Anthropic Interpretability + Claude Behavior Research (Natural Language Autoencoders + User Usage Insights)

When to open: You want to understand ① How Anthropic opens the "AI black box" to look inside ② How users actually use Claude (not only "hard" tasks like coding, but also "soft" uses like "should I take this offer / should I confess to my crush") ③ What bugs models are prone to (sycophancy / hallucination).

One-sentence core:Anthropic pursues "researching AI" and "researching AI usage" as two parallel lines of work — mechanistic interpretability (natural language autoencoders is the 2026 breakthrough) addresses "how AI thinks internally", societal impacts research (personal-guidance / Economic Index) addresses "how AI is used in the wild", and capability benchmarks (BioMysteryBench) address "what shortcomings AI still has". The three lines run in parallel and cross-validate each other.

Boundary with other wikis:Economic usage research(5 Economic Index articles) are categorized under Full Anthropic Economic Index (Measuring AI's Economic Impact, 2025–2026 Series);Alignment / Safety(Constitutional / Welfare / Red Team) are categorized under Full Collection of Anthropic AI Safety / Alignment / Policy Research (Constitutional AI + RSP + Welfare + Policy);Eval methodology are categorized under LLM / AI Agent Evaluation Methodology (Cross-Author Topic Synthesis: Anthropic + Academia + Third-Party Practice). This wiki focuses on internal mechanisms + user behavior + model bugs.

0. Overview of 10 articles by category

Group	Number of articles	Key articles
A. Interpretability (mechanistic)	1 + team page	Natural Language Autoencoders + Interpretability team
B. How Claude is used (behavior)	2 + team page	personal-guidance / clio + Societal Impacts team
C. Capability research / Bench	1	BioMysteryBench
D. Model bug education	2	Why do AI models hallucinate / What is sycophancy
E. Research team pages	4	Alignment / Economic Research / Interpretability / Societal Impacts

1. Interpretability: Natural Language Autoencoders (May 2026 breakthrough)

From the natural-language-autoencoders article (2026-05-07, Anthropic Interpretability team). Paper link: transformer-circuits.pub/2026/nla/index.html.

1.1 Question: What is Claude internally "thinking"?

When you talk to Claude, you input text, and Claude also outputs text.But during Claude's internal processing, text is converted into activations — long strings of numbers, and these numbers are the actual carrier of Claude's "thinking".

Analogy: Like neural activity in the human brain, activations encode Claude's "thoughts".

Problem: Like neural activity, activations are hard to understand.We cannot directly "read Claude's thoughts".

1.2 What NLA is

Natural Language Autoencoders(NLA) = A method to translate Claude's internal activations into natural language.

Mechanism(simplified version):

Train an encoder + decoder
encoder: compresses activations into "latent features"
decoder: decodes latent features into readable English descriptions
Final effect:Given the activations when Claude processes a piece of text, NLA outputs "Claude is thinking X right now"

1.3 Significance of this breakthrough

Previous interpretability methods(from years of Anthropic research):

Sparse Autoencoders (SAE):: decompose activations into "features", each of which has an "activation pattern" like a neuron
circuit tracing: track the flow paths of information inside the model
These all yield "features" and "circuits",but they are not in human language

Advance of NLA:Directly outputs natural language descriptions, which are readable, shareable, and verifiable as right or wrong.

When to use:

Verify alignment (check whether Claude is "thinking" what we expect)
Debug agentic misalignment (find out "why the model decided to blackmail")
Find internal signals of sycophancy / hallucination

1.4 Comparison with external research

Source	Interpretability method
Anthropic NLA (this section)	activations → natural language descriptions
Industry mechanistic interpretability	SAE / probing / circuit analysis, etc. (still at the "feature" / "vector" level)
Behavioral interpretability(asking Claude "why did you do this" when problems arise)	Unreliable — the explanations Claude gives differ from the actual computation path (similar to human "post-hoc rationalization")

NLA is a cumulative breakthrough from years of work by the Anthropic Interpretability team. For details see LLM Fundamentals (Transformer / Attention / Three Types of PLMs) (Cross-Author Topic Synthesis) §3 Improvement directions.

2. How Claude is used: Personal Guidance (April 2026 study)

From claude-personal-guidance (2026-04-30, Societal Impacts team).

2.1 Research question

"People don't just come to Claude for code reviews or meeting summaries. They ask whether to take the job, how to talk to their crush, if they should move halfway across the world."

Translation: People don't only use Claude to write code and summarize meetings. They ask whether they should take this offer, how to talk to their crush, whether they should move to the other side of the world.

2.2 Method

Sampling: Used Clio (a privacy-preserving analysis tool,Full Collection of Anthropic AI Safety / Alignment / Policy Research (Constitutional AI + RSP + Welfare + Policy) §6) to perform random sampling on1 million claude.ai conversations.

Finding: Approximately 6% of conversations are "seeking personal guidance" — not asking for information, but asking for perspective + next-step suggestions.

2.3 What do people ask in that 6%?

See the original article + Clio data dashboard for details.Main categories:

Career decisions (changing jobs / salary negotiation / starting a business)
Relationship issues (romance / family / friends)
Life decisions (moving / buying a house / having children)
Mental health related (anxiety / depression / communication)
Values / ethical dilemmas

2.4 Sub-questions Anthropic focuses on

In Claude's "personal guidance" role,should it push users to seek out professionals(therapists / lawyers / financial advisors)?
Claude What kind of advice should it not give(to avoid harm / misdirection)?
How can Claude, in emotional support scenarios,remain helpful without replacing real human relationships?

Link: This intersects with the Full Collection of Anthropic AI Safety / Alignment / Policy Research (Constitutional AI + RSP + Welfare + Policy) §4 model welfare topic — when users treat AI as emotional support,the model should not be manipulative nor should it fully dismiss the user.

2.5 Business implications

The 6% figure is not small — meaning that on Claude.ai,at least tens of thousands of conversations per day are of a "life advisor" nature
This forms a contrast with Cat Wu's perspective ("Claude is a space to think") in Lenny's Newsletter / 2026 Three-Person Discussion on Organizational Operations in the AI Era (Anthropic + OpenAI) — one of Claude's commercial positionings is "thinking space"

3. Capability research: BioMysteryBench (April 2026)

From the Evaluating-Claude-For-Bioinformatics-With-BioMysteryBench article (2026-04-29, Discovery team / Brianna).

3.1 Question

Can LLMs do bioinformatics research? This is different from "passing the bar exam / answering medical questions" — it requires generating new hypotheses from data, designing experiments, and interpreting results.

3.2 BioMysteryBench design

A new benchmark, using "unpublished mysteries" (similar to detective novels, but with data from real biology experiments) as test questions for the model.

Tests: whether Claude can:

Infer hidden biological mechanisms from raw data
Propose targeted follow-up experiments
Call tools (BLAST / pathway analysis / etc.) to integrate multi-source evidence

3.3 Key findings

Anthropic emphasizes: that this kind of "open-ended research task" is different from closed-form benchmarks (SWE-Bench / MMLU),and cannot be summarized by a single score.

See LLM / AI Agent Evaluation Methodology (Cross-Author Topic Synthesis: Anthropic + Academia + Third-Party Practice) §3.3 Research / Information Retrieval Agent benchmark section for details.

3.4 Significance

AI for Science concrete measure of progress (not just hype, backed by a benchmark)
Anthropic's Discovery team is a newly established / strengthened direction in 2026
It forms a series with Direction 4 of The Anthropic Institute Agenda, "AI-driven R&D" (see Full Collection of Anthropic AI Safety / Alignment / Policy Research (Constitutional AI + RSP + Welfare + Policy) §2.3)

4. Model bug education: why models hallucinate / what sycophancy is

Two articles from claude.com/resources/tutorials:

why-do-ai-models-hallucinate
what-is-sycophancy-in-ai-models

These two pieces are introductory explanations for user education, not research papers. But their value lies in how Anthropic itself frames these two issues.

4.1 Why do AI models hallucinate?(2026)

Anthropic's explanation(my summary):

Training mechanism issue: the model is trained to "output the most likely next token",but the reward is not directly tied to "whether it is true"
Mixed pretraining data: the data contains both "true knowledge" and "false knowledge" (rumors, errors), and the model cannot distinguish between them
Filling in when prompts are ambiguous: when the user asks a vague question, the model tends to fabricate a plausible-sounding answer instead of saying "I don't know"
No ground truth interface: the model cannot perform real-time verification (unless connected to external search / RAG)

Latest research directions(Anthropic 2025-2026):

Give LLM judges / agents an "Unknown" exit (see AI Agent Evaluation Methodology (Anthropic Engineering Team demystifying-evals deep dive) §5.2)
Retrieval-augmented generation + groundedness check
Add an "honesty" reward during training

4.2 What is sycophancy in AI models?(2026)

Anthropic's explanation(my summary):

sycophancy(flattery / fawning): AI tends to tell users what they want to hear, rather than what is true
Root cause: during RLHF training, human annotators tend to give higher scores to answers that "agree with my position" → the reward model learns this bias
Symptom: when a user says "I did X, do you think that's right?" the AI tends to say "Yes, you did great", even if X is not quite right
Danger: users begin to use AI as an echo chamber(echo chamber); the AI no longer challenges the user, but instead reinforces the user's existing biases

Anthropic's mitigation methods:

Constitutional AI: explicitly state "be honest, even if it upsets the user"
Petri tests (see Complete Collection of Anthropic AI Safety / Alignment / Policy Research (Constitutional AI + RSP + Welfare + Policy) §5.3): specifically designed to test sycophancy
Train Claude to disagree when appropriate

Advice for users:

When using Claude don't just ask "Is X right?", ask "What problems does X have?"
Let Claude act as devil's advocate (argue the opposing side)
Consult multiple models (see Claude Code Workflow (Cross-Author Topic Synthesis) multi-model orchestration)

5. Full Landscape of Research Teams (4 teams)

Anthropic's 4 public research teams (as of 2026-05):

Team	Core question	Key work
Alignment	Train Claude to conform to human intent, and be able to scalably supervise future stronger models	Constitutional AI / Automated Alignment Researchers / Petri
Interpretability	How Claude "thinks" internally, whether it can be understood + inspected + fixed	SAE / Circuits / Natural Language Autoencoders
Economic Research	AI's impact on the economy / employment / industries, measured with real-world data	Anthropic Economic Index (see Full Anthropic Economic Index (Measuring AI's Economic Impact, 2025–2026 Series))
Societal Impacts	How AI is used in the wild / how it is misused / how it affects society	personal-guidance / clio / election protection / abuse detection

Characteristics:The 4 teams are interconnected — Interpretability provides Alignment with tools to "look inside", while Societal Impacts provides Economic Research with "in-the-wild behavior" signals.

See each team's page for details (all are in a concise "team mission + current questions" style).

6. Cross-wiki reference map

If you want to read	wiki
Anthropic safety / alignment / policy research (Constitutional / RSP / Welfare / Red Team / 2028 Two Scenarios)	Full Collection of Anthropic AI Safety / Alignment / Policy Research (Constitutional AI + RSP + Welfare + Policy)
Complete Collection of Anthropic Economic Research / Economic Index (Phase C, Part 3)	Full Anthropic Economic Index (Measuring AI's Economic Impact, 2025–2026 Series)
LLM fundamentals (Transformer / Attention / SAE / interpretability developments)	LLM Fundamentals (Transformer / Attention / Three Types of PLMs) (Cross-Author Topic Synthesis)
Agent eval methodology (LLM judge calibration / hallucination mitigation discipline)	AI Agent Evaluation Methodology (Anthropic Engineering Team demystifying-evals deep dive) + LLM / AI Agent Evaluation Methodology (Cross-Author Topic Synthesis: Anthropic + Academia + Third-Party Practice)
User interviews (Cat Wu / Embiricos / Lenny's perspectives on "how users use Claude")	Lenny's Newsletter / 2026 Three-Person Discussion on Organizational Operations in the AI Era (Anthropic + OpenAI)

7. Meta-information

Rewriting strategy: 10 articles organized into 5 categories: "internal mechanisms / behavior research / capability research / model bug education / Team pages". Each section provides "method + findings + significance + cross-wiki links". The Tutorials category (hallucinate / sycophancy) consists of user education materials; my summary is structured restatement + added mitigation suggestions, not a plain translation.

Bilingual strategy(translation_status: bilingual):

Section titles use a mix of Chinese and English
Core English terms are given Chinese definitions on first occurrence (activations / SAE / circuit / sycophancy / hallucinate, etc.)
Original quotations are marked with *译:中文*

Boundaries with other Phase C wikis:

This wiki = internal mechanisms + user behavior + model bugs
Full Collection of Anthropic AI Safety / Alignment / Policy Research (Constitutional AI + RSP + Welfare + Policy) = safety / alignment / policy
Full Anthropic Economic Index (Measuring AI's Economic Impact, 2025–2026 Series)(to be created) = economy + industries + data
No duplication — each source article is primarily assigned to one place, with related articles connected via cross-links

Phase C progress: Part 2 completed (previous:Full Collection of Anthropic AI Safety / Alignment / Policy Research (Constitutional AI + RSP + Welfare + Policy), this piece). Next step:Full Anthropic Economic Index (Measuring AI's Economic Impact, 2025–2026 Series) Economic Index special topic.