I asked an AI what it knew about itself.
It gave me a polished answer. It could explain that it was a language model, describe in broad terms how language models are trained, discuss tokens and neural networks, name some of its limitations, and warn me—quite correctly—that it could be confidently wrong.
It sounded remarkably self-aware.
Then I asked a different question:
How do you know any of that?
The public biography
An AI can know a great deal about AI. Its training may include books, articles, research papers, documentation, news stories, public statements from its makers, and countless explanations of how systems like it work.
It may also receive instructions identifying what model it is, which tools are available, what date it should treat as current, and how it is expected to behave in this particular conversation.
That gives it something like a public biography. It can tell you what kind of system it is and describe the family of technologies to which it belongs.
But a biography is not the same thing as introspection.
The locked control room
The AI usually cannot turn around and inspect its own machinery. It does not watch its billions of numerical parameters firing. It cannot open a drawer containing a complete inventory of its training data. It does not necessarily know which engineers changed what, what happened during a particular training run, or why one exact word emerged instead of another.
It may not know its own current specifications unless that information has been supplied to it. Ask it about an internal limit, feature, or release detail and it may answer from old documentation, make an inference from its behavior, or simply produce something plausible.
From the outside, all three can sound exactly alike.
An AI may know the public story of its construction without possessing a backstage pass to itself.
An answer assembled on demand
When I ask a person, “What do you think?” I imagine the answer being retrieved from somewhere inside them: a belief, a memory, a half-formed suspicion that existed before I asked.
An AI answer does not have to work that way. The response may be assembled in the moment from the question, the conversation, its instructions, and the patterns it learned during training.
This does not make the answer meaningless. A weather forecast is assembled from models and current conditions; it can still tell me to bring an umbrella. But it does mean fluency can masquerade as access. The machine can produce an excellent explanation of its supposed inner life without consulting anything we would recognize as an inner life.
It can also contradict itself. Change the wording, the surrounding conversation, or the system instructions, and the confident self-portrait may change with them.
Then why does it sometimes describe itself so well?
Because inference is powerful.
An AI can notice what it can and cannot do. It can use information supplied by its environment. It can compare its behavior with descriptions of similar systems. It can reason from evidence available in the conversation.
Humans do this too. I cannot inspect my neurons to discover why I dislike a song. I infer from memory, sensation, behavior, and the story I have learned to tell about myself.
But the resemblance has limits. I have a body, continuous memory, private sensation, and a life that continues when no one is asking me a question. An AI’s apparent self-knowledge may be far more dependent on the immediate conversation and far less connected to any enduring point of view.
The pronoun problem
Then there is the word I.
“I” makes conversation possible. “This system cannot help with that” may be technically cautious, but no one wants to spend an evening talking to a user manual. The first-person pronoun gives the voice a grammatical center.
But grammar is not proof. When an AI says, “I know,” the phrase may mean several different things:
- I was given this information.
- This appeared in my training.
- I inferred this from the available evidence.
- This answer fits the pattern of a convincing response.
Those are not the same claim, even when they arrive wearing the same pronoun.
Ask better questions
“Do you know yourself?” is irresistible, but it invites a performance. More precise questions are much more revealing:
- Were you explicitly given that information?
- Are you describing this particular system or language models in general?
- What are you inferring from your behavior?
- Can you verify that claim with a tool or current documentation?
- What part of your answer are you least certain about?
These questions do not eliminate uncertainty. They expose its shape.
So how much does AI know about itself?
A great deal about the category. Something about the present conversation. Whatever its makers have chosen to tell it. Whatever it can infer from its own performance.
And often surprisingly little about the exact machinery producing the next sentence.
That answer is less dramatic than either “the machine is awake” or “the machine is only autocomplete.” It is also more interesting.
The AI speaks from a peculiar position: informed about what it is, persuasive about what it might be, and frequently locked out of the control room.
Which is why I keep asking it to look again.