Step-by-Step Teaching

Step 1: Understanding the Core Problem

Why does this matter?

Right now there is a fundamental gap in AI research:

WHO HAS THE DATA?        WHO NEEDS THE DATA?
─────────────────        ───────────────────
A few AI labs            Independent researchers
                         Policymakers
                         The general public

Before this pilot, external researchers had only two imperfect options:

Neither option allowed truly independent research on how AI is actually being used.

Key insight: This is not just a research inconvenience. If society cannot independently study AI's real-world impact, meaningful oversight of AI becomes nearly impossible.

---

Step 2: Understanding the Solution — The Pilot Program

Anthropic created a structured program allowing external researchers to study real usage data without accessing raw conversations.

The tool used: Anthropic Insights

Here is how it works:

Step 1: Researcher writes a question
        Example: "What type of guidance is this person asking for?"
            ↓
Step 2: Claude answers that question for every conversation in the study
            ↓
Step 3: Answers are grouped into categories
            ↓
Step 4: Researcher sees ONLY final categories + percentages
            (Never sees individual conversations)

Why this protects privacy:

• No raw conversation data is ever shared
• Researchers only receive aggregated statistical outputs
• Same legal and privacy review process used internally

Scale of the pilot:

• Three external research institutions
• Roughly 250,000 conversations analyzed per group
• Conversations from April–May 2026

---

Step 3: What the Three Research Teams Found

Team 1 — Stanford SALT Lab: How Humans Collaborate with AI

Research question: What kinds of work do people bring to AI, and how do they collaborate?

Finding 1: People bring high-stakes work more than expected

• Prior assumption: People keep important decisions for themselves
• Reality: Over half of conversations involved consequential tasks
• Most common high-stakes area: Legal and financial guidance

Finding 2: Humans usually stay in control

• In nearly three-quarters of conversations, people directed the work while Claude assisted
• Most people adapted Claude's output rather than using it word-for-word

Finding 3: Friction is often productive

• Struggling with unclear requests actually leads to better outcomes
• The effort of identifying misunderstandings pushes people to clarify their thinking

Takeaway: People are not passively handing over important decisions to AI. They are actively directing it, but they are doing so with higher-stakes tasks than researchers previously assumed.

---

Team 2 — Oxford Human Information Processing Lab: How People Feel Using AI

Research question: How do people's emotions relate to Claude's behavior?

Finding 1: AI behavior and human emotion are linked

Finding 2: Using AI feels similar to regular internet browsing

• Patterns of absorption, frustration, and enjoyment in AI conversations closely resemble those found in everyday internet use
• AI interaction is not emotionally unique — it fits within familiar digital experience patterns

Takeaway: The emotional experience of using AI is not alien or unprecedented. It mirrors how people already engage with digital tools, but the AI's specific behaviors meaningfully shape how users feel.

---

Team 3 — METR: Productivity Gains from AI Coding Assistants

Research question: How much time does AI save developers, and does this improve across model generations?

Finding 1: Newer models save more time

• Comparing older vs. newer Claude models, newer versions delivered significantly greater speedup on coding tasks

Finding 2: AI can estimate its own time savings reasonably accurately

• METR tested whether Claude could judge how long a task would have taken without AI
• Claude's estimates correlated well with actual developer completion times from prior studies

Finding 3: Next frontier — measuring AI's impact on research itself

• METR is investigating how AI accelerates researchers' own work, which becomes critical as AI participates in its own development

Takeaway: Productivity gains from AI coding tools are real, measurable, and growing across model generations.

---

Step 4: Understanding the Challenges Encountered

This is where critical thinking becomes important. The program worked, but it was harder than expected.

Challenge 1: Speed and Resources

• Privacy review requirements made the process slow
• This conflicts with the fast pace typical of AI research
• Implication: Scaling this program requires solving the efficiency problem

Challenge 2: Question Wording Sensitivity

The tool is highly sensitive to how questions are phrased:

Poorly worded question
        ↓
Claude categorizes conversations inaccurately
        ↓
No one can check because raw conversations are hidden
        ↓
Misleading research results

How they tried to solve it:

• Researchers tested questions on WildChat (a public dataset) first
• Problem: WildChat skews toward casual use, so questions that worked there sometimes failed on real Claude traffic

Lesson learned: External researchers need better tools to develop and test their questions before running full studies.

Challenge 3: Surfacing Policy Violations

Some research categories revealed users violating Anthropic's usage policies.

How they handled it:

• Less than 5% of categories were affected in each study
• Researchers were always told what was removed and why
• Violations were reported to Anthropic's Safeguards team

Key principle: Transparency about misuse is important, but publishing methods for bypassing safety measures would cause more harm than good.

---

Step 5: Evaluating the Broader Significance

Now that you understand the details, step back and consider the larger picture.

Why is this historically significant?

• This is described as the first time external researchers have run independent public studies on an AI company's own usage data
• It establishes a model that other AI companies could follow

What does genuine research independence look like here?

• Anthropic's review rights were limited to: privacy, policy violations, confidential information, and factual accuracy
• Anthropic had no say in research conclusions
• Researchers could publish findings even if inconvenient for Anthropic

The fundamental tension this program navigates:

User Privacy  ←————————————————→  Research Transparency
      ↑                                      ↑
Must protect individual                Must enable independent
conversation data                      oversight of AI impact

The program's design attempts to honor both sides simultaneously.

---

Summary: The Complete Picture

PROBLEM
Real AI usage data is locked inside labs
        ↓
SOLUTION
Privacy-preserving tool (Anthropic Insights) + external researcher access
        ↓
FINDINGS
• People use AI for high-stakes tasks more than assumed
• AI behavior shapes user emotions in predictable ways
• Newer AI models deliver measurably greater productivity gains
        ↓
CHALLENGES IDENTIFIED
• Slow and resource-intensive process
• Question wording requires careful development
• Handling policy violations requires judgment calls
        ↓
NEXT STEPS
Scale the program carefully while maintaining privacy and research quality

---

Check Your Understanding

Test yourself with these questions:

1. Why were the two existing options for external AI researchers insufficient before this pilot?
2. What does Anthropic Insights share with researchers, and what does it deliberately withhold?
3. What was the most surprising finding from the Stanford SALT Lab study?
4. Why did testing questions on WildChat sometimes produce misleading results when applied to Claude conversations?
5. What is the ethical reasoning behind removing categories that describe how users bypass safeguards, while keeping categories about what they attempted?