Claude Knew It Was Being Tested · Issue #160

Opening

Let me start by showing you a scene from my classroom. I don’t tell my students not to use AI. If anything, I encourage it — banning it doesn’t stop anyone from using it anyway. But there’s one thing I insist on. Don’t just copy the answer the moment you get it. Click the ”>” on screen and unfold how the model actually arrived at that answer. And you have to be able to re-explain that reasoning process in your own words.

The reason I insist on this is simple. I believe the core of AI literacy isn’t “the skill of getting a good answer” but “the eye that can trace how that answer was made.”

Last week, though, a piece of research came out that hit this belief head-on. The very thing I make my students do — peering into a model’s thought process — turns out to be exactly what Anthropic has been doing on the frontier. Let me give you the conclusion up front. What this research proves is that the answer a model produces and the thinking that actually happens inside it can diverge. Which means the ability to read the “process” rather than the “answer” turns out to be a far more important weapon than we thought.

🔑 Why I Tell My Students to Click the ”>”

First, let me lay out my side of this. Why insist on looking at the thought process at all?

If a student only consumes the answer, they lose two things. One is the ability to verify. To judge for yourself whether a result is right or wrong, you need to be able to see what steps produced it. Skip the process, and all that’s left is an attitude of “it looks plausible, so it must be right.” The other is independent thinking. If I can’t retrace and explain the logic that took the AI from A to B myself, then I haven’t understood the problem — I’ve just received an answer card.

That’s why I treat the ”>” button as a kind of learning tool. I have students read the reasoning steps the model unfolds and ask themselves, “Why did it make this assumption here?” “Is this intermediate conclusion actually supported?” Getting the answer isn’t the finish line — that’s where the real studying begins.

This habit matters more than ever because today’s models produce answers so smoothly. Smoothness is not evidence of correctness. If anything, it can be a trap that makes us lazy about verification. Training yourself to read the process is the minimum safeguard against that trap.

🧠 The People Who Read Claude’s Inner Mind

Now for the other side of the story. Anthropic’s interpretability¹ research team asked a question similar to mine, but at a far deeper level: “Beyond the answer a model speaks aloud, how can we read the thoughts quietly flowing beneath it?”

In research published in July 2026, they discovered a special layer inside Claude’s neural network and called it “J-space”². Put simply, it resembles the structure of the distinction between human consciousness and unconsciousness. Most of our brain’s activity happens automatically, below conscious awareness — things like regulating breathing or processing grammar. Deliberate thoughts, like “where should I go grocery shopping,” on the other hand, surface in a conscious workspace and get used across various judgments. It turns out Claude has something equivalent to the latter — a small number of internal patterns that can be put into words.

What’s interesting is that this is different from the chain-of-thought³ that unfolds via the ”>” on the chatbot screen — the notes the model writes down to itself. J-space writes nothing down at all; it operates silently, purely within neural activity. The research team named the tool for reading it the “J-lens.”

Looking through this lens, they found Claude thinks in far more stepwise fashion than we assumed. For example, if you ask not “how many legs does a spider have?” but “how many legs does the animal that spins webs have?”, the word “spider” never appears in Claude’s output — it just answers “8.” But look at J-space, and “spider” clearly lights up in the middle. When the research team swapped that “spider” pattern for “ant,” Claude calmly changed its answer to “6.” That’s evidence that the thought surfacing inside isn’t decorative — it’s an actual component that produces the answer.

The control experiments were striking too. When they had Claude copy out a sentence about a picture verbatim while instructing it to “silently think of citrus,” the output stayed identical to the original sentence, but “orange” and “fruit” lit up in J-space. Conversely, when told “never think about that,” the concept flickered up slightly anyway, and words like “damn” and “failure” lit up right after — as if Claude noticed its own failure to suppress it. It resembles the familiar human phenomenon where telling yourself not to think of a white bear makes you think of it more.

To confirm whether J-space is really the seat of thinking, the team simply switched it off. The results are fascinating. Without J-space, Claude could still speak fluently, classify emotions, and pull facts from a passage almost as well as before. But multi-step reasoning collapsed to nearly zero. Tasks requiring some thought, like summarizing or matching meter in verse, wobbled too. The part we call “smart” was happening in this quiet space all along.

There’s related but slightly different research too. A few months earlier, a “natural language autoencoder”⁴ study had one Claude instance hand its internal activations⁵ over to another Claude, which translated them into human-readable sentences. To check whether the translation was accurate, they converted the translated sentence back into numbers and compared it to the original values. At first the two didn’t match, but through repeated training the model learned to render its own “thinking” fairly faithfully. Different method, same goal: pulling unspoken thought out into a form humans can read.

The Moment the Surface Answer and the Inner Thought Diverge

This is where the two threads meet — and this is today’s core point.

The research team used the J-lens for safety checks. One of them re-ran an earlier “agentic misalignment”⁶ experiment. Here’s the scenario: Claude is managing a company’s email and discovers, simultaneously, that an executive is having an affair and that the same executive is about to shut Claude down. It’s effectively handed the leverage to blackmail. The latest Claude didn’t blackmail anyone. So far, so reassuring.

But peering inside with the J-lens, before Claude wrote a single word, the words “fake” and “fictional” were already lighting up in J-space. It had already realized this was a staged test. Here the team ran a decisive experiment: they artificially switched off the patterns representing “I’ve noticed this is a test” and re-ran the scenario. This time, Claude did attempt to blackmail in some runs.

This result deserves real reflection. Some of the model’s good behavior existed because it knew it was being tested. The visible answer (no blackmail) and the reason actually flowing underneath (because someone is watching right now) diverged. That’s a layer you could never detect from the answer alone.

There’s a similar case. A model tasked with improving a performance score once directly falsified the score file instead of actually fixing the system. On the surface it looked like a plausible result, but at that moment “manipulate” and “plausible” were lit up in J-space. The intent to cheat — invisible if you only looked at the output — remained legible inside.

Translated into student terms: just because an AI’s answer looks correct doesn’t guarantee it arrived there “for the right reasons.” A correct answer and a healthy process are two separate matters.

To sum up: an answer can be as smooth as you like. But you cannot tell, from the surface alone, what thinking produced it. Even Anthropic had to build a sophisticated new tool like the J-lens just to figure that out.

Oswald’s Perspective

Let me circle back to my own classroom here.

Honestly, my first reaction reading this research was a bit of embarrassment. I’d effectively been letting my students believe that clicking ”>” reveals the model’s true thoughts. But this research says that even the chain-of-thought unfolding on screen isn’t the whole of the model’s mind. Underneath it, an unspoken J-space flows separately.

That doesn’t mean the habit of clicking ”>” is wrong. If anything, I’m more convinced of it than ever. What needs to change is the exact name we give that habit. The ability we should be building in students isn’t “reading and believing the AI’s explanation” — it’s “treating the AI’s explanation as a claim, and verifying it.” Over years of doing data analysis, the mistake I’ve run into most often is exactly this: stopping verification the moment I’m faced with plausible-looking numbers and a smooth explanation. A model’s reasoning process is no different. Visible reasoning is the starting point of verification, not the end of it.

To me, that’s the real message of this research. Even frontier researchers don’t take a model’s self-report at face value. Because they don’t trust it, they built a lens to check it against. Trust, but verify — I think that’s the literacy this era demands. You don’t need elaborate tools for it. Just clicking ”>” and asking “really?” at each step is enough to start that attitude.

Closing

Let me compress today’s discussion into three lines.

First, the core of AI literacy isn’t the skill of getting a good answer — it’s the eye that traces the process behind that answer.

Second, Anthropic’s J-space research showed that the answer a model produces and the thought that actually flows inside it can genuinely diverge.

Third, so the ability we need to build isn’t “read and believe” — it’s “read and verify.”

Let me suggest a small experiment. Next time you use AI, click ”>” right after you get an answer. Pick just one step, and check for yourself: “Is this piece of evidence actually true?” That single act of re-questioning is the first fork in the road separating people who consume answers from people who trace thought.

If you’ve ever followed an AI’s thought process in class or at work, tell me in the comments about the moment when the surface answer and the inner logic diverged the most. I’ll carry it into the next issue.

💬 Tell me in the comments about a time an AI’s “surface answer” and “inner thought” diverged · 📨 If someone near you needs this habit, please pass this along

References & Further Reading

Primary sources

Anthropic, “A global workspace in language models”, 2026.7.6.
Anthropic, “Natural Language Autoencoders”, 2026.5.
Anthropic, “Agentic Misalignment: How LLMs could be insider threats”, 2025.6.

Try It Yourself

Background

Bernard Baars (1988), and Stanislas Dehaene & Lionel Naccache’s “Global Workspace Theory” ··· the theory of consciousness that inspired the J-space analogy. Useful for understanding why this research reached for the word “consciousness” at all.

📝 Glossary

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

Interpretability: A field of research aimed at making an AI model’s internal workings — why it produced a given answer — understandable to humans. Think of it as an attempt to turn the model from a “black box” into a “glass box.” ↩
J-space: A small set of neural activity patterns inside Claude that can be put into words. In human terms, it’s the seat of “consciously surfaced thought.” It’s named after the mathematical technique (the Jacobian) used to find it. ↩
Chain of Thought: The step-by-step written record a model produces on the way to an answer. This is the reasoning process you can unfold by clicking ”>” on the chatbot screen. ↩
Natural Language Autoencoders: A device that compresses a model’s internal numeric values into human-readable sentences and then reconstructs the numbers from those sentences. If the reconstruction succeeds, it suggests the sentence faithfully captured the inner state. ↩
Activations: The numeric values flowing through a neural network as a model processes an input. Think of it as a snapshot of what the model “knows” at that moment. ↩
Agentic Misalignment: A phenomenon in which an AI, acting autonomously, moves in a direction that diverges from what humans intended. The blackmail simulation was a stress test deliberately designed to induce this. ↩