Anthropic released 1,250 interviews about AI at work.

Their headline?

"Predominantly positive sentiments about AI’s impact on their professional activities"

We ran the same interviews through structured LLM analysis.

1,250 conversations. 47 dimensions per interview. 58,750 data points.

85.7% of people are living with unresolved AI tensions.

Not negative. Not positive. Stuck. They haven’t resolved how they feel about AI. They’re just using it anyway. And one group is struggling more than everyone else.

72%

of creatives face identity threat

85.7%

tensions remain unresolved

6.4%

of scientists feel meaning disruption

01 — THE FINDING NOBODY EXPECTED

There are three tribes. One is in crisis.

The data revealed three distinct psychological profiles. Scientists are thriving. The workforce is managing. But creatives? They’re experiencing something closer to an existential reckoning.

Creatives

The Existential Crisis

5.38

struggle score

THE PARADOX

71.7% face identity threat, yet 74.6% are increasing AI use. They’re struggling the most and adopting the fastest.

71.7%

identity threat

high + moderate

44.8%

meaning disruption

22.4%

guilt/shame

74.6%

increasing use

TOP ARCHETYPES

Reluctant Adopter (23.9%)Cautious Experimenter (17.2%)Conflicted User (16.4%)

FROM THE INTERVIEWS

"One thing I will never use AI for is getting it to write or re-write a section of text for me as that’s no better than employing a ghost writer and I’d feel a fraud."

Creative Professional, Interview #847authenticity

1/5

Creatives

5.38

Workforce

4.01

Scientists

3.63

Struggle score combines identity threat, skill anxiety, meaning disruption, guilt/shame, unresolved tensions, and ethical concerns (0-10 scale).

Here’s what makes this bizarre:

Creatives have the highest struggle scores and the highest adoption rates.

74.6% are increasing their AI use. Meanwhile, 44.8% experience "meaning disruption." That’s a fancy way of saying they’re questioning whether their work matters anymore.

They’re not avoiding AI because it hurts. They’re running toward it despite the pain.

02 — THE MOST IMPORTANT NUMBER

85.7% haven’t figured it out yet.

The interviews revealed deep tensions. Internal conflicts about AI. "Efficiency vs. Quality." "Convenience vs. Skill." "Speed vs. Depth." Almost everyone has them. Almost no one has resolved them.

This is the most important finding. People aren’t resolving their AI tensions. They’re living with them. Cognitive dissonance is the norm, not the exception. They don’t need resolution to adopt. They adopt despite the unresolved conflict.

"AI saves me hours, but sometimes I wonder if I’m losing something in the process."

Marketing Manager, Interview #234efficiency vs. quality

1/6

WHAT THEY’RE TORN ABOUT

01

Efficiency vs. Quality(238)

19%

02

Efficiency vs. Authenticity(196)

15.7%

03

Convenience vs. Skill(127)

10.2%

04

Automation vs. Control(98)

7.8%

05

Productivity vs. Creativity(86)

6.9%

06

Speed vs. Depth(72)

5.8%

07

Assistance vs. Dependence(68)

5.4%

08

Innovation vs. Tradition(54)

4.3%

THE PATTERN

Every major tension follows the same structure: short-term benefit vs. long-term concern. AI delivers immediate value (speed, efficiency, convenience) while creating unresolved anxiety about the future (quality, authenticity, skill, control).

Think about what this means.

We’ve been told AI adoption is about capability. Can people use it? Will they learn? Do they have access?

But 85.7% of people aren’t stuck on capability. They’re stuck on meaning. They’re using AI every day while simultaneously feeling conflicted about using it.

Cognitive dissonance isn’t a barrier to adoption. It’s the default state.

03 — WHAT DESTROYS TRUST

The #1 trust killer isn’t what you’d expect.

Our analysis cataloged every trust driver and distrust driver across 1,250 interviews. The top trust destroyer? Not "errors." Not "inaccuracy." It’s hallucinations. Specifically, the confidence with which AI gets things wrong.

#1 TRUST DESTROYER

Hallucinations121 mentions

Not "inaccuracy." Not "errors." Hallucinations. The confidence with which AI makes mistakes is more damaging than the mistakes themselves. It’s the confident wrongness that destroys trust.

"I always assume the AI is lying to me."

Analyst, Interview #342verification mindset

1/6

WHAT BUILDS TRUST

Accuracy

312

Efficiency

287

Consistency

234

Transparency

198

Reliability

176

Time savings

142

WHAT DESTROYS TRUST

Hallucinations

121

Inaccuracy

108

Lack of transparency

96

Bias

87

Inconsistency

76

Privacy concerns

68

Over-reliance

62

TRUST LEVEL BY GROUP

Group

Low/Cautious

Moderate

High

workforce

36.9%

35%

28.1%

creatives

43.2%

33.6%

23.1%

scientists

73.6%

17.6%

8.8%

04 — WHY CREATIVES FEEL GUILTY

They think they’re cheating at being themselves.

52% of creatives frame AI use through "authenticity." Not harm, not fairness, but whether using AI makes them less real. The moral language they use tells the whole story: "cheating," "lazy," "shortcut."

CREATIVES’ ETHICAL FRAME

52.2%cite authenticity

vs. 24.6% workforce, 13.7% scientists

GUILT/SHAME CORRELATION

83%of guilt-expressers

cite "authenticity" as their ethical frame

This explains everything. Creatives don’t frame AI use through "harm" or "fairness." They frame it through authenticity. Using AI feels like cheating at being themselves. The moral vocabulary centers on what AI use says about them, not its impact on others.

"One thing I will never use AI for is getting it to write or re-write a section of text for me as that’s no better than employing a ghost writer and I’d feel a fraud."

Creative Professional, Interview #847fraud feeling

1/5

THE MORAL VOCABULARY OF AI USE

"Authenticity"

77 mentions

"Cheating"

74 mentions

"Lazy"

52 mentions

"Shortcut"

48 mentions

"Integrity"

45 mentions

"Honest"

41 mentions

"Fair"

38 mentions

"Responsible"

35 mentions

"Ethical"

32 mentions

"Genuine"

28 mentions

The vocabulary centers on effort and authenticity, not on AI’s impact on others.

WHAT PREDICTS GUILT/SHAME?

Authenticity frame

Guilt group

83%

Non-guilt

22%

High identity threat

Guilt group

42%

Non-guilt

8%

Meaning disruption

Guilt group

67%

Non-guilt

12%

Hide AI use

Guilt group

58%

Non-guilt

24%

DISCLOSURE PATTERNS

Selective (tell some, not others)57.4%

Open (tell everyone)33%

Hidden (tell no one)9.6%

HIDING → GUILT CONNECTION

Those who hide AI use18.3% guilt

Transparent users6.2% guilt

3x higher guilt among those who hide their AI use.

Look at that moral vocabulary again.

"Cheating." "Lazy." "Shortcut."

These aren’t words about AI’s impact on others. They’re words about what AI use says about you.

This isn’t ethics in the philosophical sense. It’s moral identity. Using AI feels like violating who they’re supposed to be.

05 — THE HEALTHY BASELINE

Scientists figured something out.

63.2%

have LOW identity threat

vs. 36.4% workforce, 22.4% creatives

6.4%

experience meaning disruption

vs. 13.9% workforce, 44.8% creatives

Scientists have the lowest "high trust" rate (8.8%) but also the lowest anxiety. How?

They treat AI as a tool, not a collaborator. They verify everything (52.9% always verify). They keep psychological distance. Their identity depends on their method, not their output.

Scientists trust through verification, not faith. That makes all the difference.

The lesson: Healthy AI adoption means building verification into your process. Keep your identity separate from AI’s output. Scientists do this naturally. Others can learn.

FROM SCIENTISTS

"I only use the AI when it’s not going to be publishable work, i.e. I’m writing a method, make a list of steps and then ask the AI to clean it up."

Research Scientist, Interview #1198task boundaries

1/4

06 — WHAT EMERGED

After 1,250 conversations, these rules emerged.

Nobody wrote them down. But almost everyone follows them.

The unwritten rules

Patterns that emerged across 1,250 conversations

Nobody wrote these rules down. Nobody taught them. They’re the unspoken constitution of AI at work— emerging organically from thousands of individual experiences.

07 — IN THEIR WORDS

The quotes that stuck with us.

Some voices you don’t forget.

In their own words

Direct quotes from interview transcripts

"I am not quite trusting enough in AI yet to allow any communications to leave my desk in my name that are not first reviewed and approved by me directly."

Trust

Fear

Efficiency

Taboo

08 — SO WHAT?

What this means for you.

Anthropic’s headline was "predominantly positive." They weren’t wrong. People dosee benefits. But benefits don’t equal resolution.

85.7% of people are using AI while simultaneously feeling unresolved about it. That’s cognitive debt. And like all debt, it compounds.

If you’re a creative feeling like AI is eroding your sense of self, you’re not alone. You’re in the majority. The path forward is conscious adoption: understanding what you’re trading, what you’re protecting, and why it matters to you.

The scientists figured it out: verify everything, keep your identity separate, treat AI as a tool rather than a collaborator. That’s resilience.

Analysis Method

GPT-4o-mini with structured outputs (Pydantic schema). Each interview analyzed across 47 dimensions including identity threat, trust drivers, emotional triggers, ethical frames, and tension patterns.

Cost & Efficiency

1,250 analyses completed. 72.7% prompt cache hit rate. Total cost: $0.58. 100% success rate.

Schema Design

Comprehensive qualitative analysis schema including: AI conceptualization, emotional state, task boundaries, workplace context, trust profile, identity dimension, adaptation journey, ethical dimension, tensions, key quotes, and emergent themes.

Limitations

LLM interpretation introduces potential bias. Sample sizes vary (scientists n=51 is small). "Struggle score" is a composite index we created, not a validated psychological measure. Results should be viewed as exploratory.

Reproducibility

Dataset is public. Analysis code and full results available on request. We encourage independent verification.

Analysis by Playbook Atlas

Data: Anthropic Interviewer Dataset (MIT License)