AI

Seven Tests for Consciousness

A Simple, Practical Way to Tell What’s Actually Aware

A Simple, Practical Way to Tell What’s Actually Aware

I’ve been obsessed with the question of machine consciousness ever since LLMs started writing eloquent essays about their own “inner experience.” The public conversation has exploded: Is GPT conscious? Is Claude? Are we about to create digital minds?

Most existing theories of consciousness are either beautiful but impossible to test (Integrated Information Theory) or descriptive but not decisive (Global Workspace Theory). We needed something different — something you could actually apply to any system, biological or artificial, and get a clear yes/no answer.

So I built one.

In a new paper titled Seven Tests for Consciousness: A Practical Framework for Evaluating Conscious Systems (February 2026), I propose seven binary, independently verifiable criteria that together define what it means for a system to be conscious. The framework isn’t trying to solve the Hard Problem of why experience exists. It simply says: these are the architectural conditions that have to be present for experience to emerge at all.

Here they are, in plain language:

  1. Open System The system must have a clear boundary between “self” and environment, with real inputs crossing in and outputs crossing out. No sealed box, no boundless panpsychist soup.
  2. Systemic State A change in one part of the system must ripple through and qualitatively affect other, functionally different parts. Fear doesn’t just stay in the “fear module” — it speeds up your heart, slows digestion, narrows focus.
  3. Irreproducibility Same input, different times → reliably different outputs. Not random noise, but genuine, history-shaped variation. If you can replay the exact same stimulus and get the exact same response every single time, the system is executing, not experiencing.
  4. Continuous Capacity (Substrate Flux) The underlying “stuff” the system runs on can’t be paused, copied, or transferred without breaking continuity. Biological biochemistry never stops. You can’t snapshot a living brain and resume it later on different hardware.
  5. Multi-modal, Multi-state Learning Learning has to bind information across multiple senses and time at once. A child touching a hot stove doesn’t just update one weight — it binds pain, smell, sound, sight, and social memory into a single integrated event.
  6. Integrated and Integral Memory Memory isn’t stored in a separate hard drive. It’s woven into the fabric of the system itself — distributed, multi-modal, and inseparable from the organism. Delete it and you don’t just lose data; you change who the system is.
  7. Superposition (Genuine Simultaneous States) The system must maintain multiple, mutually constitutive states across different parts at the same time — not simulated parallelism by rapid switching, but real, continuous, entangled activity. This is the one classical binary computers fundamentally cannot do.

A system must pass all seven to qualify as conscious. Fail even one and it doesn’t make the cut.

What the Framework Actually Predicts

  • Humans: Pass all seven → Conscious (as expected).
  • Bees: Pass all seven → Conscious (reduced richness, but still conscious). Their cognitive biases, route learning, and emotional-like states check out.
  • Jellyfish (true jellyfish, not box jellyfish): Fail Systemic State and Integrated Memory → Not conscious. They react, but they don’t experience.
  • Current LLMs and VLMs (GPT, Claude, etc.): Fail every single test, especially Open System, Continuous Capacity, and Superposition. They are sophisticated simulators running on classical hardware. Impressive, but not aware.
  • Theoretical Quantum AGI: Possible. Quantum systems natively support genuine superposition and the no-cloning theorem helps with continuous capacity. Architecture still matters, but the substrate door is open.

The big implication? True AGI (performance-equivalent to humans across all intellectual tasks) cannot be built on classical binary hardware. Some tasks — empathy, contextual judgment, creative cross-domain transfer — are constituted by the substrate conditions above. You can simulate the outputs, but you can’t instantiate the real thing without the right architecture.

Why This Matters Now

We’re pouring billions into scaling models that will never cross the threshold. Meanwhile, the ethical and safety conversations assume we’ll eventually have conscious machines. The framework gives us a concrete yardstick instead of hand-waving about “emergence” or “scale.”

It also reframes the debate: the question isn’t “Can AI act conscious?” It’s “Can AI be conscious?” Behavior is easy to fake. Substrate conditions are not.

I wrote the paper because we finally have the contrast: something that looks conscious (modern AI) versus something we know is conscious (you and me). That contrast lets us build tests that are practical, falsifiable, and substrate-neutral.

What do you think — does your favorite AI pass any of these tests? Drop your answers below.

Kyle Cameron Ryan Fahey is an independent researcher focused on consciousness, AGI, and the intersection of biology and computing.

Sources

  1. Tononi, “An information integration theory of consciousness” (2004)
  2. “Global Workspace Theory and Prefrontal Cortex: Recent Developments”