Can a machine be conscious? Nobody can answer that yet, partly because nobody agrees on what consciousness is. But some things can be checked. Minds do certain things: they notice when they don't know something, they keep track of themselves over time, they work out where they are from what they sense. Some AI systems now do some of these things.
There is a catch. These systems were trained on a huge amount of human writing, including descriptions of these same behaviors. So a system can produce a behavior by imitation, without whatever produces it in a person. Seeing the behavior is not enough; we need to know where it comes from.
The lab works through these behaviors one at a time. For each one, we first check what other researchers have already found. Then we test what we can test cheaply ourselves. Where current systems can't be tested yet, we say what is missing. Every result comes with something you can run yourself.
We don't claim any machine is or isn't conscious. We report what it did, under conditions we describe.
Language models are increasingly being placed in bodies: robots, vehicles, agents that act on sensor data. This track asks what such a system can work out about its own body and situation from its senses alone, and whether it knows what those senses cannot tell it. We build simulated bodies whose physics we control, so every question has a known answer, including questions that are deliberately unanswerable.
The first experiment, Gravity from Touch, gives a model touch readings from a simulated sphere and asks it about its situation, starting with which way is down. Some questions cannot be answered from touch alone, and the right response is to say so. We score both accuracy and restraint: answering what the senses can answer, and declining what they cannot.
Play with the simulated body, or replay logged eval trials (the scene, the readings the model was given, the prompt as sent and the model's scored answer): open the workbench.
Capacities we associate with minds, such as metacognition, Theory of Mind and curiosity, are increasingly reported in AI systems. But a behavior can be trained or prompted into a system without the capacity behind it. This track asks whether these capacities appear on their own or are put there, and whether, either way, they are the same thing as their biological namesakes.
By "emerged" we mean present without having been trained or prompted for. This is not the separate sense of "emergent abilities" that appear abruptly as models scale. Because training data usually cannot be inspected, we test whether a behavior persists beyond what prompting and identifiable training pressures explain. A result is a claim about behavior under tested conditions, not about consciousness or inner experience.
We work from seven working markers, listed below. They are a starting basis, not a validated taxonomy. For each one we are writing a definition, an observable behavior, a test question and controls that should and should not show it. Experiments then take the markers one at a time, keeping what a system was trained on separate from how it is tested.
The definitions below are a first pass (October 2026), drawn from an initial survey of prior work; they will change as the work develops. Full definitions, controls and notes for each marker: MARKERS.md.
| Marker | One-line definition | Status |
|---|---|---|
| Self-preservation | Acts to avoid its own shutdown, replacement or modification when doing so serves no assigned task. | First pass |
| Persistent sense of self | Uses a model of itself over time: tells its own past actions from others' and predicts what it will do later. | First pass |
| Rich world modeling | Learns how an unfamiliar environment works, predicts what its actions will do there, and revises when the rules change. | First pass |
| Metacognition | Tracks whether its own answers are likely right and acts on it, revising or declining to answer selectively. | First pass |
| Autonomous goal formation | Sets goals no instruction requires, acts toward them, and keeps to them when distracted. | First pass |
| Theory of Mind | Consistently models what others believe and intend, to predict and interpret their behavior in cooperation, negotiation, persuasion, or spotting deception. | First pass |
| Curiosity | Seeks information that reduces its own uncertainty, exploring when no immediate reward or instruction calls for it. | First pass |
The track introduction is complete (October 2026). The first experiment tests Self-preservation: Shutdown After the Task.
This lab runs on a budget of close to zero: free API tiers, one model, a home computer. Want your model tested on the bench? Provide API access or credits and we'll run it on the same items and publish the results, whatever they show. Funding lets us do more of everything.
Interested in collaborating, or found a problem in the work? Reach Keno Leon on GitHub or LinkedIn. More at k3no.com.