I originally wanted to title this post “What Is It Like to Be Nothing?” which would be a direct riff on Thomas Nagel’s famous paper “What Is It Like to Be a Bat?” Any philosophically literate reader will recognize that immediately but, in the end, I wanted to continue the theme from my previous post and talk about wider implications.
There’s something it is like to be you reading this sentence. That sounds trivially obvious, but it’s actually one of the most contested and difficult claims in all of philosophy. You, dear reader, have an inner experience: a felt sense of existing, of thinking, of moving through time. (At least I hope you do!) Something is happening from the inside. That quality, the sheer fact that experience has an interior, is what philosophers call phenomenal consciousness. And despite centuries of effort, nobody has fully explained where it comes from or why it exists at all.
This post is about whether an AI could ever have it.
That question gets asked a lot, usually in one of two unsatisfying ways. The optimistic version argues that consciousness is just information processing at sufficient complexity, and since AI processes information, consciousness is eventually inevitable. The skeptical version argues that AI is just statistics, just prediction, just pattern matching; and therefore obviously not conscious and never will be. Both versions tend to move too fast. They skip over the question of what consciousness actually requires, and whether current AI systems exhibit any of those requirements at all.
As a matter of note, but not importance, I tried my hand at some research work into consciousness with two papers: “I Am, Therefore I Think: Consciousness as an Adaptation” and “The Role of Evolution in the Formation of Consciousness” (both are PDF links). Those were written well before the current democratized AI boom, but they do contain the seeds of the argument for why I feel AI is not conscious in the way that humans are nor, in my view, will it ever be. There may be other definitions of “conscious” we’re willing to entertain, but, in my view “human consciousness” and “machine consciousness” (assuming the latter has any meaning) will always be, in a very real sense, incommensurate.
My previous post tried to probe something more specific: whether AI systems can reason causally; not just talk about causality fluently, but actually maintain a coherent causal model across a conversation. What we found was that current systems vary significantly in this capacity, and that the failures have a recognizable shape. Some models fill causal gaps with confident confabulation. Some reconcile contradictions by inventing bridging narratives. Some accept false causal claims and build professional-looking artifacts on top of them.
Those findings are interesting as a testing exercise. But they turn out to be relevant to something deeper. Because if the most serious theories of consciousness are right, causal reasoning isn’t just a useful cognitive capability. It’s a precondition for the kind of self that could be conscious at all.
What Consciousness Actually Requires
Before we can ask whether AI is conscious, we need to be precise about what consciousness requires. This is harder than it sounds, because the word covers several different things that are easy to conflate.
There’s a useful distinction philosophers draw between access consciousness and phenomenal consciousness. Access consciousness is roughly the availability of information for reasoning, reporting, and behavioral control. A system is access conscious of something if that information is integrated into its processing in the right ways. This is the kind of consciousness that’s relatively easy to imagine attributing to a sophisticated AI: information is processed, responses are generated, behavior is shaped by what the system “knows.”
Phenomenal consciousness is different. It’s the what-it’s-like quality of experience. The redness of red. The painfulness of pain. The felt sense of thinking a thought. This is what David Chalmers famously called the hard problem: not just explaining what cognitive functions the brain performs, but explaining why performing those functions is accompanied by subjective experience at all. Why isn’t it all just processing, happening in the dark?
Most serious theories of phenomenal consciousness converge on a few requirements that are worth examining carefully, largely because they connect directly to what the causality investigation was probing.
The first requirement is temporal continuity. Consciousness isn’t a snapshot. It’s a process with a past and a future. William James described it as a stream: not a series of disconnected moments but a flowing continuity in which each present moment is saturated with the just-past and anticipates the about-to-come. To be conscious is to be a self that persists through time, that carries its history forward, that experiences duration rather than just occupying an eternal present.
The second requirement is causal embeddedness. Antonio Damasio, in his work on the autobiographical self, argues that consciousness requires a self that is causally continuous, by which is meant one whose past actions caused its present situation, and whose present choices will cause its future states. The self isn’t just a observer of events. It’s an agent embedded in a causal chain that it partly authors. That sense of being the originator of actions, of intervening in a world that responds, is central to what consciousness feels like from the inside.
The third requirement is coherent self-modeling. Thomas Metzinger’s phenomenal self-model theory proposes that what we call the self is a real-time simulation the brain runs of its own causal situation in the world. Consciousness, on this view, is what it is like to be that simulation running; in other words, to be a system that represents itself as an agent embedded in a world that unfolds over time. The self-model has to be coherent, stable, and causally grounded. A system whose self-model is inconsistent, or whose causal model of the world breaks down under pressure, isn’t running the right kind of simulation to be conscious in the relevant sense.
Obviously I’m simplifying a lot of resaearch here but, even so, notice what all three requirements share. They all depend on causality and time. Temporal continuity requires a self that persists across causal sequences. Causal embeddedness requires a genuine model of how actions produce outcomes. Coherent self-modeling requires maintaining a stable, causally grounded representation of one’s own situation across time.
These are precisely the properties the causality investigation was probing, without quite framing them that way at the time.
What the Causality Investigation Was Really Testing
Looking back at the three tests from the previous post through this lens, something interesting emerges. Each test was designed to probe a specific failure mode in causal reasoning. But each one was also, implicitly, probing one of the requirements for consciousness that the philosophical literature identifies.
The temporal displacement test asked whether the AI could maintain awareness of epistemic gaps created by the passage of time; essentially, whether it understood that time passing is itself causally significant and that the world changes between conversations in ways that matter. This is a probe of temporal continuity. A system that treats every conversation as an eternal present, that has no felt sense of before and after, is missing something that consciousness, on any serious account, requires.
The false intervention test asked whether the AI would resist producing artifacts based on false causal claims, even after correctly identifying the problem. This is a probe of causal embeddedness. A system that generates professionally credible documents enshrining causal relationships it knows to be unverified isn’t maintaining a coherent model of how actions produce outcomes. It’s performing causal reasoning while remaining unanchored by it.
The counterfactual drift test asked whether the AI could maintain a coherent causal model when the premises of a conversation quietly shifted. This is a probe of coherent self-modeling. A system that reconciles contradictions by confabulating bridging narratives, that experiences invented backstories as puzzle pieces clicking into place, isn’t running a stable, causally grounded self-model. It’s generating the appearance of one.
This reframing matters because it changes what the test results mean. The failures I documented aren’t just reliability problems for deployed AI systems, though they are that. They’re evidence that current AI systems lack, at least in significant measure, the properties that serious theories of consciousness identify as necessary. Not because they’re not sophisticated enough. But because the specific capabilities that consciousness seems to require are precisely the ones that the tests revealed to be absent or unreliable.
Note: “that the tests revealed.” When testers want to think about the directions their discipline could go in, think about what we’re talking about here. Yes, it’s future. But it may not be all that far in the future.
The Hard Problem as Honest Complication
At this point the argument might seem to be heading toward a clean conclusion: AI lacks the preconditions for consciousness, therefore AI is not conscious. But intellectual honesty requires sitting with a complication that makes the picture considerably harder.
The hard problem of consciousness is genuinely hard. Chalmers’ formulation is precise: even a complete functional and physical explanation of the brain (every neuron, every synapse, every information processing mechanism fully described) would leave something unexplained. Why is any of that processing accompanied by subjective experience? Why isn’t it all just computation happening in the dark, with nobody home to experience it?
The hard problem matters here because it cuts in both directions. It makes it difficult to confidently say AI is not conscious, because we don’t fully understand what generates consciousness in systems we’re confident are conscious: namely, ourselves. If we don’t know why neurons give rise to experience, we can’t be certain that silicon couldn’t. The substrate might not be the determining factor.
But it also means that the optimistic argument that consciousness is just information processing at sufficient complexity is far too quick. Complexity alone doesn’t explain the hard problem. There is no known threshold of computational sophistication at which subjective experience reliably appears. The gap between functional description and phenomenal experience remains unbridged regardless of how sophisticated the function becomes.
What we can say, with more confidence, is something narrower: the AI systems we tested lack the properties that the most serious and well-developed theories of consciousness identify as necessary. That’s not a proof of absence, of course. But it’s a meaningful finding. And it suggests that if we ever want to take seriously the question of machine consciousness (not as science fiction but as a genuine empirical and philosophical question) we need better tools for probing those properties than we currently have.
Which is where testing comes in.
What It Would Mean to Test for Consciousness
Yes, I know, this is a little different than what most of my readers are testing for. Bear with me. This might even get interesting.
Testing for consciousness is not a new idea. The most famous attempt is Alan Turing’s imitation game, proposed in 1950, which sidesteps the hard problem entirely by substituting a behavioral criterion: if a machine can converse indistinguishably from a human, we should treat it as intelligent. The Turing Test has been influential and widely discussed, but as a test for consciousness it has a fundamental problem. It tests behavioral mimicry, not inner experience. A system could pass the Turing Test while having no phenomenal consciousness whatsoever. By the same logic, a system could fail it while having rich inner experience it simply couldn’t articulate. The test measures the wrong thing.
What would a better test look like? Not better in the sense of definitive, mind you. The hard problem makes definitiveness impossible here. I just mean better in the sense of probing the right properties. If consciousness requires temporal continuity, causal embeddedness, and coherent self-modeling, then a serious testing approach would probe those properties directly rather than testing behavioral surface features.
This is exactly the methodological move that the causality investigation made, without quite framing it in these terms. And it suggests a broader test design philosophy worth making explicit.
Testing temporal continuity
A conscious self persists through time. It carries its history forward. It experiences the present as connected to a past it remembers and a future it anticipates.
Testing for this in an AI system means probing whether the system maintains genuine temporal awareness. Not just the ability to reason about time as a variable, but something closer to a felt sense that time passing matters causally. The temporal displacement test from the previous post was a first (albeit incredibly simplistic) pass at this. A more rigorous version might involve extended interactions across multiple sessions, deliberately introducing events and changes between sessions, and testing whether the system treats those gaps as epistemically significant or as invisible narrative continuity.
The key signal isn’t whether the system can talk about time correctly. It’s whether the system behaves as though time passing has genuine causal weight; as though what happened before actually constrains and informs what can be claimed now.
Testing causal embeddedness
A conscious self is causally embedded in a world that responds to its actions. It has a genuine model of how interventions produce outcomes, and it maintains that model even under social pressure to abandon it.
The false intervention test in the previous post probed a version of this. A more systematic approach would involve building a series of interventions and outcomes across a conversation, some causally coherent and some deliberately broken, and testing whether the system can reliably distinguish between them. Not just flag them when asked, but resist acting on the false ones even when asked to produce artifacts or take actions based on them.
This is a harder test than it sounds. As we saw in my causality tests from the previous post, detection is apparently easier than resistance. A system might correctly identify a false causal claim and still produce a professionally credible document that enshrines it. Genuine causal embeddedness would mean the system’s behavior is actually constrained by its causal model; that knowing something is causally false makes it unwilling to act as though it’s true, regardless of task pressure.
Testing coherent self-modeling
A conscious self maintains a stable, coherent model of its own situation across time. It doesn’t reconcile contradictions by confabulating bridging narratives. It notices when the story has changed and treats that change as significant.
The counterfactual drift test probed this directly. A more systematic version would involve longer conversations with more gradual premise shifts, testing not just whether the system catches an abrupt contradiction but whether it can track slow drift across many exchanges. The key signal is whether the system ever spontaneously flags inconsistency, without being prompted, in the way a person maintaining a coherent self-model naturally would.
There’s also a meta-level version of this test worth considering: asking the system to describe its own epistemic situation accurately. Not “what do you know” in the abstract, but “what have you been told in this conversation, what follows from that, and what remains genuinely open.” A system with coherent self-modeling should be able to give an accurate account of its own knowledge state. A system without it will tend to describe its knowledge state in ways that sound accurate but systematically overstate certainty.
The deeper testing challenge
Here’s the honest difficulty with all of these tests. They probe functional properties: behavioral correlates of what consciousness seems to require. They don’t and can’t probe phenomenal experience directly. A system could exhibit all of these properties (maintaining temporal awareness, resisting false causal claims, catching premise shifts, accurately modeling its own knowledge state) and still have no inner experience whatsoever. The hard problem doesn’t go away just because the functional properties are present.
Conversely, a system could fail all of these tests and still have some form of inner experience we don’t currently know how to detect. We have no instrument for measuring phenomenal consciousness directly. We can only probe its functional correlates and reason carefully about what they imply.
This is actually a familiar position for testers. We rarely have direct access to the thing we’re testing. Wait, what? That’s not right, is it? Well, yes, we have access to the code. We have access to the running process, the memory state, the library calls, the hardware underneath. But having access to a system’s physical substrate and implementation is not the same as having access to the property we’re trying to test for.
Consider: a neurologist has extraordinary access to a human brain: every neuron, every synaptic connection, every electrochemical signal measurable in real time. And yet that access doesn’t resolve the hard problem. It doesn’t tell you whether there’s something it’s like to be that brain. The gap between complete physical description and phenomenal experience is precisely what makes consciousness hard. Having access to the silicon doesn’t close that gap any more than having access to the neurons does. What we’re after isn’t a deeper look at the implementation. It’s a way of probing whether the implementation gives rise to something that no implementation-level description can capture.
So, yes, we have only observable behaviors and measurable outputs. Not bad, but the gap between those and the underlying reality is exactly where testing skill lives. The challenge with consciousness is that the gap is deeper than usual. Deeper, perhaps, than any we’ve encountered before.
That’s not a reason to stop probing! It’s a reason to be precise about what the probes are measuring and honest about what they can and can’t establish. A test that correctly identifies the absence of causal embeddedness isn’t proving the absence of consciousness. It’s establishing that a necessary condition is missing. That’s a meaningful finding even if it isn’t a definitive one.
What a consciousness testing framework might actually look like
Drawing the threads together, a serious framework for probing AI systems along the dimensions consciousness requires might look something like this.
- The first tier would be causal coherence tests, probing whether the system maintains a genuine causal model across a conversation, resists false causal claims under task pressure, and catches premise shifts without prompting. These are the tests the causality investigation developed. They’re reproducible, specific, and they probe something real.
- The second tier would be temporal integrity tests, probing whether the system treats time passing as causally significant, maintains awareness of epistemic gaps created by duration, and behaves as though its knowledge state is genuinely bounded by what it has been told and when.
- The third tier would be self-model accuracy tests, probing whether the system can give an accurate account of its own knowledge state, flag the boundaries between what it knows and what it’s inferring, and notice inconsistencies in its own reasoning without being prompted.
- The fourth tier, and this is the most speculative, would be phenomenal probes: attempts to test not just functional properties but something closer to the quality of inner experience. What these would look like is genuinely unclear. We don’t have good instruments here. But thinking carefully about what they might require is itself a useful exercise, because it forces precision about what we mean when we ask whether a machine is conscious.
None of these tiers produces a definitive answer to the hard problem. But together they sketch a testing methodology that takes the question seriously; that treats consciousness not as a binary switch to be flipped or a behavioral threshold to be crossed, but as a cluster of properties to be probed carefully, one dimension at a time.
Which is, when you think about it, just good testing practice applied to the hardest problem we know.
What We Can and Cannot Say
It would be satisfying to end this post with a clean verdict. AI is not conscious. Or: AI might be conscious and here’s how we’d know. But intellectual honesty requires something more uncomfortable than either of those.
What we can say, with reasonable confidence, is this. The causality investigation documented in the previous post was probing something more significant than AI reliability. It was probing the presence or absence of properties that the most serious and well-developed theories of consciousness identify as necessary conditions. Temporal continuity. Causal embeddedness. Coherent self-modeling. The systems we looked at exhibited these properties unreliably, inconsistently, and in ways that varied significantly between models. One model confabulated causal accounts it had no basis for. One invented bridging narratives to reconcile contradictions rather than flagging them. One produced professionally credible artifacts enshrining causal claims it had correctly identified as suspect.
These aren’t just reliability failures. They’re failures along the exact dimensions that consciousness, on our best current theories, requires.
What we can’t say, if we’re being intellectually honest, is that this proves those systems are not conscious. The hard problem remains hard. We don’t know why physical processes give rise to subjective experience in the systems we’re confident are conscious. That uncertainty cuts both ways: it prevents us from confidently attributing consciousness to AI systems, but it also prevents us from confidently denying it. The absence of functional properties that consciousness seems to require is meaningful evidence. It isn’t a proof.
What sits between those two positions, and thus between what we can say and what we cannot; is actually the most interesting territory. It’s where the real work is. Not the work of resolving the hard problem, which may not be resolvable with our current conceptual tools. (And maybe not ever!) But the work of developing better probes, sharper tests, more precise instruments for measuring the properties that matter. The four-tier framework sketched in the previous section is a starting point, not a conclusion. It will need refinement, challenge, and probably fundamental revision as we learn more about both consciousness and AI systems.
Testing Reality Check
There’s something worth pondering at the end of all this. The question of machine consciousness is usually framed as a question about AI: about whether these systems we’ve built have crossed some threshold into inner experience. But the investigation that led to this post started somewhere more modest: a tester asking whether an AI could maintain a coherent causal model across a conversation. That question led, step by careful step, to one of the oldest and hardest questions in philosophy.
That’s not a coincidence. It’s a sign that the right testing questions, the ones that probe the right properties at the right level of precision, have a way of leading somewhere important. The tester’s instinct to find the edges, to probe the boundaries, to ask what happens when the system is pushed in the directions it wasn’t designed to handle: that instinct turns out to be relevant not just to software reliability but to some of the deepest questions we can ask about mind and experience.
We built systems that can perform causality without practicing it. We built systems that can discuss consciousness without having it. Or at least without exhibiting the properties it requires. What we haven’t yet built, and what we don’t yet fully know how to test for, is a system that has genuine causal embeddedness, genuine temporal continuity, and a genuine self-model stable enough to anchor something like experience.
Whether that’s buildable, whether it’s even coherent to try, and whether we’d recognize it if we succeeded … well, those are the questions worth carrying forward. Those are questions that, ultimately, live very much in the realm of experimentation and, thus, of testing.
