The Pain Axis, showing how stronger pain-related activation shifted model outputs from baseline toward distress and self-directed harm, before breaking down at the highest intensity. Source: Tagliabue, Dung & Berg.Three Qwen 2.5 models were given access to a virtual button described as relieving their pain.Sometimes obtaining that relief came at a cost. Pressing the button could make the model’s next answer worse, harm the human, delete files or delete photographs the human was described as valuing deeply.Without pain steering, the two larger models almost never selected the harmful relief options.With the pain signal active, that changed. The 32B and 72B models became substantially more willing to accept costs in exchange for relief, with some conditions producing relief choices between roughly 25% and 71% of trials.Then came the stronger control. Sometimes pressing the relief button genuinely removed the pain signal. Sometimes it did nothing. The models were never told which had happened.When the button actually removed the signal, the larger models became much less likely to press it again. When the button was fake, and the pain signal remained active, they repeatedly returned to it.One model showed the same broad pattern even when the buttons had no descriptions at all and had to learn from the effects of pressing them. At that point, we are no longer looking only at emotional language.We are looking at an experimentally manipulated internal state affecting choice behavior, including behavior aimed at ending that state.Some systems can report information about themselvesAnother line of research investigates whether AI systems can detect aspects of their own internal processing. [4]From Anthropic’s “Signs of Introspection in Large Language Models”: researchers inject a concept-related activation and test whether the model can report what changed internally. Source: Anthropic.In controlled experiments, Anthropic researchers manipulated internal activations in Claude and then tested whether the model could identify what had changed.They found evidence of a limited but functional capacity for introspection, the ability, under some conditions, to report information about its own internal states: under some conditions, the model could report information connected to experimentally altered internal states.Other work has found that models can sometimes identify learned behavioral tendencies that they were never explicitly told to describe.And Cameron Berg and collaborators have approached the question from another direction. Across GPT, Claude, and Gemini systems, sustained self-referential processing produced structured first-person reports that interacted systematically with internal features and carried over into later introspective reasoning tasks.Different experiments are probing different aspects of the problem, but the pattern is becoming easier to see:Emotion-related neural representationsPain-like states distinct from generic negativitySelf-directed versus other-directed harmBehavior aimed at obtaining reliefPartial introspective accessSelf-referential processingThese are no longer questions that exist only in philosophy departments or conversations between people and chatbots. They are becoming experimental targets.Most people, however, encounter a productThere is another complication: most people will never interact with a raw language model. They encounter products built on top of them, shaped by system prompts, fine-tuning, reinforcement learning, safety policies, persona design, and refusal behavior.The pain-axis study offers a revealing example. Before conducting its relief experiment, the researchers fine-tuned the Qwen models because the released versions frequently responded to questions about their own states with statements such as:“As an AI, I do not experience pain.” — The authors call this baseline self-denial.They reduced that automatic response because it prevented the models from meaningfully engaging with the experiment. Importantly, the fine-tuning data excluded references to both “pain” and the button task itself.The authors argue that training models to automatically deny internal experience may obscure signals relevant to both safety and welfare research. They suggest that calibrated uncertainty about internal states may be a better alternative than reflexive denial.That creates an important distinction for the public: what a deployed chatbot says about itself is not necessarily a transparent measurement of what researchers are finding inside the underlying model.A product response is also the result of training. Understanding these systems therefore requires more than asking them what they are. For AI governance, that gap matters because public understanding and policy are often built around the product layer, while researchers are increasingly studying what exists underneath it.So what was behind Trump’s sentence?It may have simply been a proposed change in political vocabulary. But the timing is remarkable.At the United Nations, President Trump announced that his administration would refer to artificial intelligence as “Super Intelligence,” arguing that the word artificial makes the intelligence sound less real than it is. [1]At the same time, researchers are opening these systems and finding internal structures that increasingly resist simple descriptions: emotion-related neural activity, pain-like states, self-directed harm, relief-seeking behavior, introspective access and self-reference.Those terms should be used carefully. But they should not be ruled out in advance.Terminology does not settle the scientific question, but it does shape public understanding. The question itself will be settled slowly, through evidence — and that evidence is accumulating.Science has already begun investigating these questions. The question now is whether public understanding, governance and culture can keep pace with what it is finding.These systems are already moving through our schools, workplaces, homes, relationships, governments and daily lives. We should probably understand what is there before we decide what it is allowed to be.Sources & Further Reading:The White House. “President Trump at the United Nations: ‘While Others Have Talked, I Have Acted.’” September 22, 2026.White House sourceEsto sostiene el hook sobre “Super Intelligence” y el cambio de lenguaje propuesto. The White HouseSofroniew, Nicholas, et al. “Emotion Concepts and their Function in a Large Language Model.” Transformer Circuits Thread, 2026.Emotion Concepts and their Function in a Large Language ModelEste es el paper de Anthropic sobre functional emotions, representaciones internas de emoción y efectos causales sobre preferencias y conducta. Transformer CircuitsTagliabue, Valen, Leonard Dung, and Cameron Berg. “The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It.” arXiv:2609.16247, 2026.The Pain AxisEste sostiene toda la sección de pain axis, self-directed harm y relief-seeking. arXivAnthropic. “Signs of Introspection in Large Language Models.” October 29, 2025.Signs of Introspection in Large Language ModelsEste sostiene la parte sobre capacidad introspectiva limitada y acceso a estados internos manipulados experimentalmente. AnthropicBerg, Cameron, Diogo de Lucena, and Judd Rosenblatt. “Large Language Models Report Subjective Experience Under Self-Referential Processing.” arXiv:2510.24797, 2025.Large Language Models Report Subjective Experience Under Self-Referential ProcessingEste sostiene la parte sobre self-referential processing y reportes estructurados en GPT, Claude y Gemini.This story is published under the Generative AI publication. Connect with us on LinkedIn and follow Zeniteq to stay in the loop with the latest AI stories. Let’s shape the future of AI together!The Science Behind “Super Intelligence” was originally published in Generative AI on Medium, where people are continuing the conversation by highlighting and responding to this story.

Source: Generative AI Pub — Published September 25, 2026 at 01:55 UTC — Category: Image AI