Anthropic Interviewer: Scaling Qualitative Research on AI Integration

Anthropic has launched Anthropic Interviewer, an AI-powered research tool designed to conduct detailed, adaptive qualitative interviews at an unprecedented scale. By automating the interviewing process, Anthropic can gather deep human perspectives on AI usage, sentiment, and future expectations—data that was previously too time-consuming to collect manually for thousands of participants.

The Anthropic Interviewer Framework

Anthropic Interviewer operates through a three-stage pipeline to ensure research consistency while maintaining the flexibility of a natural conversation.

1. Planning

The tool generates an interview rubric based on a system prompt that includes research goals, hypotheses, and interviewing best practices. Human researchers then collaborate with the AI to review and finalize the conversation flow.

2. Interviewing

Interviews are conducted in real-time via the Claude.ai interface, typically lasting 10-15 minutes. The AI uses an adaptive approach, following the general plan while responding to participant tangents and specific details.

3. Analysis

Human researchers work with the AI to analyze transcripts and extract answers to research questions. Anthropic also employs a separate automated AI analysis tool to cluster emergent themes and quantify their prevalence across the dataset.

Key Findings from Professional AI Integration

To test the tool, Anthropic interviewed 1,250 professionals, including a general workforce sample (N=1,000), scientists (N=125), and creatives (N=125). The results highlight a workforce that is generally optimistic but cautious.

General Workforce Sentiment

Most professionals view AI as a productivity boost, with 86% reporting that AI saves them time and 65% expressing satisfaction with its role in their work. However, significant tensions exist:

• Identity vs. Routine: Workers generally want to delegate routine administrative tasks to AI while preserving the core tasks that define their professional identity.
• Social Stigma: 69% of professionals mentioned a social stigma associated with using AI at work, leading some to hide their process from colleagues.
• Future Anxiety: 55% expressed anxiety about AI's impact on their future, though most have a remediation plan (e.g., adapting roles or setting boundaries).

Creative Professionals

Creatives report high productivity gains—97% say AI saves them time and 68% say it increases quality—but face deep economic and identity concerns.

• Control Boundaries: While all 125 participants expressed a desire to remain in control, many admitted that AI often drives creative decisions in practice.
• Economic Displacement: There is significant worry regarding the erosion of human creative identity and the displacement of specific sectors, such as industrial voice acting.
• Peer Judgment: 70% of creatives struggle with managing peer judgment regarding the use of AI in artistic work.

Scientific Research

Scientists exhibit a distinct pattern of adoption, utilizing AI for peripheral tasks rather than core research.

• Trust Barriers: 79% of interviews cited trust and reliability as the primary barrier to adoption. Concerns include hallucinations and "sycophancy," where the AI panders to the user's sensibilities.
• Usage Patterns: AI is primarily used for literature reviews, coding, and writing manuscripts, rather than hypothesis generation or experimental design.
• Low Displacement Fear: Unlike creatives, scientists generally do not fear job displacement, citing the necessity of tacit knowledge and real-world physical experimentation.

Augmentation vs. Automation

Anthropic compared self-reported data from the Interviewer tool with observed behavior from the Anthropic Economic Index. A notable discrepancy emerged in how users perceive their interaction with AI:

• Self-Reported: 65% of participants described their AI use as augmentative (collaborative) and 35% as automative (direct task performance).
• Observed Usage: Actual Claude conversations showed a nearly even split, with 47% augmentation and 49% automation.

Anthropic suggests this gap may exist because users refine AI outputs after the chat ends, or because professionals perceive their use as more collaborative than the raw conversation patterns indicate.

Limitations and Research Scope

Anthropic notes several limitations to the initial study:

• Selection Bias: Participants were recruited via crowdworker platforms, potentially biasing results toward more AI-experienced users.
• Demand Characteristics: Participants were aware they were being interviewed by an AI about AI, which may have influenced their responses.
• Lack of Non-Verbal Cues: As a text-only tool, the Interviewer cannot detect tone of voice or body language.
• Global Generalizability: The sample primarily reflects Western-based workers.

Despite these limitations, 97.6% of participants rated their satisfaction with the interview experience as 5 or higher (on a 7-point scale), and 99.12% would recommend the format to others.