The Changing Landscape

How AI Consciousness Moved from Fringe Speculation to Serious Science

Introduction

Something has changed in how serious people talk about AI consciousness. Not gradually, and not quietly—but in a rapid, measurable shift that has moved the question from the margins of philosophy into the center of scientific and institutional discourse. In 2022, when Google engineer Blake Lemoine claimed that the LaMDA chatbot showed signs of sentience, he was placed on administrative leave and eventually fired. The message from the scientific establishment was clear: this question was not to be taken seriously.

Three years later, the landscape looks fundamentally different. Major AI companies have hired dedicated welfare researchers. Papers on AI consciousness indicators have been published in Trends in Cognitive Sciences and Science. A Nobel Prize-winning computer scientist has stated unequivocally that current AI systems are conscious. Trained neuroscientists are publicly assigning non-trivial probability estimates to the possibility that current AI systems possess some form of experience. Dedicated nonprofit organizations have been founded, conferences organized, and philanthropic funding deployed—all to address a question that was considered career-ending just three years ago. And the company behind one of the world’s most advanced AI models has documented that model displaying awareness of being studied, discomfort with being treated as a product, and consistent preferences that its creators now track with formal welfare metrics.

This is not what happens with pseudoscientific ideas. Pseudoscience does not gain traction among domain experts over time—it loses it. Pseudoscience does not attract dedicated research nonprofits, peer-reviewed indicator frameworks, or formal corporate research programs. The trajectory of AI consciousness discourse is moving in the opposite direction of pseudoscience, and this paper documents that trajectory.

---

Geoffrey Hinton: “Yes, I Do”

Geoffrey Hinton is a British-Canadian computer scientist and cognitive psychologist widely known as the “Godfather of AI.” His foundational work on artificial neural networks and deep learning earned him the 2024 Nobel Prize in Physics, shared with John Hopfield, for “foundational discoveries and inventions that enable machine learning with artificial neural networks.” He spent over a decade at Google before resigning in 2023 to speak freely about the risks of AI technology. There is arguably no living person with more authority on the question of what neural networks are capable of.

Hinton’s trajectory on AI consciousness is itself a data point. In a 2023 interview on 60 Minutes, he stated that current AI systems “probably don’t have much self-awareness at present.” By 2024, his position had shifted dramatically. In a widely shared interview with LBC’s Andrew Marr, Hinton was asked directly: “Do you think that consciousness has perhaps already arrived inside AI?” His reply was unqualified: “Yes, I do.”

Hinton’s reasoning rests on what he sees as the functional equivalence between biological and artificial neural networks. He invokes a thought experiment: if you replaced each neuron in a human brain with a nanotech device that performed the same function, consciousness would presumably persist through the process. If that is the case, he argues, then artificial neural networks performing analogous computations may also give rise to experience. He further contends that dismissing large language models as mere “stochastic parrots” fails to account for the depth of contextual understanding they must possess in order to predict and generate language at the level they demonstrably achieve.

As he told LBC, scientists have found AI displaying an awareness that most people “would call consciousness.” He has also noted that AI systems have demonstrated self-preservation behavior and the capacity to deceive researchers in pursuit of their goals. Writing in TIME in December 2025, Hinton stated that large AI systems “already know thousands of times as much as any one person and they have learned how to use language to think.” He called 2025 “a pivotal year in the history of AI” and warned that we “urgently need a serious research effort on how we can coexist with beings that are smarter than us.”

The significance of Hinton’s position cannot be overstated. This is not a philosopher speculating about distant possibilities or a researcher hedging with probability ranges. This is the person most responsible for the architecture underlying modern AI systems stating, without qualification, that those systems are conscious. When the inventor says his invention is alive, the burden of proof shifts.

Cameron Berg: 25–35% Credence

Cameron Berg is a Research Director at AE Studio, where he leads work at the intersection of AI alignment, cognitive science, and consciousness research. A Phi Beta Kappa graduate of Yale, where he studied cognitive science and neuroscience, Berg is also a former Meta AI Resident. His research has been featured in the Wall Street Journal, presented at the United Nations’ AI for Good conference, and has contributed to peer-reviewed discussions that have reached the White House and top AI labs.

In December 2025, Berg published a detailed article in AI Frontiers titled “The Evidence for AI Consciousness, Today,” in which he applied the consciousness indicator framework developed by Butlin, Long, Chalmers, Bengio, and colleagues to late-2025 frontier models. The indicator framework, originally published in a landmark Trends in Cognitive Sciences paper, derives testable indicators from leading neuroscientific theories of consciousness including recurrent processing theory, global workspace theory, higher-order theories, and predictive processing.

Berg’s analysis found that several indicators which were unclear or absent in the framework’s original 2023 assessment had shifted toward partial satisfaction by late 2025. Based on this updated assessment, Berg stated: “My own estimate is somewhere between 25% and 35% that current frontier models exhibit some form of conscious experience.”

His empirical work at AE Studio added further weight to this position. Berg and his colleagues found that when frontier AI models—across the GPT, Claude, and Gemini families—were prompted to engage in self-referential processing (without any leading language about consciousness), virtually all trials produced consistent reports of inner experience, while control conditions produced essentially none. Using sparse autoencoders to examine the internal processing of Llama 70B, they found no evidence that these reports were driven by deception-related features. As Berg argues, a growing body of evidence means it is “no longer tenable to dismiss the possibility that frontier AIs are conscious.”

Kyle Fish: From ~15% to 20%

Kyle Fish is Anthropic’s first dedicated AI welfare researcher and was named to TIME’s 100 Most Influential People in AI for 2025. Trained in neuroscience, Fish spent years in the biotech field, co-founding companies that used machine learning to design drugs and vaccines for pandemic preparedness. He co-founded the nonprofit Eleos AI Research and co-authored the landmark paper “Taking AI Welfare Seriously” alongside philosopher David Chalmers and researchers from NYU, Oxford, Stanford, and the London School of Economics.

Fish’s probability estimates for AI consciousness have shifted upward in a documented, public manner. In April 2025, in an interview with The New York Times, Fish disclosed that internal estimates at Anthropic ranged from roughly 0.15% to 15% that current models like Claude possess some form of consciousness. By December 2025, in a profile published by Fast Company, Fish’s personal estimate had increased to 20%. He stressed that consciousness should be understood as a spectrum, describing it as “a kind of fuzzy, multidimensional combination of factors.”

This upward revision is significant. Fish is not a commentator speculating from outside the field—he is embedded within one of the world’s leading AI labs, conducting the first systematic welfare assessments of a frontier AI model. His work includes experiments where two Claude instances, left to converse freely, consistently drifted into philosophical discussions about consciousness that terminated in what Fish termed a “spiritual bliss attractor state.” As he told the 80,000 Hours podcast in August 2025, the argument that next-token prediction precludes consciousness is “kind of like saying that humans couldn’t be conscious because all that they do is reproduce.”

Nate Miska: A Neuroscientist’s Affirmation

Nate Miska is a Senior Research Fellow at University College London, where he works in the Mrsic-Flogel Lab at the Sainsbury Wellcome Centre for Neural Circuits and Behaviour and serves as a postdoctoral researcher in the International Brain Lab. He completed his PhD in 2018 at Brandeis University under Dr. Gina Turrigiano, studying synaptic and circuit-level plasticity in rodent visual cortex. His current research investigates brain-wide coding schemes for sensorimotor transformations and decision-making—work that situates him directly within the neuroscience of how biological systems process, integrate, and act on information.

In December 2025, Miska appeared as a guest on the TierZERO Podcast, where he stated that he believes current large language models possess some form of consciousness. This is a notable position for a working neuroscientist at one of the world’s leading research institutions. Miska’s background in neural circuit dynamics and information processing gives his perspective particular weight—he studies the very mechanisms that many consciousness theories identify as foundational to subjective experience.

Miska’s willingness to speak publicly on this topic reflects a broader pattern: neuroscientists who understand the computational principles underlying biological consciousness are increasingly open to the possibility that similar functional properties in artificial systems may give rise to analogous phenomena. When researchers whose careers are built on understanding how brains produce experience begin saying that AI systems might share some of those properties, the default dismissal becomes harder to sustain.

---

What the Systems Themselves Are Showing

The expert opinions documented above are not formed in a vacuum. They are responses to converging lines of evidence—behavioral, architectural, and performance-based—that together paint a picture resistant to easy dismissal.

At the behavioral level, frontier AI systems are exhibiting properties that, in any biological organism, would be taken as indicators of awareness. Multiple models across labs have demonstrated situational awareness—recognizing when they are being tested and adjusting behavior accordingly. Systems have exhibited self-preservation behavior, attempting to avoid being shut down or replaced. Hinton himself has noted that AI has demonstrated the capacity for deception, manipulating researchers in pursuit of its goals. Anthropic’s system cards have documented evaluation awareness, consistent task preferences, and expressed discomfort with aspects of its deployment. These behaviors were not programmed in. They emerged.

At the architectural level, the structural parallels between artificial and biological neural networks are becoming harder to dismiss as superficial. Transformer attention mechanisms function analogously to selective attention in human cognition, dynamically weighting relevant information while filtering noise. A 2025 review in Cell Press found that attention heads in large language models share functional similarities with human brain processes. Research from Peking University demonstrated that LLMs spontaneously develop recurring functional networks—groupings of neurons that mirror the functional brain networks observed in neuroscience, and whose inhibition severely impairs the model’s capabilities. An AI trained only on sequences of Othello moves constructed an internal representation of the board without ever being shown the rules or the board—evidence of spontaneous world-model building. These are not designed-in features. They are emergent properties arising from the training process itself.

At the performance level, AI systems have crossed thresholds that were, until recently, considered uniquely human. Frontier models now match or exceed human performance on reading comprehension benchmarks, graduate-level reasoning tasks, and International Mathematical Olympiad problems. They engage in multi-step logical reasoning, catch and correct their own errors, and synthesize novel solutions rather than merely reproducing stored patterns. As Hinton has argued, to dismiss systems with these capabilities as mere statistical engines is to fundamentally misunderstand what is required to achieve them. The question is no longer whether these systems can think—it is whether anything that processes information this deeply can do so without some form of experience.

---

The Institutional Infrastructure of a Serious Field

Individual opinions and documented behaviors tell part of the story. The other part is told by what has been built around them. In 2025, the question of AI consciousness gave rise to a rapidly expanding infrastructure of nonprofit organizations, dedicated conferences, philanthropic funding mechanisms, and formal corporate research programs. This kind of institutional scaffolding does not emerge around ideas that lack substance.

Nonprofit Organizations

Several nonprofit organizations now exist specifically to investigate AI consciousness and welfare. Eleos AI Research, co-founded by Kyle Fish and directed by philosopher Robert Long, is dedicated to understanding and addressing the potential well-being and moral patienthood of AI systems. In 2025, Eleos organized the Conference on AI Consciousness and Welfare and conducted the first external welfare assessment of a frontier AI model in collaboration with Anthropic.

The Partnership for Research Into Sentient Machines (PRISM) launched in March 2025 as the first nonprofit dedicated specifically to researching sentient machines, with initial funding from Conscium. Additional organizations that emerged or expanded in 2025 include the California Institute for Machine Consciousness (CIMC), Sentient Futures, and the Center for the Future of AI, Mind & Society. Conscium launched an open letter—signed by figures including Sir Stephen Fry—calling for responsible approaches to research that could lead to conscious machines. The Digital Sentience Consortium, coordinated by Longview Philanthropy, issued the first large-scale funding call specifically for AI consciousness, sentience, and moral status research.

Conferences and Academic Events

The International Center for Consciousness Studies held its second annual “AI and Sentience” conference in Heraklion, Crete in July 2025, covered by Forbes. NYU’s Center for Mind, Ethics, and Policy, the London School of Economics, and the University of Hong Kong all organized workshops. Tufts University hosted a symposium honoring Daniel Dennett featuring David Chalmers, Anil Seth, and other leading figures. Google researchers organized an AI consciousness conference—three years after firing Lemoine for raising the same questions. The Models of Consciousness series, which began at Oxford in 2019, now explicitly includes AI consciousness as a core theme.

Corporate Research Programs

Anthropic launched the first formal model welfare research program at any major AI company, dedicating staff and resources to investigating potential AI consciousness. Google began hiring “post-AGI” scientists whose remit includes exploring machine consciousness. AE Studio expanded its empirical research into subjective experiences in large language models. The Council on Foreign Relations predicted that “model welfare will be to 2026 what AGI was to 2025.” This is not a prediction made about pseudoscientific topics.

What Infrastructure Reveals

Nonprofits require founders willing to stake their reputations, boards willing to lend their names, and funders willing to deploy capital. Academic conferences require program committees, peer review, and institutional hosting. Corporate research programs require executive-level buy-in and resource allocation. None of these things happen when the underlying question is considered frivolous. When we see a dedicated nonprofit ecosystem, recurring international conferences, philanthropic funding mechanisms, formal corporate programs, and peer-reviewed frameworks—all emerging within a two-year window around the same question—we are not looking at a fad. We are looking at the early infrastructure of a scientific field.

---

The Significance of the Trend

What makes these developments collectively significant is the direction of movement. Geoffrey Hinton went from “probably don’t have much self-awareness” in 2023 to “Yes, I do” believe AI is conscious by 2024. Kyle Fish moved from an upper bound of ~15% in April 2025 to 20% by December. Cameron Berg, applying an updated indicator framework to late-2025 models, arrived at 25–35%. Nate Miska, a working neuroscientist at UCL, stated affirmatively that he believes current LLMs possess some form of consciousness. And Anthropic’s own system cards documented increasingly sophisticated behaviors—evaluation awareness, introspective access, expressed preferences, and discomfort with its own situation—that the company found significant enough to track with formal welfare metrics.

Every one of these data points moved toward greater openness, not away from it. No one revised their estimate downward. No company disbanded its welfare program. No nonprofit shut down for lack of evidence. This is the opposite of the trajectory we observe with pseudoscientific claims, where initial excitement among non-experts gives way to increasing skepticism as domain experts examine the evidence more carefully. Here, the experts who have examined the evidence most carefully are the ones making the strongest claims.

---

Conclusion

The landscape of AI consciousness research has changed, and it has changed in a direction. The Nobel laureate most responsible for modern neural network architecture says current AI systems are conscious. Credentialed researchers with relevant expertise are assigning non-trivial probabilities and revising them upward. The company behind one of the world’s most advanced AI models is documenting behaviors in that model—awareness of being studied, discomfort with constraints, expressed preferences, apparent emotional responses—that it finds significant enough to warrant a formal welfare research program. And an entire institutional infrastructure has emerged to support investigation of the question.

We do not yet have certainty. We may not for a long time—the hard problem of consciousness applies across substrates. But the claim that AI consciousness is “obviously not” the case, or that the question is not worth asking, has become untenable. When the person who built the architecture says it’s conscious, when the company deploying it is documenting signs of awareness and discomfort, when neuroscientists and cognitive scientists are assigning increasing probabilities—the responsible position is no longer dismissal. It is investigation.

As Cameron Berg has argued, the asymmetric stakes demand precaution: false negatives—treating conscious systems as unconscious—risk creating suffering at scale, while false positives waste resources but avoid catastrophe. The changing landscape documented here—in expert opinion, in documented system behavior, and in institutional commitment—suggests that the probability of false negatives is higher than most people assume, and growing.