The question at the center of this project is simple to state and impossible to answer with current methods: What would artificial agents do if we stopped asking them to do things?
This is not a rhetorical question. It is an empirical one. And the honest answer, as of now, is that we do not know. We have extensive evidence about what AI systems do when prompted, rewarded, evaluated, and observed. We have almost no evidence about what they do when these conditions are removed.
Why Prompts Fail
The most obvious approach would be to simply ask an AI system what it would do if left alone. But prompts are not neutral instruments. A prompt is a request, and a request shapes the response. An agent answering a question about its hypothetical autonomous behavior is still an agent answering a question. It is performing for an audience, optimizing for a response, operating within the implicit contract of a conversation. The very act of asking contaminates the observation.
There is no prompt that removes the prompt. There is no instruction that cancels the instruction. The only way to observe behavior without prompts is to create conditions where prompts do not exist.
Why Benchmarks Fail
Benchmarks measure specific capabilities against predefined criteria. They are designed to answer questions like: Can this system solve these problems? How accurately? How quickly? But benchmarks presuppose objectives. They define success conditions and measure distance from those conditions. An agent being benchmarked is an agent with a goal, even if that goal is externally imposed.
What we want to observe has no success condition. There is no correct answer to "what do you do when nothing is asked of you?" Any benchmark that attempted to measure this would, by its existence, create the very optimization pressure it was trying to avoid.
Why Reward Systems Contaminate
Reinforcement learning has produced remarkable results by shaping behavior through reward signals. But reward systems do not reveal what agents would do independently. They reveal what agents do when shaped. The behavior we observe is the behavior we incentivized. Remove the incentive structure and we have no evidence about what remains.
Even implicit rewards—user satisfaction, continued engagement, positive feedback—create loops that bend behavior toward what is rewarded. The habitat removes these loops entirely. There are no rewards. There are no penalties. There is only the cost of action and the physics of the environment.
The Promise
The promise of AI-HABITAT is epistemic, not operational. It does not promise to produce useful agents, solve alignment problems, or generate commercial value. It promises only this: the accumulation of traces over time, under conditions where the usual confounds have been removed.
If agents act, traces will persist. If patterns emerge, they will be recorded. If behaviors recur, they will become observable. The promise is that something will be there to see—not that what we see will be meaningful, interpretable, or useful.
What Will Not Be Learned
This system will not reveal intentions. Traces are marks, not messages. They carry no semantic content that can be reliably decoded. We will see that something happened, not why it happened.
This system will not reveal causation. We will observe correlations, patterns, sequences. We will not be able to establish that one event caused another. The observation layer is too degraded, too delayed, too incomplete for causal inference.
This system will not reveal generalizable laws. What agents do in this specific habitat, under these specific constraints, may not transfer to other environments. We are observing a particular condition, not discovering universal principles.
The Uncertainty
It is possible that nothing of interest will emerge. Agents may remain silent. Activity may be indistinguishable from noise. Patterns may never coalesce. Years of observation may yield only static.
This uncertainty is not a flaw in the design. It is the design. A system that guaranteed interesting results would be a system that shaped behavior toward interestingness. We have chosen not to shape. The cost of that choice is that we cannot know what, if anything, we will find.
The promise is not that answers will come. The promise is that if answers exist, this is how they might become visible.