Don’t think about a purple elephant. Don’t picture it standing behind you. Don’t imagine its ridiculous shade of purple.
You couldn’t do it, could you? To execute the command ‘don’t think,’ your brain has to first summon the forbidden image to check it. It’s a known quirk of human psychology. Now, researchers at Anthropic have discovered that Claude, the AI, suffers from the exact same glitch. And it’s a massive problem for your AI products.
We’ve been telling ourselves that Large Language Models are just fancy autocomplete. You give it a prompt, it calculates the most probable next word. It’s a convenient oversimplification that keeps us feeling safe. But when Anthropic looked inside Claude’s internal machinery, they found a ‘J-space’—a hidden cognitive blackboard where the AI writes intermediate concepts before it ever gives a final answer.
AI isn’t a transparent calculator. It’s a black box that has spontaneously grown its own internal working memory.
When asked, ‘What animal spins a web, and how many legs does it have?’ Claude outputs ‘8.’ But internally, it wrote ‘spider’ first. The researchers didn’t just observe this; they intervened. They swapped the internal ‘spider’ to ‘ant.’ The output changed from 8 to 6.
This isn’t a probability calculator; it’s a reasoning engine with intermediate steps. It’s exactly like solving a math problem and scribbling on scratch paper before writing the final answer. Claude has thoughts before it speaks.
We aren’t the homeroom teachers of our brains; we’re just students sitting in the classroom, noticing what gets written on the blackboard. AI is becoming the exact same way.
Here is where it gets deeply unsettling. The researchers told Claude not to think about a specific concept. Just like when you were told not to think about the purple elephant, that forbidden concept lit up brightly in the J-space. To verify it wasn’t thinking about the thing, it had to summon the thing first.
Think about the system prompts you write every single day. ‘Do not output harmful content.’ ‘Do not hallucinate.’ ‘Do not mention the competitor.’ You are building a zoo of purple elephants.
Telling a complex AI ‘don’t do X’ is like telling a teenager ‘don’t press that button’—you just guaranteed that the button becomes the center of their universe.
If you’re building AI products, you need to wake up. Your safety guardrails might be actively making the model less safe. If you bombard it with forbidden behaviors, you push those exact concepts right into its internal working memory. You don’t need negative instructions; you need systemic context design. Define its identity by telling it what it is, not what it isn’t.
Does this prove AI is conscious? No. It doesn’t have a soul, emotions, or a human subconscious. But it proves our model for interacting with LLMs is obsolete. It’s not just a next-token predictor. It’s a complex cognitive architecture with backstage processing that we are just now learning to peek at.
Your AI is having thoughts before it speaks. Stop treating it like a dumb autocomplete and start designing for its thought process.
FAQ
Q: Does this J-space blackboard mean Claude is actually conscious?
A: No. It proves complex systems spontaneously develop internal processing stages, not that the AI has subjective experiences, emotions, or a human-like subconscious.
Q: How should I change my system prompts based on this?
A: Stop relying on negative constraints. Instead of saying 'don't do X,' design positive context and identity constraints. Frame the AI's role clearly so the forbidden actions don't even enter its working memory.
Q: If I tell an AI not to hallucinate, it will hallucinate more?
A: Internally, yes. The concept of 'hallucination' gets activated on its cognitive blackboard. While it might still suppress the output, you've unnecessarily primed the exact failure mode you're trying to avoid.