Introducing Anthropic Interviewer: What 1,250 professionals told us about working with AI

interviewsqualitative researchresearch methodsworksocietal impacts

Anthropic introduces Anthropic Interviewer, a Claude-based tool that plans, conducts and helps analyze qualitative interviews, and tests it on 1,250 professionals (a general workforce sample and samples of creatives and scientists); participants were mostly positive about AI at work, reported stigma and anxiety, described their use as more augmentative than Anthropic's usage data show, and rated the interview format highly; the transcripts are released publicly.

Kunal Handa et al. (Anthropic) · 2025-12-04 · original

Anthropic's earlier studies of how people use Claude worked from conversation logs. Those logs show what happens inside the chat window but not what people do with Claude's output afterwards, how they feel about it, or what they want from AI in future. Asking would take many hundreds of interviews. In December 2025 Anthropic introduced Anthropic Interviewer, a tool powered by Claude that conducts such interviews automatically and feeds the results back to human researchers, and reported a first test with 1,250 professionals (introduction).

For this library the post is relevant in two ways. It describes a research instrument whose behavior is set by system prompts and a written research goal, and it reports what a large sample of workers said about AI, with an explicit comparison between what they said and what Anthropic's usage data show.

How the tool works

Anthropic Interviewer works in three stages ("How Anthropic Interviewer works"):

  1. Planning. A system prompt, written with Anthropic's user research team, gives the tool its interview methodology, the researchers' hypotheses about each sample and best practices for interview plans. From the research goal, the tool generates questions and a planned conversation flow, which human researchers then review and edit with it. The plan is meant to keep many interviews on the same questions while allowing tangents.
  2. Interviewing. A second system prompt instructs the tool in interviewing practice. Interviews ran in Claude.ai, adaptively and in real time, and lasted about 10 to 15 minutes.
  3. Analysis. A human researcher works with the tool, which takes the interview plan as input and produces answers to the research questions with illustrative quotations. Separately, Anthropic's automated analysis tool, Clio, identified emergent themes across transcripts and measured how many participants raised each.

The written research goals

The tool was made aware of the research goals through its system prompt (section "Research goals"). For the general workforce sample the goal begins "Understand how individuals integrate AI tools into their professional workflows"; the creative and science goals are worded in parallel, the science goal adding trust levels and barriers to adoption. The post does not publish the system prompts themselves. The authors say the main intention of this first study was a practical test of the tool, and that the goals nonetheless produced data worth analyzing.

The test sample

Participants were recruited through crowdworker platforms; all had a main occupation other than crowdwork ("Participants"). The sample had three parts:

  • General workforce, 1,000 people from unselected occupations, the largest groups being educational instruction (17%), computer and mathematical occupations (16%) and arts, design, entertainment and media (14%).
  • Creatives, 125 people, mostly writers and authors (48%) and visual artists (21%).
  • Scientists, 125 people, including physicists (9%), chemists (9%), chemical engineers (7%) and data scientists (6%), with more than 50 other disciplines represented.

All participants consented to analysis and to public release of their transcripts, which Anthropic published as a dataset. Interviews were supplemented by surveys on participants' backgrounds.

What it finds

General workforce

86% said AI saves them time and 65% were satisfied with the role AI plays in their work ("AI's impact in the general workforce"). 69% mentioned the social stigma that can come with using AI at work. 41% said they felt secure and that human skills are irreplaceable, while 55% expressed anxiety about AI's effect on their future; of the anxious group, 25% set boundaries on their AI use and 25% adapted their roles, and only 8% of the general workforce sample expressed anxiety with no clear plan in response. Emotional profiles were similar across occupations, combining high satisfaction with frustration.

Augmentation and automation. Using the categories from Anthropic's Economic Index, where augmentation means AI collaborating with a user on a task and automation means AI performing the task directly, 65% of interviewees described AI's main role as augmentative and 35% as automative ("Augmentation versus automation"). Anthropic's most recent usage analysis at the time showed a more even split, 47% augmentation and 49% automation (some interactions unclassified). The post lists possible explanations without choosing among them: sample differences, users refining outputs after the chat ends, use of other AI providers, gaps between self-report and practice, and professionals perceiving their use as more collaborative than it is. 48% of interviewees had considered moving toward roles that manage and oversee AI systems.

Creatives

97% said AI saved them time and 68% said it improved the quality of their work ("AI's impact on creative professions"). 70% mentioned managing peer judgment of their AI use. Economic anxiety ran through the interviews, from voice actors to composers; the post notes that Claude does not produce images, video or music, so these concerns are about AI in general. All 125 said they wanted to remain in control of their creative output, yet many described moments when AI drove creative decisions. Emotional profiles varied by discipline, with trust consistently low.

Scientists

Scientists mainly used AI for literature review, coding and writing, not for hypothesis generation or experimentation ("AI's impact on scientific work"). Trust and reliability concerns were the main barrier in 79% of interviews and technical limitations appeared in 27%; regulatory constraints, skill atrophy and cost were each raised in fewer than 10%. Scientists in general did not fear displacement, citing tacit knowledge, the human nature of research decisions, and constraints such as security processes and budgets. 91% wanted more AI help in their research; roughly a third mainly with writing, the majority across their research, including generating new ideas.

Participants' view of the interviews

After the interviews, 97.6% rated their satisfaction at 5 or higher on a 1-to-7 scale (49.6% at the top), 96.96% said the conversation captured their thoughts well, and 99.12% said they would recommend the format (Appendix).

Limitations and open questions

The authors list limitations in "Conclusions and limitations": selection bias from crowdworker recruitment, possibly toward more positive or experienced views; demand characteristics, since participants knew an AI was interviewing them about AI; a single snapshot in time; a text-only interviewer that cannot read tone or body language; the gap between self-report and actual use, which the augmentation comparison illustrates; researcher interpretation; a mainly Western sample; and a non-experimental design that cannot establish that AI caused the reported gains. The post gives no comparison with human-conducted interviews on the same questions, does not publish the system prompts or interview plans, and does not describe how the percentages from thematic clustering were validated.

Research questions

  1. Do AI-conducted interviews produce different answers from human-conducted ones? The authors name demand characteristics as a limitation. Randomizing participants to the AI interviewer or a human interviewer with the same plan, and comparing themes and sentiment, would measure the effect.
  2. How much do the system prompts shape what participants say? The tool's methodology and the researchers' hypotheses are written into its system prompts. Running the same sample with variants of the prompt, for example with and without the hypotheses, and comparing the resulting themes would show their influence.
  3. How reliable are the thematic percentages? Figures such as 69% mentioning stigma come from automated clustering. Having human coders label a random subset of the released transcripts would give an agreement rate.
  4. What explains the augmentation gap? The post lists five explanations. Linking self-reports with the same participants' consented usage data would test the leading ones directly.
  5. Does high participant satisfaction mean the interviews captured their views? Satisfaction ratings were very high. Asking participants to review an AI-written summary of their own interview and mark errors would test accuracy rather than satisfaction.
  6. Do the findings hold for Claude.ai users? This test used crowdworker-recruited professionals; the tool is intended for Claude.ai users. Results from the public pilot the post launches would show how the samples differ.

How it connects

References

  • Handa, K., Stern, M., Huang, S., Hong, J., Durmus, E., McCain, M., Yun, G., Alt, A., Millar, T., Tamkin, A., Leibrock, J., Ritchie, S., & Ganguli, D. (2025, December 4). Introducing Anthropic Interviewer: What 1,250 professionals told us about working with AI. Anthropic. https://www.anthropic.com/research/anthropic-interviewer
  • Anthropic. (2025). AnthropicInterviewer dataset. https://huggingface.co/datasets/Anthropic/AnthropicInterviewer (cited in the post; not examined for this note).