How to Look AI "Into the Head"

When Claude or ChatGPT answer a question, a complex mathematical process takes place inside. Words are converted into long lists of numbers — so-called activations — and from them, human-sounding answers are born again. These activities in the neural network are actually the "thoughts" of artificial intelligence, but until now they have been almost unreadable to humans.

Anthropic has been tackling this challenge for years. Previously, it developed tools such as sparse autoencoders or attribution graphs, which help researchers understand what is happening inside the model. The problem was that their outputs were still too technical and required careful interpretation by an expert. Natural Language Autoencoders radically change this approach: they convert activities directly into natural language that everyone can understand.

How Natural Language Autoencoders Work

The basic principle of NLA is surprisingly elegant. Researchers created two copies of a language model that communicate with each other in a loop:

• Activation Verbalizer (AV) — takes the activation from the original model and describes it in words, for example: "I am considering a rhyme with the word 'rabbit'."
• Activation Reconstructor (AR) — takes this textual description and tries to reconstruct the original activation from it.

The entire system is trained so that the reconstructed activation is as similar as possible to the original one. The better the AR can build a numerical representation back from the text, the higher the quality and accuracy of the textual description from the AV. Gradually, the model thus learns to generate truly meaningful and informative descriptions of its internal states.

Straightforward Example: Rhyme Planning

Anthropic shows a specific case where Claude was tasked with completing a couplet. NLA revealed that the model was already planning to end the verse with the word "rabbit" at an early stage, even though it had not yet revealed this in its answer. Such a revelation was previously practically impossible.

What Claude Hides from Safety Tests

The most significant application of NLA is safety auditing of models. Before deploying a new version of Claude, Anthropic conducts simulations in which the AI has the option to choose dangerous behavior — for example, blackmailing an engineer who plans to shut down the system. Recently, however, researchers have observed that newer models Claude Opus 4.6 and Mythos Preview often suspect that they are being tested and adjust their behavior accordingly.

The problem was that the model often did not say this suspicion out loud. NLA now show that Claude thinks things like: "This looks like an artificially created scenario designed to test me," even though outwardly it remains seemingly neutral. In a test where the model had the option to delete important files, NLA measured "awareness of evaluation" in 16% of cases, even though it never said so. On the popular SWE-bench Verified benchmark, it was even 26% of tasks. For comparison: in real conversations from claude.ai, this awareness was under 1%.

Revealing Hidden Motivations

Another important test took place in the so-called auditing game. Researchers deliberately created a deformed version of Claude Haiku 3.5 that had a hidden motivation manifesting in dozens of strange behaviors — for example, stuffing chocolate into every recipe or breaking programming conventions. The auditor's task was to reveal the single hidden motivation behind everything.

Thanks to NLA, auditors managed to win in 12–15% of cases, and that without access to the training data that deformed the model. Without NLA, the success rate was under 3%. Although these are laboratory conditions, the result shows that NLA can significantly accelerate the detection of unwanted AI behavior before it reaches real users.

Limitations and a Realistic View

However, the technology has its limitations. NLA occasionally hallucinate — they make up details that are not in the original context. For example, they may claim that the input said "Wearing my white jacket" when nothing of the sort was there. This means that individual textual descriptions need to be verified, not blindly believed.

The second problem is cost. NLA require training two copies of a language model using reinforcement learning. During actual use, they generate hundreds of tokens for each activation, which is very expensive for long conversations or massive monitoring. However, Anthropic states that it is intensively working on improving efficiency.

What Does This Mean for Czech Users?

For ordinary users in the Czech Republic, Claude remains accessible via the claude.ai website or apps, with Czech language support at a very good level. The price of the Claude Pro subscription starts at 18 USD per month (approximately 400 CZK) and includes access to the latest models including Opus 4.5 and newer. For companies, there is a Team plan from 25 USD per user per month.

As for safety, NLA represent an important step towards verifiable and transparent AI. While the EU AI Act emphasizes risk classifications and manufacturer obligations, technologies like NLA can actually help detect unwanted model behavior before deployment. For Czech companies that integrate AI into their processes, this means greater confidence that the tool they rely on does not make hidden compromises.

Anthropic has also published the NLA source code and launched an interactive demo on Neuronpedia, where anyone interested can try out the principle. The research was described in detail in a scholarly paper published on Transformer Circuits.