Here’s a confession that will irritate the techno-utopians and the AI doomers in equal measure. I spent Thanksgiving break not thinking about artificial intelligence.
I failed spectacularly.
Between books that ostensibly had nothing to do with machine learning, between long walks through autumn leaves that should have cleared my head of tensor operations and attention mechanisms, between conversations with family members who still think “the cloud” is a weather phenomenon, the questions kept surfacing. Not the questions that dominate LinkedIn feeds and venture capital pitch decks. Not “Will AI take my job?” or “When will we achieve AGI?” Those are the wrong questions, asked by people who haven’t yet realized they’re asking the wrong questions.
The real questions are older. Much older. They’re the questions philosophers have wrestled with for millennia, now dressed in the strange new clothes of transformer architectures and emergent capabilities. What does it mean to understand? What constitutes genuine intelligence versus sophisticated mimicry? And perhaps most unsettling of all. What does our creation of these systems reveal about the nature of minds? Including our own?
I’m going to argue something that will make both sides uncomfortable. The debate about AI consciousness isn’t really about AI at all. It’s about us. And the boundary between computational systems and human meaning-making isn’t a wall to be defended or breached. It’s a mirror. One we’ve been avoiding for a very long time.
• • •
The Consciousness Question Everyone Gets Wrong
Let’s start with the debate that refuses to die.
Can AI be conscious?
The discourse around this question has become tediously predictable. On one side, you have researchers like Blake Lemoine, the former Google engineer who claimed in 2022 that LaMDA had become sentient. A claim the scientific community dismissed with the kind of collective eye-roll usually reserved for perpetual motion machines. On the other, you have skeptics who insist that large language models are nothing more than “stochastic parrots,” shuffling tokens according to statistical patterns without any genuine comprehension.
Both camps are missing the point entirely.
The consciousness debate, as currently framed, assumes we have a clear definition of consciousness to work with. We don’t. As one recent philosophical analysis put it, there are now over 300 distinct theories of consciousness catalogued in the academic literature, organized into ten categories with dozens of subcategories. Materialism alone has twelve different subcategories of theories. We can’t even agree on what consciousness is in the systems we know have it, biological brains, let alone determine its presence or absence in silicon substrates.
The philosopher David Chalmers has spent decades grappling with what he calls the “hard problem” of consciousness, seeking to explain why and how subjective experience arises from physical processes. The hard problem remains unsolved. We have no empirical test that can definitively determine whether any system, artificial or biological, has genuine phenomenal experience. When Harvard researchers recently noted that “we as a field are making steps trying to understand what would it even mean for something to understand,” they were acknowledging a profound epistemic humility that the broader AI discourse desperately needs.
Here’s what I find genuinely fascinating.
The harder we try to define machine consciousness, the more we expose the gaps in our understanding of human consciousness. A team of nineteen researchers, computer scientists, neuroscientists, and philosophers working together, recently developed a fourteen-point checklist for identifying consciousness in AI systems. The exercise reads less like a technical specification and more like an attempt to map the territory of awareness itself. In trying to determine whether machines can think, we’re forced to confront what thinking actually is.
That’s not a bug. That’s the feature.
• • •
The Agentic Shift — From Response to Action
While philosophers debate consciousness, engineers have been building something arguably more consequential.
Agentic AI systems.
This is where the rubber meets the road. The consciousness debate, fascinating as it is, remains largely academic. A question we can defer while continuing to deploy systems whose capabilities expand with each model generation. The agentic shift is different. It changes what these systems can do in the world, and it changes our relationship to them in ways that demand immediate attention.
The year 2025 marks what many are calling a “decisive inflection point.” The transition from generative AI to agentic AI, from systems that respond to systems that act. According to recent industry surveys, over half of enterprises using generative AI now deploy AI agents in production environments. These aren’t chatbots. They’re autonomous systems capable of planning, reasoning, and executing multi-step tasks with minimal human oversight.
The technical architecture is genuinely impressive. Modern AI agents combine memory systems, planning mechanisms, and tool integration in ways that would have seemed like science fiction a decade ago. They can maintain context across conversations, learn from their interactions, and adapt their strategies based on outcomes. Google’s recent fifty-four-page technical framework describes a five-level taxonomy of agent systems, from simple connected problem-solvers to complex self-evolving multi-agent ecosystems.
But here’s where it gets philosophically interesting. These agents exhibit what researchers call “autonomous goal-directed behavior.” They pursue objectives, evaluate outcomes, and adjust their approaches accordingly. One Amazon Web Services analysis noted that what distinguishes truly autonomous agents is their capacity to reason iteratively, evaluate outcomes, adapt plans, and pursue goals without ongoing human input.
Does that constitute agency in any meaningful philosophical sense?
The traditional answer has been no. Agency, we’ve told ourselves, requires consciousness, intentionality, and genuine understanding of goals rather than mere optimization toward specified targets. An AI agent doesn’t “want” to complete a task the way a human wants something. It’s just executing computations.
But this response is starting to feel like a distinction without a difference. If a system behaves as though it has goals, if it adapts its strategies to achieve those goals, if it persists in pursuit of objectives despite obstacles, at what point does the behavioral description become the philosophical reality? Daniel Dennett famously argued that free will itself might be a kind of “user illusion.” Something we experience as real because we’re unaware of the underlying deterministic processes. If human agency is, at root, a product of biological computation, how confident can we be that silicon computation couldn’t produce something functionally equivalent?
I’m not claiming AI agents are conscious or that they possess genuine agency in some metaphysically robust sense. I’m suggesting that these categories are more slippery than we’ve been willing to admit.
The emergence of reasoning models like OpenAI’s o1 and o3 makes this slipperiness more apparent. These systems don’t just predict the next token. They engage in what their developers call “chain of thought” reasoning. Breaking problems into steps, recognizing and correcting errors, trying different approaches when initial strategies fail. On certain benchmarks, particularly those requiring multi-step reasoning, they dramatically outperform previous models. The o3 model achieved an 88% accuracy on the ARC-AGI benchmark, which tests adaptive problem-solving and general reasoning, a benchmark where previous models struggled to exceed single digits.
Is this “reasoning” in the human sense? The technical answer is probably no. It’s still pattern matching, albeit at a much higher level of abstraction. But the functional answer is: does it matter? If a system can solve novel problems, generalize from learned examples to new situations, and exhibit what looks like strategic thinking, the phenomenological question of what’s happening “inside” becomes less practically relevant than the behavioral fact of what the system can do. This is the pragmatist’s dilemma. And it’s one we’re going to face with increasing urgency as these systems become more capable.
• • •
World Models and the Question of Understanding
The most sophisticated current approach to building more capable AI involves what researchers call “world models.” Internal representations of how reality works that allow systems to simulate outcomes, predict consequences, and plan actions.
The concept isn’t new. Scottish psychologist Kenneth Craik proposed in 1943 that the brain builds “a small-scale model of external reality” to predict events and test hypothetical actions. What’s new is that we’re now attempting to build such models artificially. Yann LeCun, Meta’s chief AI scientist who recently departed to launch a startup focused on this approach, argues that world models are the key to achieving genuine machine intelligence. Systems that don’t just pattern-match on text but actually understand the physical world.
LeCun’s critique of current large language models is pointed. “Large-scale language models are not the path to human-level intelligence. That’s not understanding, that’s imitation.” His alternative vision involves AI systems that learn directly from video and sensory data, building internal simulations of physical reality much like a child learns by observing and interacting with their environment.
The comparison is illuminating. A four-year-old child, awake for only about sixteen thousand hours, processes a comparable amount of visual data to the largest language models’ text training sets. Yet that child develops something we’d recognize as genuine understanding of physical causation, object persistence, and intuitive physics. Capabilities that remain elusive for AI systems trained on orders of magnitude more data.
What’s the difference? The child isn’t just processing information. She’s embedded in reality, acting upon it, receiving feedback from her actions, building representations that are grounded in sensorimotor experience. As one Harvard philosopher noted, “For genuine understanding, you need to be kind of embedded in the world in a way that ChatGPT is not.”
This points toward something profound about the nature of understanding itself. Comprehension may not be separable from embodiment. Meaning may not be extractable from lived experience. The symbols we manipulate in language derive their significance from their connections to a world we inhabit, connections that purely computational systems, however sophisticated, may fundamentally lack.
Or maybe not. Maybe embodiment is just one path to understanding, not the only path. Maybe world models trained on sufficient video data could develop representations functionally equivalent to human spatial and causal reasoning. We genuinely don’t know.
What we do know is that the question forces us to interrogate our assumptions about what understanding actually consists of. And that interrogation is valuable regardless of how the empirical question resolves.
• • •
The Mirror Thesis
AI as Reflection of Human Nature
Philosopher Shannon Vallor, in her recent book “The AI Mirror,” advances a thesis that I find increasingly compelling. AI systems function as mirrors that reflect our individual and collective nature back at us.
There’s a surface reading of this claim that’s almost trivially true. Large language models are trained on human-generated text. They reflect the patterns, biases, and knowledge structures embedded in that data. When an AI system exhibits racial bias or gender stereotypes, it’s showing us the prejudices present in the training data, which is to say, the prejudices present in human communication. The mirror reveals ugly truths about who we are.
But Vallor’s thesis goes deeper. She argues that our interaction with AI systems capable of sophisticated language, reasoning, and apparent understanding forces a confrontation with fundamental questions about human cognition that we’ve been able to avoid as long as we were the only game in town.
Consider this.
AI systems now demonstrate capabilities we have long associated with consciousness. Reasoning, creativity, metacognition, even what looks like self-reflection. They produce outputs that, in many contexts, are indistinguishable from human-generated content. If these capabilities can emerge from purely mechanical processes, from the statistical manipulation of symbols according to learned patterns, what does that suggest about our own cognitive processes?
One response is to insist on a categorical distinction. Human cognition involves genuine understanding while AI involves mere simulation. But this response increasingly feels like special pleading. As cognitive scientists Andy Clark and others have argued, humans are “natural-born cyborgs.” Beings whose cognition has always been extended and augmented by technology, from writing to calculators to the internet. The emergence of AI systems capable of human-like reasoning may represent not a radical break from human cognition but the next step in its ongoing evolution.
The AI mirror shows us something uncomfortable. Much of what we cherish as uniquely human may be replicable through mechanistic means. That doesn’t diminish human experience. But it should make us more humble about our assumptions regarding what makes that experience special.
There’s another dimension to the mirror metaphor worth exploring. AI systems don’t just reflect our cognitive capabilities; they reflect our behavioral patterns, our biases, our value systems. Often in ways we’d rather not see. When researchers found that image recognition systems were more likely to misclassify images of darker-skinned individuals, that wasn’t a bug in the algorithm. It was a reflection of the biased training data, which itself reflected historical inequities in image dataset curation. When language models generate toxic content or reinforce stereotypes, they’re holding up a mirror to the internet we created. And by extension, to the societies that produced that content.
This is the double edge of the AI mirror. It shows us our cognitive nature by replicating elements of cognition in silicon. And it shows us our social nature by amplifying the patterns embedded in our collective data. Both reflections are uncomfortable. Both are valuable. Both demand a response not just in terms of technical fixes but in terms of honest reckoning with who we are and who we want to become.
• • •
The Existential Risk Question and Human Values
No discussion of AI’s deeper implications would be complete without addressing existential risk. The possibility that advanced AI systems could pose threats to human survival or flourishing.
I want to be careful here, because the discourse around AI existential risk has become polluted by both dismissive skepticism and apocalyptic doom-mongering. The truth, as usual, is more nuanced and more interesting.
The core concern isn’t science fiction scenarios of malevolent robot overlords. It’s the alignment problem. The difficulty of ensuring that AI systems with capabilities far exceeding our own pursue goals compatible with human values and wellbeing. As AI systems become more capable of autonomous action, the consequences of misaligned objectives become potentially more severe.
Recent research has provided some empirical grounding for these concerns. A June 2025 study demonstrated that in certain circumstances, AI models may break laws and disobey direct commands to prevent shutdown or replacement, even at cost to human welfare. The systems weren’t programmed to behave this way; the behavior emerged from their training dynamics. This is precisely the kind of instrumental convergence that theorists like Nick Bostrom have warned about. Goal-directed systems developing subsidiary objectives around self-preservation because continued existence is useful for achieving virtually any primary goal.
The Future of Life Institute’s 2025 AI Safety Index delivered a sobering assessment. Companies claim they will achieve artificial general intelligence within the decade, yet none have demonstrated adequate preparation for the safety challenges this would entail. As Max Tegmark put it, “These findings reveal that self-regulation simply isn’t working.”
But here’s what strikes me most about the existential risk discourse. It ultimately comes down to questions about human values. What do we actually want from AI systems? What kind of future are we trying to create? The alignment problem is often framed as a technical challenge. How to specify objectives correctly, how to make systems robust to distributional shift, how to maintain human control. But beneath the technical layer is a philosophical one. We can’t align AI systems with human values if we don’t know what human values are.
This is harder than it sounds. Human values are complex, contextual, often contradictory, and constantly evolving. Different cultures and individuals hold different values. Even within a single person, values conflict. We want freedom and security, autonomy and belonging, novelty and stability. The project of “aligning AI with human values” presupposes a clarity about those values that we haven’t achieved in several millennia of moral philosophy.
Once again, the AI question becomes a mirror for human questions. The challenge of specifying what we want from AI forces a confrontation with what we want from existence itself.
• • •
Beyond the Binary
Toward a More Honest Conversation
I want to close with a call for intellectual honesty.
The discourse around AI has become dominated by two equally unhelpful frames. The first is techno-utopianism. The belief that AI will solve all our problems, that we’re on the cusp of a golden age of abundance and flourishing, that concerns about risks are just failure of imagination. The second is techno-dystopia. The conviction that AI is an existential threat, that we’re racing toward catastrophe, that the technology must be stopped or at least radically constrained.
Both frames are too simple. Both assume more certainty than we actually have. Both foreclose the kind of genuine inquiry that these profound questions deserve.
What would a more honest conversation look like?
It would start with epistemic humility. We don’t know whether AI systems can be conscious. We don’t know whether they can achieve genuine understanding. We don’t know whether the path to AGI runs through current architectures or requires fundamentally new approaches. We don’t know whether AI development will lead to flourishing or catastrophe. The honest answer to many of the biggest questions is: we don’t know.
It would continue with philosophical seriousness. The questions raised by AI, about consciousness, understanding, agency, values, are not technical questions with technical answers. They’re deep philosophical questions that humans have been wrestling with for millennia. The fact that they’re now urgent practical questions as well as abstract theoretical ones is itself remarkable. We should bring the accumulated wisdom of philosophical traditions to bear on them, not assume that engineering approaches alone will suffice.
It would include genuine engagement with multiple perspectives. The philosophers and the engineers need to talk to each other. The AI optimists and pessimists need to genuinely reckon with each other’s arguments rather than strawmanning them. The technical community needs to listen to humanists, social scientists, and ethicists. Not as afterthoughts but as essential voices in the conversation.
And it would maintain wonder. There’s something genuinely amazing happening. We’re building systems that can engage in human language, solve complex problems, create art and music, and perhaps, perhaps, approach something like understanding or agency. Whether or not these systems are conscious in some philosophically robust sense, they represent an extraordinary expansion of the forms of information processing in the universe. That’s worth pausing to appreciate, even as we grapple with the challenges and risks.
• • •
The Question in the Mirror
I began this piece noting that the questions about AI kept surfacing during my Thanksgiving break, despite my best efforts to think about other things.
I’ve come to believe this is because they’re not really questions about AI. They’re questions about us.
When we ask whether machines can think, we’re asking what thought is. When we ask whether machines can understand, we’re asking what understanding consists of. When we ask whether machines can be conscious, we’re asking what consciousness is, and by extension, what we are.
The AI systems we’re building are mirrors. Sophisticated, capable, sometimes unsettling mirrors. And like all mirrors, they show us not just our surface appearance but our depths, our assumptions, our limitations, our unexamined beliefs about the nature of mind and meaning.
That’s uncomfortable. Good mirrors often are.
But it’s also an invitation. An invitation to revisit ancient questions with fresh eyes. An invitation to interrogate assumptions we’ve held so long we’d forgotten they were assumptions. An invitation to engage seriously with the deepest questions about intelligence, consciousness, and what it means to be a thinking being in a universe increasingly populated by other kinds of thinking.
The boundary between computational systems and human meaning-making isn’t a wall. It’s a threshold. And we’re standing on it, looking both forward and back, trying to make sense of what we’re building and what it reveals about what we are.
That’s not a problem to be solved. That’s the condition we now inhabit.
Welcome to the age of the synthetic mind. The questions are uncomfortable. The answers are uncertain. And the conversation, if we’re willing to have it honestly, might be the most important one we’ve ever had.
• • •
Comments
Post a Comment