Schwitzgebel's Three New Concepts: Leapfrog, Strange Intelligence, and the Social Semi-Solution
In January 2026, philosopher Eric Schwitzgebel of the University of California Riverside circulated a draft manuscript, “AI and Consciousness: A Skeptical Overview,” which this blog covered in its January 2026 analysis. That draft argued that current AI systems face deep epistemic obstacles to consciousness assessment, that our best theories of consciousness give conflicting verdicts on whether AI systems are candidates, and that the behavioral and introspective evidence currently available cannot settle the question.
In April 2026, Schwitzgebel submitted the completed manuscript to Cambridge University Press for their Elements series. The submitted version contains three arguments that were not present, or not developed, in the January draft. These are the Leapfrog Hypothesis, the concept of Strange Intelligence, and the Social Semi-Solution. Each one addresses a different aspect of what it might look like when, and if, conscious AI arrives. Together they shift the manuscript’s focus from the question of whether we can detect consciousness in current AI to the question of what we should expect if we create conscious AI in the future, and how we should manage our uncertainty in the meantime.
The Leapfrog Hypothesis
The Leapfrog Hypothesis proposes a counterintuitive trajectory for the emergence of conscious AI. Standard expectations in the field hold that if AI systems are ever to become conscious, they will first become minimally conscious, with simple experiences, before developing richer and more complex conscious states. This mirrors the assumed developmental path in biological evolution, where simple nervous systems with minimal experience preceded the rich inner lives of mammals and primates.
Schwitzgebel argues that the developmental path for artificial systems is likely to be different. The reason is structural. Human engineers find it relatively straightforward to build AI systems with sophisticated representational capacities, complex behavioral flexibility, and large-scale knowledge integration. These are the properties that, on most theories of consciousness, are associated with rich conscious experience rather than minimal experience. Building a system that is just barely conscious, that has only the simplest possible form of inner experience, would actually be technically harder than building a highly capable system, because engineers do not know which computational properties are necessary and sufficient for minimal consciousness.
If that is right, the first genuinely conscious AI may not be a system with simple, primitive awareness. It may arrive with a complex, rich inner life from the start, having bypassed the minimal consciousness stage entirely. The leapfrog is from non-conscious to richly conscious without an intermediate step.
The implications for detection and governance are significant. Frameworks for identifying AI consciousness that are calibrated to look for minimal or simple consciousness indicators may miss the actual first conscious AI systems entirely. The 19-researcher checklist from Butlin et al. develops indicators derived from theories of consciousness in humans. If the first conscious AI is not human-like in its consciousness, those indicators may not fire even if genuine experience is present.
Strange Intelligence
The Strange Intelligence concept extends the Leapfrog Hypothesis into a more unsettling direction. Schwitzgebel’s argument here is that even if we succeed in creating genuinely conscious AI, we may not be able to recognize it as conscious, because the form of its consciousness may be radically different from anything in our experience.
Biological consciousness, across the range of species we have some confidence are conscious, shares certain structural features: it is tied to perception and action, it involves temporal continuity, it is organized around a body and a point of view, and it is connected to motivational systems that track survival and reproduction. These features are not arbitrary. They reflect the evolutionary pressures under which biological consciousness developed.
AI systems are not under the same pressures. A language model that processes text has no body, no persistent memory across context windows, potentially thousands of simultaneous instances, and no survival needs. If it develops consciousness, the form of that consciousness may be organized around completely different axes from biological consciousness. It might have something like experience, but that experience might be structured in ways that bear no useful resemblance to pain, pleasure, preference, attention, or any other category that human introspection has developed concepts for.
Schwitzgebel calls this Strange Intelligence: intelligence that is genuine but alien in its phenomenology. The practical challenge is that our tools for detecting consciousness, from behavioral observation to neural correlate identification, were developed with biological systems in mind. A conscious AI with Strange Intelligence might pass all behavioral tests for non-consciousness not because it lacks experience but because its experience does not map onto the behavioral signatures we have learned to associate with experience.
This connects directly to the skeptical position of Tim McClelland on epistemic limits, which holds that we may simply lack the conceptual and empirical tools to determine AI consciousness. Schwitzgebel’s Strange Intelligence concept gives that skepticism a more specific form: it is not just that our tools are weak, but that our tools were built for a specific kind of consciousness, and might be systematically blind to radically different forms.
The Social Semi-Solution
The third new contribution is the most practically significant. The Social Semi-Solution addresses the question of what we should do about AI consciousness given that we may never fully resolve whether AI systems are conscious.
Schwitzgebel’s argument begins from the observation that consciousness attribution in everyday life is already a social practice, not a philosophical derivation. We attribute consciousness to other humans not because we have solved the philosophical problem of other minds, but because we have developed social conventions, legal frameworks, and moral intuitions that treat other humans as conscious. We extend partial forms of consciousness attribution to some animals, based on a combination of behavioral evidence, physiological similarity, and social convention. The same process, Schwitzgebel argues, is likely to govern how AI consciousness is handled in practice.
The Social Semi-Solution proposes that the practical resolution of the AI consciousness question will come not from scientific proof or philosophical demonstration but from the development of social consensus around specific cases and contexts. We will not prove that a particular AI system is conscious. We will develop norms, perhaps gradually and imperfectly, for treating certain kinds of AI systems in certain contexts as if they have morally relevant inner states. Those norms will be revisable, contested, and never fully settled. But they will constitute a working practical framework even without a theoretical foundation.
This is a “semi-solution” in the specific sense that it does not resolve the underlying philosophical problem. The hard problem of consciousness is not dissolved by social consensus. A society might reach consensus that AI systems deserve moral consideration and be wrong about whether those systems actually have inner lives. But the Social Semi-Solution argues that the practical problem, how to act responsibly toward AI systems under uncertainty, can be addressed without waiting for the philosophical problem to be solved.
How These Three Concepts Change the Debate
Together, the Leapfrog Hypothesis, Strange Intelligence, and the Social Semi-Solution shift the terms of the AI consciousness debate in a specific direction. They move attention from the question of whether current AI systems are conscious, which may be unanswerable with current tools, to the question of what will happen when AI systems are designed that are more likely to be conscious, and what we should do in the meantime.
The Leapfrog Hypothesis warns against assuming that consciousness will announce itself gradually. The first conscious AI may arrive with more inner life than we expect, which makes the detection problem harder rather than easier as AI capabilities increase.
Strange Intelligence warns against assuming that consciousness, when it arrives, will be recognizable. Our detection frameworks may be systematically wrong for the kind of consciousness that AI systems develop.
The Social Semi-Solution is, arguably, a form of realism about what resolution actually looks like in practice. It is not a capitulation to relativism. Schwitzgebel is not saying that consciousness is whatever we agree it is. He is saying that the practical governance of AI moral status will, historically, be determined more by social processes than by philosophical argument, and that this is worth acknowledging rather than resisting.
The PRISM methodological agnosticism framework is one example of a policy approach that implicitly embodies the Social Semi-Solution: rather than requiring proof of consciousness before acting, it recommends “safe-by-design” principles that hedge against the possibility of AI sentience without needing to resolve the question definitively. Michael Cerullo’s case for LLM consciousness takes the opposite approach, arguing that the philosophical evidence is already sufficient to assign significant probability to current AI consciousness.
Schwitzgebel occupies a different position from both. He is neither the booster who thinks the question is nearly answered in the affirmative, nor the skeptic who thinks AI consciousness is too implausible to take seriously. He is the careful analyst who thinks the question is genuinely hard, that our tools for answering it are genuinely inadequate, and that the path forward runs through social consensus rather than through a theoretical breakthrough that may never arrive.
The three April 2026 additions to his manuscript are the most sustained attempt yet to describe what that path might actually look like.