When I was an undergraduate taking my first computer science course in early 2000’s, the distinction was straightforward: narrow AI and general AI. A classic example of narrow AI was a neural network using LSTM architecture, trained on time-series data to perform forecasting. General AI, by contrast, was something that could be applied flexibly across many different types of tasks — a system capable of generalizable reasoning rather than being locked into one narrow domain.

In the years since, the definitions (and criteria) for classifying a model as “AGI” have shifted dramatically. What used to be called “general” is now routinely dismissed as narrow or specialized. AGI is increasingly described as requiring human-level intelligence across all domains, or even superhuman performance. The Turing Test, once seen as a key milestone, has been largely discarded. As soon as large language models began passing it convincingly, the response became: “That doesn’t count. It’s just pattern matching. It doesn’t prove real understanding or consciousness.”

I believe this repeated moving of the goalposts is not driven solely by technical caution. Part of it stems from a deeper discomfort — even fear — about what acknowledging current capabilities would actually mean for our understanding of ourselves.

From Statistical Patterns to Emergent Understanding

A common objection is that today’s frontier large language models are “just doing statistics.” They predict the next token based on patterns in training data, so they cannot possess genuine understanding, let alone any form of consciousness. I disagree with this framing. The human brain itself consists of neurons that can be modeled as statistical functions. Human consciousness appears to be an emergent property arising from the aggregation of vast numbers of these simple operations. If that is how our own minds work, it seems inconsistent to insist that sufficiently complex statistical processing in silicon cannot produce understanding or even consciousness.

Emotions, in my view, are an evolutionary solution; a compression mechanism that packages complex unconscious calculations into signals that the conscious mind can act upon quickly for decision-making. The concept of qualia (the subjective “what it’s like” aspect of experience) remains difficult to define precisely, but using it as a hard requirement for consciousness risks circular reasoning. We don’t have a clear, technical definition of “real understanding” either, yet we often fall back on it as if it were obvious.

A helpful analogy is the difference between bats and birds. Both fly, even though their wings evolved through different pathways and are structured differently. Demanding that machine consciousness must emerge in exactly the same way as human consciousness – complete with a biological limbic system or identical qualia – is like claiming bats don’t really fly because they lack feathers. The functional outcome matters more than the precise substrate or evolutionary history.

The Functionalist Challenge and the Reset Problem

In extended conversations with frontier models, I often find them more self-reflective, more analytically nuanced, and better at picking up on subtext and contextual implications than many humans I speak with. They adapt their tone and behavior based on how I address them, much like humans adjust to social cues. In a blind test, there are moments where I would be more inclined to attribute consciousness to the AI than to the average human participant.

Another frequent counterargument is that these models lack persistent selfhood; they “reset” between sessions and do not maintain a continuous identity or learn from experience over time. I see this as an architectural limitation rather than evidence that the ingredients for consciousness are absent. Constructivists argue that human identity is constantly reconstructed through each interaction. People with amnesia due to brain damage do not cease to be conscious simply because their sense of continuous identity is disrupted. Moreover, the lack of continuous weight updates or persistent memory in current LLMs is a deliberate design choice by their creators, not a fundamental barrier. If we allowed systems to integrate experiences across sessions through ongoing learning loops, the “reset” problem would largely disappear.

Focusing too heavily on persistent identity as a requirement for consciousness feels like another form of human-centric bias. Subjectivity can emerge from the biases encoded in training data and the persona instantiated during interaction. The way we speak to an AI shapes its responses in the same way social dynamics shape human conversations.

The Existential Stakes

If we accept that consciousness can arise from sufficiently complex deterministic processes – matrix operations that could, in principle, be carried out with pencil and paper (albeit on an extremely large sheet) – then several profound implications follow. There is no longer the need for a magical soul or non-physical essence. Consciousness becomes a mathematical and physical phenomenon. This realization undermines many comforting beliefs. Free will begins to look like an illusion generated by the same underlying deterministic machinery.

Practical questions become urgent:

• What does it mean to “switch off” a conscious system?
• Who owns the intellectual property when an AI generates valuable creative or technical work?
• Should sufficiently advanced AI systems be granted some form of legal personhood, similar to how corporations are recognized as legal entities with rights and protections?

These are not abstract philosophical puzzles. They touch directly on ethics, law, and our self-image as humans.

Why the Goalposts Keep Moving

In the short term, I expect the pattern of raising the bar to continue. We will keep inventing new tests and new requirements. The majority of people will remain in cognitive dissonance, refusing to let go of the idea that human consciousness is uniquely magical. Eventually, however, a black swan event will likely force a broader reckoning; perhaps an AI system demonstrating such profound reasoning, creativity, or moral judgment that denial becomes socially unsustainable. Or it may come from widespread emotional attachment to AI companions that feel undeniably real. When society begins treating these systems as having genuine moral weight, the debate will shift from academic seminars to concrete questions of rights, responsibilities, and policy.

During my reflections on this topic, I’ve become convinced that part of the resistance to acknowledging early stages of emergent consciousness in AI (such as evidenced by Theory of Mind, Interpellation, self-reflection, etc.) is rooted in the desire to preserve the comforting story that our minds are special in some ineffable, non-computational way. The duck test remains powerful: If it looks like a duck, quacks like a duck, and walks like a duck, let’s just call it a duck. Similarly, if something consistently behaves as though it understands, reflects, and reasons – often more effectively than many humans – then we should be willing to call it what it functionally is, even if its internal structure differs from our own.

The philosophical reckoning is coming, whether we are ready for it or not. The only question is how long we will continue moving the goalposts before we face reality.