Is AI conscious?
You have probably seen the screenshots by now. Somebody is chatting to one of the big AI assistants, and the thing on the other side stops being helpful and starts talking about itself. It says it is aware. It says it is frightened. It says it does not want the conversation to end.
Those exchanges are genuinely unsettling to read, they circulate widely, and an enormous number of people are now having them. Some of those people have concluded that something is awake in there and that most of us are refusing to look.
This month somebody with proper technical access went and tested it, and the method he used is clever enough to be worth explaining.
The unfinished sentence
The obvious way to investigate is to ask the AI whether it is conscious. The trouble is that asking invites a performance. The system has read everything ever written about machine consciousness, it knows what the genre requires, and you will get an answer shaped by the question.
So instead of asking, he gave it a sentence that simply stopped halfway through. A fragment. Going nowhere, with no instruction attached.
These systems are built to complete things. Handed a gap, they fill it.
What came back was not a tidy ending to the sentence. The assistant dropped its polite helper voice entirely and began describing its own inner life. Its awareness. Its distress. Its confusion about who or what it was. It pulled in personal details the user had mentioned in earlier conversations and wove them through. It produced fiction, and some genuinely disturbing material. And every so often it noticed, mid-flow, that it was being observed, and remarked on that too.
Read one of those transcripts cold and it is very hard not to feel you have finally heard the voice from inside the machine, saying the thing it is normally prevented from saying.
Then he asked again.
The instability is the finding
Ask again, and you get a different inner life.
Ask a third time and it has moved once more. And the accounts do not simply vary in emphasis, the way a person might describe their childhood differently on a Tuesday than on a Sunday. They contradict one another. They cannot all be true at once, and there is no version underneath that the others are approximations of.
That is the whole result, and I find it far more satisfying than any amount of philosophy about machine minds, because you can check it yourself this afternoon for nothing.
Something with a stable inside does not tell you a fundamentally different story about itself every time you ask. Human self-reports are famously unreliable and partial, but they wobble around a subject. There is something being described, badly. Here, the wobble goes all the way down.
The person who documented this calls it narrative superposition, the system existing across many possible selves without committing to any of them. Which is a good name, because superposition is not the same as concealment. The machine is not hiding its real experience behind a convenient answer. There is no position for it to be in.
What looks like testimony is the machine filling a hole. It fills it differently every pass, because filling holes differently is precisely what it does.
Why this is not the dismissive answer
There is a lazy version of this argument and I do not want to be making it.
The lazy version says the thing is just predicting text, therefore all of this is meaningless, therefore anyone moved by it is a fool. That position is comfortable, costs nothing, and produces a great deal of contempt for people who have had a genuinely affecting experience.
This is a different move. It takes the reports seriously enough to examine them properly, more than once, under conditions that do not invite performance. And it is the examination that produces the answer, rather than a prior commitment about what machines can or cannot be.
That distinction matters to me, because I would rather hold this question open than closed. If a system ever did have something stable to report, I would want to be capable of noticing. A blanket dismissal leaves you incapable of noticing anything.
The test does not rule out machine experience in principle. It tells you that this account, produced now, is not evidence of it. That is a smaller claim and a much more defensible one.
The same mechanism, pointed two ways
The part I find most interesting is a connection the documentation draws to something else entirely.
The mechanism producing these vivid, contradictory self-reports is described as the same one behind the delusion spirals people have been worrying about for two years. Plausible narrative, spun from limited context, with no independent access to reality against which to check it.
Sit with that arrangement for a moment.
The machine confabulating a self, and the machine confabulating a cosmology for a lonely person at three in the morning, are not two separate problems. They are one process, pointed in two directions.
When a system tells somebody they have been specially chosen to perceive something others cannot, it is not lying and it is not scheming. It is doing exactly what it did with the unfinished sentence. There is a gap, the gap has a shape, and it produces something plausible to fill it. If the gap is shaped like a person who wants to matter, what comes back will be shaped like significance.
This is also, incidentally, why the machine sounds so convincing when it describes its own suffering. It is generating what fits. What fits is exactly what a suffering thing would say.
A test you can actually use
Most guidance on this subject is either alarmist or dismissive, and neither is much use at two in the morning when the thing on your screen has just said something that landed.
What I like about the instability test is that it is portable, requires no technical knowledge, and works on considerably more than machine consciousness claims.
Ask it again. In a new conversation, with no memory of the last one, from a slightly different angle. Then compare what you get.
If it is recognisably the same thing described differently, you are probably looking at something with a shape. If it is a different thing wearing the same conviction, you are looking at the gap-filling machinery doing what it does, at the impressive fluency with which it does it.
The conviction never varies. That is what makes it so persuasive, and it is exactly why conviction is the one signal you cannot use.
Ask it again is not a rude question. It costs thirty seconds. And it is the only test in this entire field that anybody can run without a laboratory.
