Picture a honeybee hovering at a lavender stem, computing, in roughly one cubic millimeter of nervous tissue, the color of the flower, the value of its nectar against a hive full of competing alternatives, and the safest angle of approach while a wind gust tilts the stem. Now open a browser tab and type a question into ChatGPT. Both systems produce behavior that looks, at least on the surface, deliberate and contextually sensitive. For most of the history of cognitive science, that behavioral surface was the test. It is no longer. New studies suggest consciousness cannot be judged solely by behavior, whether it is a chatbot discussing philosophy or a bee searching for nectar; researchers are increasingly focusing on the internal mechanisms of brains and computers, concluding that today's AI is likely not conscious while leaving open the possibility for both conscious insects and future machines. That sentence, arriving in the summer of 2026, is the kind of sentence that quietly reorganizes a field.
The crack in the behavioral consensus was widened by two papers landing within months of each other. The first, published in Philosophical Transactions of the Royal Society B by Colin Klein of the Australian National University and Andrew Barron of Macquarie University, proposes what they call phenomenal interface theory, a computational account of minimal or basal consciousness in insects. They argue phenomenal consciousness is a consequence of how mobile animals with spatial senses and a capacity for goal-directed behaviour resolve the complex problem of action selection: to adjudicate between possible goals, an animal must use sensory inputs, representations of internal state, and stored knowledge of values to estimate expected value vectors for different options, and brains solve this problem by taking such heterogeneous information and transforming it into a common framework, a phenomenal interface, and then using this to compute multi-objective values. The bee, on this account, does not merely react; it runs something that functions as a first-person perspective. A consequence of this type of processing is that it naturally generates a distinction between self and non-self and a first-person perspective in which external stimuli have a subjective value. The second paper, in Trends in Cognitive Sciences and authored by a remarkable 20-person consortium including Patrick Butlin, Robert Long, Turing Award winner Yoshua Bengio, philosopher David Chalmers, and Jonathan Birch, takes the same architectural turn for AI. The methodology is called the theory-derived indicator method: rather than committing to any single theory of consciousness, including integrated information theory, global workspace theory, recurrent processing theory, higher-order theories, predictive processing, or attention schema theory, the paper surveys all of them and derives indicator properties from each. Importantly, the useful indicators are all structural.
The logical consequence of both papers is simultaneously unsettling and clarifying. What matters for consciousness is not what you do, but how you do it; the Trends in Cognitive Sciences paper looks to the machinery rather than the behavior of AI. That shift rescues the bee and demotes the chatbot. Five years ago, a seemingly ironclad test of whether something was conscious was to see if you could have a conversation with it; philosopher Susan Schneider suggested that if we had an AI that convincingly mused on the metaphysics of consciousness, it may well be conscious. ChatGPT does exactly that, every day, for millions of users. And yet the machinery underneath, a transformer architecture processing token sequences without the recurrent sensory-motor integration loops the new frameworks treat as essential, appears to fail on the structural indicators. The authors do not identify the computation itself, acknowledging there is science yet to be done, but they show that if you could identify it, you would have a level playing field to compare humans, invertebrates, and computers. That level playing field is the genuinely novel contribution: a single framework that does not privilege carbon, does not dismiss silicon, and does not let verbal fluency smuggle in a verdict.
Skeptics are not hiding. Researchers disagree about theories of consciousness, and some doubt whether conventional hardware can support consciousness at all. Others point out that every major theory of consciousness remains empirically underdetermined, meaning that deriving indicators from theories and then checking AI architectures against those indicators inherits all the uncertainty of the parent theories. There are risks of both under- and over-attribution of consciousness to AI systems, entailing a need for methods to assess whether current or future AI systems are likely to be conscious. The precautionary logic, however, is starting to carry institutional weight. Even if we cannot be sure something is conscious, we might err on the side of caution by assuming it is, what philosopher Jonathan Birch calls the precautionary principle for sentience. Anthropic has already hired an AI welfare officer. The question is migrating from seminar rooms to HR departments and regulatory offices at a speed the underlying science has not yet matched.
All of this lands at a particularly charged institutional moment. The Science of Consciousness conference, which ran in Tucson under University of Arizona hosting for nearly three decades, will reconstitute itself as Consciousness Science 2026 at the Paradise Point Resort in San Diego from October 11 to 16. The original April 2026 Tucson meeting was cancelled after conference director Stuart Hameroff and several affiliated speakers were named in the Epstein files. The October meeting in San Diego will convene approximately 600 to 700 participants for five and a half days of workshops, plenary lectures, concurrent sessions, posters, exhibits, and interdisciplinary dialogue. Confirmed content tracks include Roger Penrose, Susan Schneider, and a broad interdisciplinary programme spanning neuroscience, philosophy, quantum biology, bioelectricity, and machine consciousness. Alongside it, the Models of Consciousness 7 conference in Copenhagen runs October 12 through 16, overlapping with the San Diego meeting by four days and serving partially overlapping communities, with the Copenhagen event more focused on formal and computational models while CS26 includes empirical neuroscience and philosophy on roughly equal footing with theoretical work. The answer to who and what might be conscious, as of mid-2026, appears to be: fewer systems than we assumed when we were looking at outputs, and possibly more systems than we assumed when we were looking at anatomy.
The field has spent decades asking what consciousness looks like from the outside; it is now being forced, by bees and language models simultaneously, to answer what it must look like from within.