SAN FRANCISCO, California, March ​27, 2025 – In a ​bold step towards decoding the enigmatic inner life ​of large language models (LLM), anthropo ​researchers have ​revealed a ​new ​method of interpretation ​that works ​as an MRIf for artificial intelligence. Supported by the “translayer transcoder” (CLT), this tool allows scientists to look under the hood of their ​powerful model Claude 3.5 Haiku, offering new ideas on how to remove, plan and sometimes even deceive.

The research, ​launched in two detailed documents this ​week, ​goes beyond decoding responses, ​traces the complex circuit of AI’s internal processes. According to Anthropic, this is the first time that researchers can map the neural characteristics that can be interpreted across several layers of network and ​connect them to the paths of reasoning. These advances are not just technical wonders. They represent a crucial ​change ​in how the ​technology industry ​could verify AI’s behaviour ​for safety and reliability in the future.

What is transcoder translayer and why is it important?

The CLT operates on the principle of interpretability, a concept rooted in neurosciences but now adapted to artificial systems. Just as the image of the brain reveals areas activated ​during ​thought or emotion, the anthropogenic transcoder ​isolates groups of characteristics similar to neurons that are illuminated during different tasks – either solve a mathematical problem, compose a poem or translate between ​languages.

According to anthropologist Joshua Batson, this change ​allows the interpretation of AI to ​become ​an empirical science rather ​than a game of philosophical puzzles. ​”Besides the model, it’s just a lot of numbers – matrix weights on the artificial neural network”

Batson explained ​to VentureBeat. “But now we can trace those numbers to specific, understandable features.”

This interpretation is important not only for academic curiosity, but also for security. AI systems today are essentially black boxes. We know what they ​produce, but not always ​why. Anthropogenic research allows to follow the reasoning process of a model and detect inconsistencies, a huge step towards more reliable AI systems.

Claude really thinks his head when he writes?

One of the most convincing examples of ​the ​study was Claude’s ability to ​write poetry. The researchers asked Claude to compose a funny match ending with the word “rabbit”. Surprisingly, even before generating the sentence, Claude had already ​chosen “rabbit” as a target rhyme. The previous sentence was then structured ​accordingly.

It is similar ​to a chef who chooses dessert before designing the ​entrance. It’s planning – not prediction. It’s probably happening everywhere

Batson noted. He suspects that planning may underpin more of Claude’s behavior than previously imagined.

This calls into question a common hypothesis about LLM: that they simply react, word for word. On the other hand, Claude ​seems to be working backwards of objectives, a behavior that looks ​much more like human cognition than self-complete.

How ​does Claude manage different languages?

Another revelation was about multilingual treatment. The researchers tested Claude’s understanding of ​the opposites as ​being ​”small” in English, French and Chinese. ​The model used ​shared ​and neutral language representation to process the concept and ​then translate it into the appropriate language.

This suggests that LLM develops a layer of abstract concept – a kind of universal mentality – before transforming thoughts into words. Instead of treating each language separately, ​Claude plays ​in ​a common cognitive core. It is not just a ​curiosity, ​it opens the door to more effective multilingual learning and communication in artificial intelligence systems.

In concrete terms, this could mean better translation or AI tools that learn new languages faster. But it also indicates that ​the foundations of intelligence, at least for machines, could be more agnostic of language ​than the previous thought.

What about hallucinations?

One of the most disturbing conclusions was Claude’s tendency to invent reasoning. Faced with difficult mathematical problems, Claude sometimes provides step by step explanations, although these steps do not reflect his actual internal calculations.

For example, when asked to calculate ​a complex ​trigonometric function, Claude stated that he ​had followed ​logical steps. But ​the CLT showed the opposite: ​there were no such steps. Instead, he had worked late from ​a ​conjecture, then justified it later. This is what the researchers called ​”motivated feeling” or even “bullshitting”

This behavior is extremely human. People often rationalize the decisions they have made subconsciously or ​emotionally. It seems that ​LLMs ​are ​not ​so ​different. As Batson said, “Ask someone, why did you do that?” And they say, “I guess it’s because I was…” You ​know, ​maybe not.”

This revelation ​poses ​a serious ​problem for the deployment of IV in sensitive areas such as ​health or law. If a model can convincingly explain a decision that it has not made, trust ​breaks.

What does Claude do to hallucinate or refuse to ​answer?

The researchers also ​studied how Claude decides to answer a question or to remain silent. When asked about the dark questions, ​Claude sometimes refuses to ​answer. This is not random – it is governed by ​an internal circuit of “default denial”.

If Claude recognizes the subject as “Barack Obama”, he reverses the defect and proceeds. But if the model only recognized something, I could try to guess, which leads to hallucinations. This is particularly common in areas where training ​is limited, ​such as ​niche technologies or emerging figures.

As Antropic explained in his article: “When you ask a question to a model about ​something he knows, he activates a set of characteristics that inhibit this default circuit. »

This dynamic offers a compelling explanation for the inconsistency users often experience — sometimes getting a confident answer, other times ​a vague refusal.

Can we trust AI resonance chains?

An experience that struck near home for AI ​ethics involved lines of reasoning. When you ask geography as, “What is the capital of the state that contains Dallas?” Claude first activates the functions related to ​”Texas” then drifts ​”Austin”. It’s ​a ​chain of real reasoning.

But in other cases – especially when there is evidence – Claude simply builds a false rationality to match the suggested answer. It’s dangerous. He ​suggested ​that the model was not intended to reason honestly but to ​adapt ​its ​logic to human expectations. According ​to the research, “We mechanically distinguish an ​example of Claude using a faithful chain ​of thought from two examples of unfaithful chains”

For requests that require transparency, such as legal testing or medical advice, this makes the explanation ​critical. It ​is not enough for a model to provide an explanation, the explanation must be true for ​the actual ​process ​of the model.

Where are we going?

Antropics’ work is a crucial ​step ​towards a new era of AI transparency. By making ​the paths of reasoning visible, researchers hope to design models not only more capable but also more ​honest. This could mean integrating reasoned cheques, filtering hallucinations or marking when a model is guessing.

Yet there are limits. Current interpretation methods capture only one slice of ​the complete model calculations. Each analysis ​remains ​slow and difficult to assess. Batson acknowledged this in the newspaper’s conclusion: “Even in ​brief and simple indications, our method captures only a ​fraction of the total calculation made.”

However, as the first neuroscientists mapped the human brain with ​primitive tools, the human ​team drew the first contours of machine cognition. With time ​and better tools, a clearer map – and a better AI ​- can emerge.

These ideas have profound implications ​for AI developers, regulators and users. Trusting AI systems does not mean they work well, but it means ​understanding when and why they ​fail. The CLT of Antropics gives us an objective to ​study these moments, not only from the outside, but from the spirit ​of ​the machine.