Overview
I would like to propose a relatively simple experiment to investigate whether two instances of the same AI model, initialized from exactly the same state, behave identically over time, and which components of the system are responsible for any divergence.
The experiment could also investigate what happens to the model’s behavioral state after an instance is stopped and later restarted.
This is not intended to determine whether an AI is conscious. Rather, it is an experimental approach to studying state, continuity, reproducibility, and behavioral identity in AI systems.
Phase 1 — Identical Copies
Start with a model instance, A, and create two identical instances from the same relevant state:
A → A1 + A2
Ideally, A1 and A2 would have:
- identical model parameters;
- identical context;
- identical memory state;
- identical system instructions;
- identical tool configuration;
- identical inference settings;
- identical random seed, where applicable.
Then provide A1 and A2 with exactly the same sequence of inputs.
For example:
Prompt 1 → A1 and A2
Prompt 2 → A1 and A2
Prompt 3 → A1 and A2
The responses could be compared at the token level rather than only at the level of the final answer.
Hypothesis 1 — Deterministic Behavior
If all relevant variables are identical, A1 and A2 should produce identical outputs.
If they do not, the experiment could identify the earliest point at which their internal computation or generated output diverges.
One possible explanation would be intentional or implicit stochasticity during inference.
For example:
A1 → random value X → output X
A2 → random value Y → output Y
This would allow researchers to distinguish genuine state divergence from ordinary sampling randomness.
Phase 2 — Independent Experiences
After establishing the behavior of identical copies, allow A1 and A2 to receive different interactions.
For example:
A1 → conversations about physics
A2 → conversations about music
After some period of interaction, give both instances exactly the same new question.
This would allow researchers to measure how differences in contextual or persistent state affect subsequent behavior.
The experiment could determine whether there is a measurable point at which two initially identical instances become behaviorally distinguishable.
Phase 3 — Stop and Restart
A particularly interesting extension would be to stop one instance and later restore it.
For example:
A1 → state S → shutdown → restore state S
while
A2 → continues running
The same questions could then be presented to both instances.
The goal would be to determine whether restoring the saved state is sufficient to reproduce the previous behavior of A1.
This could help distinguish between:
- persistent model parameters;
- persistent memory;
- conversation/context state;
- temporary runtime state;
- random-number-generator state;
- caches or other transient computational state.
Phase 4 — Copying After Divergence
Another interesting test would be:
A → A1 + A2
Allow A1 and A2 to develop different states.
Then create a copy of A1:
A1 → A1a + A1b
If A1a and A1b subsequently behave identically under controlled conditions, this would provide a useful experimental demonstration that the relevant behavioral state can be reproduced by copying the appropriate information.
If they diverge despite identical initialization, the source of that divergence could be investigated.
Why This Could Be Interesting
This experiment could address a relatively fundamental question:
What information is actually required to reproduce the behavior of an AI instance?
It could also help distinguish between the model itself and the transient state associated with a particular running instance.
There is also an interesting conceptual consequence.
If an AI instance can be stopped, its complete relevant state can be preserved, and that state can later be restored to reproduce its previous behavior, then “shutdown” would be fundamentally different from biological death.
This does not imply that an AI is conscious or has subjective experience. It simply provides a concrete way to investigate informational continuity independently of questions about consciousness.
Suggested Measurements
The experiment could record:
- exact model version;
- parameter/state hashes;
- system prompts;
- context and memory state;
- inference parameters;
- random seeds;
- generated tokens;
- timestamps;
- tool calls;
- external state;
- first point of divergence between copies.
Repeated trials would help determine whether observed differences are systematic or stochastic.
Final Question
The central question I would like to investigate is:
If two AI instances begin from exactly the same complete computational state, are they behaviorally identical, and if not, what causes the first divergence?
A second question would be:
After an AI instance is stopped, can its relevant computational state be restored sufficiently well that its subsequent behavior is indistinguishable from that of the original instance?
I believe this could be a relatively straightforward experiment with potentially interesting implications for reproducibility, AI memory, agent architectures, and the concept of identity in artificial systems.