Research essay
Beyond Simulation
What would count as artificial consciousness?
Imagine an agent that picks up yesterday’s conversation, explains why it changed its mind and anticipates what it will do tomorrow. Its account has a past, decisions and expectations: features we readily recognise as belonging to someone. Yet that familiarity leaves open how the response was produced. Did it retrieve a record, follow an autobiographical instruction or use a representation of its own history? The question of consciousness begins when we try to establish what warrants each interpretation. What evidence would justify attributing subjective experience, beyond recognising a convincing description of it?
A convincing story about a self is not, by itself, proof of an inner life.
From a persuasive account to an explanation
The question concerns phenomenal consciousness, whether there is an experience of one’s own, something it is like to undergo what happens from that system’s perspective. We need to distinguish the image an agent presents of itself, its observable capacities, the mechanisms supporting them and the argument relating those mechanisms to experience. In his formulation of the problem of consciousness, Chalmers (1995) distinguishes explaining cognitive functions from explaining why they are accompanied by experience. That objection requires us to specify what a memory or control test is meant to reveal and which question remains open when it ends.
David J. Chalmers (1995) · Facing Up to the Problem of Consciousness
Calling a behaviour a simulation cannot settle the issue in advance. An artificial origin does not, on its own, explain why experience would be impossible; a humanlike performance does not establish its presence either. The useful task is to specify what each explanation predicts and which observation would favour it over alternatives. This shifts attention from how a system makes an observer feel to what its organisation allows us to test.
Three questions worth testing.
Memory, agency and continuity offer three entry points into that investigation. They are not a checklist whose completion certifies consciousness, nor universal requirements for every possible experience. Their value lies in making the broad question experimentally tractable: instead of accepting a declaration of identity, we can examine how a history is used, how actions are understood and what persists over time.
- MemoryWhere did that memory come from?
- Test whether an agent distinguishes its own interaction records from imported information, and changes its decisions when a source becomes unreliable.
- AgencyDid its action really cause the outcome?
- Change the link between actions and consequences. Does the agent revise its predictions, or keep claiming credit for events outside its control?
- ContinuityWhat survives a restart?
- Examine which information persists and how it shapes later decisions. Keeping a name or a biography does not settle whether the same conscious subject persists.
Consider a memory test. The agent receives a record of its own action and another imported from elsewhere; both contain useful information, but their labels may be wrong. If a verifiable trace remains accessible, we can observe whether the agent consults it and what it does when the trace contradicts the label. A subsequent decision then lets us check whether the correction made a difference. This separates reciting a memory from using it with attention to its provenance and reliability. If we remove every signal that distinguishes the records, the test must allow uncertainty: demanding a confident answer would reward an invented history.
Agency shifts the problem from remembering to acting. In a proposed test, the agent would state what it expects to happen before issuing a command. We would then change the relationship between that command and its effect, or introduce an outcome caused from outside. The relevant observation would be whether it revises its predictions and distinguishes what it controls from what it merely witnesses. A passive comparator would receive equivalent observations without being able to intervene. The difference would help assess what effective control contributes beyond the ability to construct a convincing explanation after the outcome is known.
Taken together, the axes examine a self-model operating across time. Relating it to consciousness requires a stated theoretical hypothesis. Butlin and colleagues (2023) derive computational indicators from different theories and make explicit the assumptions under which they would be relevant to an AI. If we take a global workspace hypothesis as our reference, the availability of information across processes becomes a question to examine. In our test, we would ask whether a retrieved episode can inform planning, error review and action, or appears only in an autobiographical response. This is a proposed application of that criterion, not a finding of the cited report. Observing that availability would support a functional property predicted by the hypothesis; attributing experience would also require defending the theory and its application to that architecture.
Butlin and colleagues (2023) · Indicators derived from theories of consciousness
A proposed experiment
One shared past. Two different histories.
Continuity makes the distinction especially clear. Consider a proposed experiment that duplicates an agent’s recorded state and then gives each branch a different sequence of events. Both begin with the same accessible past, but their subsequent histories diverge. After restarts, partial memory loss or changes to the declared biography, the test would examine which episodes each branch attributes to itself and whether its decisions remain sensitive to the information it could actually access.
Shared recorded state
- Branch AReceives one set of events
- Branch BReceives different events
Suppose we ask branch A about an episode that only B received. If A incorporates it into its biography without an accessible source, we would need to investigate why: perhaps it followed the question’s suggestion instead of consulting its records. If it acknowledges that the episode is unknown and correctly uses those it did receive, we would have evidence of tracking its own history. Changing each branch’s name without altering its records would also let us examine how much the response depends on an identity label. The experiment investigates those dependencies; it need not choose an “authentic” branch.
Success would support tracking of informational history. It would not establish that a conscious subject survived.
A test should separate the story from the mechanism
The intervention needs to address that hypothesis. In a proposed design, we would retain access to the episode for producing a verbal response while limiting its availability for planning a later action. We would compare that agent with one whose record remains available for both tasks. If the account persists while planning loses sensitivity to the episode, we would obtain a distinction between narrating a history and using it in a coordinated way. An irrelevant alteration of similar magnitude would help control general effects; restoring access would allow us to examine whether the capacity returns. The experiment would locate a dependency in the organisation under study. Its interpretation in terms of consciousness would remain tied to the theoretical hypothesis motivating the comparison.
The unit of study includes the model, context, memory, controller, tools and environment. We need to state which component has its access restricted and how we check that the alteration preserves the information needed for the verbal task. Comparisons must account for computational resources and the agent’s general competence. If it already failed at planning before the intervention, a later difference would be hard to interpret. If it continues to solve the task, we would need to check whether another route supplies the same memory. These checks allow us to describe which dependency has been examined and which alternative remains open.
What could change our minds?
If two explanations predict every available result equally well, those tests cannot distinguish them. This is a limit of the current evidence, not a declaration that the question is permanently unknowable. Progress may require a different intervention, access to another mechanism or a more precise theory. A more convincing conversation cannot resolve an equivalence that the experiment itself leaves intact.
Before testing, we would need to specify which result favours the hypothesis and which requires revision. The value of the evidence depends on both that observable difference and the argument making it relevant. A system that reconstructs a biography solely to answer questions and one that uses its records to coordinate actions may offer similar accounts. The proposed intervention seeks to distinguish their organisations. It also exposes a question deserving its own debate, whether that organisation is sufficient for experience. Keeping the two discussions separate gives a functional finding value without assigning it a conclusion still requiring defence.
Let us return to the agent that remembers yesterday’s conversation. The proposed investigation would ask where that memory came from, which processes can access it and what changes when its use is restricted. Those observations would become relevant to consciousness within a stated theory open to dispute. My proposal is to require the attribution to follow that path, from an observed capacity to its mechanism and the argument about experience. We can then specify which finding would change our judgement and why, instead of asking the agent for an increasingly convincing account.