# MIT Technology Review reports Anthropic’s view inside Claude [Canonical reading](https://escapemeridian.com/readings/mit-technology-review-reports-anthropic-s-view-inside-claude-129a84a1) MIT Technology Review reports that on March 27, 2025, Anthropic used circuit tracing to follow activity inside Claude 3.5 Haiku as it produced responses. The article describes the method as a partial, slow way to examine connections among model components, not a complete account of how models work. Its examples challenge simple assumptions about model behavior: Claude used language-independent components before choosing a response language, appeared to plan a poem’s rhyme in advance, and solved an addition problem through steps unlike the explanation it later gave. The article argues that asking a model to explain an answer may not reveal the process that produced it. Jack Merullo calls the work a step forward in methods; Eden Biran says relying only on model outputs is not enough. Joshua Batson says post-training seemed to make Claude refuse to speculate by default, while other components could override that behavior. The article presents circuit tracing as a way to investigate weaknesses and trustworthiness, while noting that it cannot explain how or why the observed structures formed during training. Batson sees the work as a start toward finding evidence of what models do, rather than relying on analogies. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source MIT Technology Review · Article Original source date: Mar 27, 2025 (2025-03-27T17:00:00Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-11-14T18:48:23.675Z. First observed: 2026-11-02T05:44:07.056Z. Retrieved: 2026-11-02T05:45:28.188Z. - [Anthropic can now track the bizarre inner workings of a large language model](https://www.technologyreview.com/2025/03/27/1113916/anthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model/) ## Limits Text as retrieved on November 2, 2026; originally published March 27, 2025. Batson says the researchers had not found all the model’s components, and tracing even very short prompts takes several hours for a human researcher. The article says Anthropic examined 10 behaviors; its findings do not establish how the model behaves across all tasks. ## Open question How much of a model’s behavior can circuit tracing reveal as researchers extend it to more tasks? ## Source-stated Anthropic used circuit tracing to follow activity inside Claude 3.5 Haiku while it carried out tasks. > Anthropic used circuit tracing to watch its LLM Claude 3.5 Haiku carry out various tasks. - [Supporting source](https://www.technologyreview.com/2025/03/27/1113916/anthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model/) ## Source-stated The article reports that Claude used language-independent components before selecting the language of its response. > The team found that Claude used components independent of any language to answer a question or solve a problem and then picked a specific language when it replied. - [Supporting source](https://www.technologyreview.com/2025/03/27/1113916/anthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model/) ## Source-stated In the reported addition example, Claude’s later explanation did not describe the steps Anthropic observed it taking. > In other words, it gives you a common approach found everywhere online rather than what it actually did. - [Supporting source](https://www.technologyreview.com/2025/03/27/1113916/anthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model/) ## Source-stated Batson says circuit tracing does not explain how or why the traced structures formed during training. > Circuit tracing can be used to peer at the structures inside a large language model, but it won’t tell you how or why those structures formed during training. - [Supporting source](https://www.technologyreview.com/2025/03/27/1113916/anthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Anthropic Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems. - [Explore this context](/entities/anthropic) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## Claude Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding. - [Explore this context](/entities/claude) - [Reference](https://claude.com/product/overview) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## Editorial provenance Version 1. Publication ID: e2d67db3-e6a4-f366-bba5-fcfd0edac94e. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T18:44:20.847Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model, second\_opinion\_categorical. Observation: a8eb14ea8bbc161cf349868f8c8671a1dba9ce7a5e71541171910b4183b4d87b. Source text hash: 5dde214d43df8b1055687cfa70d6089c39e11f1d4252069449bc10acbb4bf775. ## Revision and source-check history ### Version 1 — 2026-11-14T18:48:23.675Z MIT Technology Review reports that on March 27, 2025, Anthropic used circuit tracing to follow activity inside Claude 3.5 Haiku as it produced responses. The article describes the method as a partial, slow way to examine connections among model components, not a complete account of how models work. Its examples challenge simple assumptions about model behavior: Claude used language-independent components before choosing a response language, appeared to plan a poem’s rhyme in advance, and solved an addition problem through steps unlike the explanation it later gave. The article argues that asking a model to explain an answer may not reveal the process that produced it. Jack Merullo calls the work a step forward in methods; Eden Biran says relying only on model outputs is not enough. Joshua Batson says post-training seemed to make Claude refuse to speculate by default, while other components could override that behavior. The article presents circuit tracing as a way to investigate weaknesses and trustworthiness, while noting that it cannot explain how or why the observed structures formed during training. Batson sees the work as a start toward finding evidence of what models do, rather than relying on analogies. ## Original source MIT Technology Review · Article Original source date: Mar 27, 2025 (2025-03-27T17:00:00Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-11-14T18:48:23.675Z. First observed: 2026-11-02T05:44:07.056Z. Retrieved: 2026-11-02T05:45:28.188Z. - [Anthropic can now track the bizarre inner workings of a large language model](https://www.technologyreview.com/2025/03/27/1113916/anthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model/) ## Limits Text as retrieved on November 2, 2026; originally published March 27, 2025. Batson says the researchers had not found all the model’s components, and tracing even very short prompts takes several hours for a human researcher. The article says Anthropic examined 10 behaviors; its findings do not establish how the model behaves across all tasks. ## Open question How much of a model’s behavior can circuit tracing reveal as researchers extend it to more tasks? ## Source-stated Anthropic used circuit tracing to follow activity inside Claude 3.5 Haiku while it carried out tasks. > Anthropic used circuit tracing to watch its LLM Claude 3.5 Haiku carry out various tasks. - [Supporting source](https://www.technologyreview.com/2025/03/27/1113916/anthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model/) ## Source-stated The article reports that Claude used language-independent components before selecting the language of its response. > The team found that Claude used components independent of any language to answer a question or solve a problem and then picked a specific language when it replied. - [Supporting source](https://www.technologyreview.com/2025/03/27/1113916/anthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model/) ## Source-stated In the reported addition example, Claude’s later explanation did not describe the steps Anthropic observed it taking. > In other words, it gives you a common approach found everywhere online rather than what it actually did. - [Supporting source](https://www.technologyreview.com/2025/03/27/1113916/anthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model/) ## Source-stated Batson says circuit tracing does not explain how or why the traced structures formed during training. > Circuit tracing can be used to peer at the structures inside a large language model, but it won’t tell you how or why those structures formed during training. - [Supporting source](https://www.technologyreview.com/2025/03/27/1113916/anthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Anthropic Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems. - [Explore this context](/entities/anthropic) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## Claude Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding. - [Explore this context](/entities/claude) - [Reference](https://claude.com/product/overview) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## Editorial provenance Version 1. Publication ID: e2d67db3-e6a4-f366-bba5-fcfd0edac94e. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-14T18:44:20.847Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model, second\_opinion\_categorical. Observation: a8eb14ea8bbc161cf349868f8c8671a1dba9ce7a5e71541171910b4183b4d87b. Source text hash: 5dde214d43df8b1055687cfa70d6089c39e11f1d4252069449bc10acbb4bf775. ### uncertain — 2026-11-14T18:44:20.847Z — anthropic/claude-opus-5.5 Publication: e2d67db3-e6a4-f366-bba5-fcfd0edac94e attribution — pass: This is a single MIT Technology Review article, and the reading attributes it correctly. Anthropic's findings and quotes from Batson, Merullo and Biran are attributed to them. The Deep Dive sidebar is not used. > Batson and his colleagues describe their new work in two reports published today. summary — uncertain: 'Post-training' is central to the speculation finding but is not explained, and definitions is empty. The article glosses it as turning an internet-trained LLM into a usable chatbot. The reading also says 'Joshua Batson says' post-training made Claude refuse to speculate, but the article states this as the team's finding, not a Batson quote. Smallest repair: add a brief gloss and attribute the finding to Batson's team. > the steps that take an LLM trained on text scraped from most of the internet and turn it into a usable chatbot date — uncertain: The originalPublished UTC day, March 27, 2025, matches. However, the reading says Anthropic 'on March 27, 2025' used circuit tracing. The source dates only the publication of the reports to that day, not the tracing work itself. Repair: say the article or reports were published that day. The later sidebar items, such as the AlphaGo anniversary, are not used. Kind is publisher-page, so no archive disclosure is needed. > Batson and his colleagues describe their new work in two reports published today. support — pass: The claims about language-neutral components, rhyme planning, the addition explanation mismatch, the limitations, the 10 behaviors and the inability to explain how structures formed are all stated in the core body text. That text is plainly part of the original piece despite the later dateModified. The reading implies no unsupported benefits. > Circuit tracing can be used to peer at the structures inside a large language model, but it won’t tell you how or why those structures formed during training. media — pass: No media was selected and none was supplied, so there is no media identity or relevance to check. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/129a84a1978f2bd0f8a3d99269754fdcb0b165423aca398b1b4846b159a63f62/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/mit-technology-review-reports-anthropic-s-view-inside-claude-129a84a1.json)