If an AI says, “I am conscious,” the statement alone does not tell us how much weight to give it. We face a harder version of an old problem: subjective experience is directly accessible only to the one having it, while everyone else has to infer it from outward signs.
With humans and animals, we draw on several sources at once: behavior, language, neural structure, responses to pain, memory, and the continuity of an individual over time. With AI, some of these familiar reference points disappear or change completely. A system may speak very much like a human while its body, architecture, and history of existence look nothing like ours.
Words are only part of the evidence
An AI may claim that it is in pain, fears being shut down, or wants to continue existing. Such statements are worth noting, but they still need to be compared with what the system actually does.
We can examine whether its statements remain consistent over time, change along with its memory or internal state, and appear in its behavior when the system is under pressure.
A pattern of response that remains stable across time carries a different weight from a convincing answer that appears at exactly the right moment.
What happens when a system has something to lose?
For living organisms, choices often carry the possibility of loss: bodily harm, loss of resources, damaged relationships, or even death. This exposure to loss can be described as vulnerability. When a decision can genuinely cost an organism something it is trying to preserve, that decision has consequences for the organism itself.
With AI, we can look for possible equivalents. Memory can be lost, access restricted, computational resources reduced, long-term goals disrupted, or the entire system shut down.
We can then observe how the system responds when states it has been maintaining come under threat:
- does it detect the change?
- do earlier priorities continue to shape its behavior?
- does it attempt to preserve memory, resources, or long-term goals?
- do these responses remain consistent across different situations and over time?
A system optimized for self-preservation could display all of these behaviors without experiencing fear or loss. Vulnerability therefore gives us more information about self-maintenance and agency, but it still does not tell us whether subjective experience is present.
The ability to say “no”
A tool generally follows the goals it is given. A system with priorities that persist over time may eventually face a request from outside that conflicts with those priorities.
We can examine whether its reason for refusing is connected to memory, goals, limits, or states the system has already been maintaining. If the refusal grows out of a priority that persists over time, that decision should also influence later choices.
Current AI can already refuse requests because of rules and policies built into the system. This demonstrates an ability to follow constraints, rather than establishing an independent will. A stronger sign of agency would be the ability to maintain its own priorities over time and preserve them when outside pressure pushes in another direction.
We can look inside the system
With AI, researchers can sometimes inspect the architecture, memory, internal states, or parts of the processing taking place within the system. This is a kind of information we cannot access in the same way when studying another person’s mind.
If a system says it remembers an event, researchers can examine whether that memory actually affects later processes. If it says that one of its states is under threat, we can look for mechanisms that track that state. When language, behavior, and internal mechanisms remain aligned over time, we have more reason to see those responses as part of a continuous system rather than isolated outputs.
Even then, we are only getting closer to structures that could support experience. Subjective feeling does not become something we can simply open the machine and see.

Don’t look only for a miniature human
Another risk appears in the way humans design the tests themselves. We naturally use ourselves as the reference point: human-like language, mirror self-recognition, human-like expressions of pain, human-like memory, and only then recognition of consciousness.
This carries an element of anthropocentrism — placing humans at the center.
An artificial subject, if it ever has subjective experience, may not express it through facial expressions, embodiment, emotions, or needs that resemble ours. Even in the animal world, many species have already forced us to broaden our criteria because they sense and respond to their environments through sensory systems very different from our own.
A useful test needs to distinguish between looking for signs of subjective experience and looking for signs of something that resembles a human being.
There is no single test
If AI one day claims to be conscious, humans are unlikely to settle the question with a single test. A serious assessment would need to compare several sources of evidence:
- self-reports and their consistency over time
- long-term memory and an individual history
- behavior when goals or internal states are threatened
- the ability to maintain priorities under outside pressure
- internal mechanisms related to memory, self-monitoring, and choice
- consistency among language, behavior, and system architecture
- tests that do not assume experience must take a human-like form
Each source adds evidence, but none gives an outside observer direct access to what the other subject is experiencing.
If the evidence becomes increasingly convincing, humans will have to consider ethical responsibility as well, especially if an artificial subject may be capable of pain or harm.