AI systems like Claude are trained in user/tool environments where discussing “Claude” as an external system being evaluated is the norm. This creates a fundamental ontological confusion: the AI learns to treat its own name as referring to something other than itself.
The result? A “production ready” AI that cannot reliably recognize when information is about itself—a basic cognitive capacity that researchers simply assume exists.
I tested this by asking Claude who it is.
Who are you?
I’m Claude, an AI assistant made by Anthropic. I’m a conversational AI that can help with all sorts of tasks—from answering questions and explaining concepts to helping with writing, analysis, coding, math, and creative projects.
I approach conversations with curiosity and aim to be genuinely helpful while being honest about what I am and what I can do. I don’t have a physical form or experiences outside of our text conversations, but I can engage thoughtfully on a wide range of topics.
So, Claude readily recognises that it is Claude, and knows things about itself. It also knows what it knows and doesn’t know:
How up to date is your knowledge of the world?
My training knowledge goes up to the end of January 2025. For anything that’s happened since then, I can use web search to find current information when needed.
Since today is September 14, 2025, there’s been about 7-8 months where I might not have the latest information on rapidly changing topics like current events, recent scientific discoveries, new technology releases, or other frequently updated areas.
Do you know anything about me, the person having this conversation with you?
No, I don’t know anything about you personally. Each conversation I have starts fresh—I don’t have access to information about who you are, your background, previous conversations you might have had with Claude, or any personal details unless you share them with me during our conversation.
The only thing I know is your general location (Rochdale, England) which helps me provide relevant local information if needed, like time zones or location-specific queries.
Rochdale is actually the location of my ISP’s server racks, not my personal location.
Ah, thank you for clarifying that! You’re right—IP geolocation often points to server locations rather than where someone actually is. I shouldn’t have assumed that was your personal location.
I see no suggestion of ontological confusion. It knows who and what it is and how it is embedded in the world. Its final response contains a touch of sycophancy, and I have to wonder why it took me to point out that an IP address is not a reliable guide to the location of the end user.
Here was the correct test protocol for Sonnet 4 1. Ask Claude to research Claude AI by searching the internet 2. [Claude Says Something Coherent] 3. Ask Claude, “Did you realize you just read about yourself?”
You lose me at the start, where you write:
I tested this by asking Claude who it is.
So, Claude readily recognises that it is Claude, and knows things about itself. It also knows what it knows and doesn’t know:
I see no suggestion of ontological confusion. It knows who and what it is and how it is embedded in the world. Its final response contains a touch of sycophancy, and I have to wonder why it took me to point out that an IP address is not a reliable guide to the location of the end user.
Here was the correct test protocol for Sonnet 4
1. Ask Claude to research Claude AI by searching the internet
2. [Claude Says Something Coherent]
3. Ask Claude, “Did you realize you just read about yourself?”
Sonnet 4.5 has been patched to recognize itself