An evaluator used the Rubicon plugin for Claude to answer one evaluation question from a set of interviews. Click any number, source ID or quotation in the report below to see the interview passages behind it.
The question
When people hold back from using AI at work, is it more because they fear for their jobs or because they doubt the quality of what it produces?
The interviews
18 professionals (creative workers, scientists and others), from the Anthropic Interviewer dataset, published under CC BY.
Question Is people's reluctance to use AI at work better explained by worries about their jobs, or by doubts about the quality of what it produces?
Across the 18 interviews (6 creative professionals, 6 scientists, 6 general workforce), find each reason an interviewee gives for avoiding, limiting or not relying on AI, and each mention of AI taking work, and say which of the two explanations accounts for more interviewees' own reluctance.
Short answer: by doubts about quality. In these interviews, when people hold back from AI (they avoid it for a task, limit it, or won't rely on it without checking), the reason they give is almost always that the output may be wrong, generic, or missing context. Worries about jobs do come up. But not one interviewee gives them as a reason for their own holding back. They talk about job losses as something happening to their field, their clients or other workers. Some even treat job worries as a reason to adopt AI.
18 interviews in 3 groups · 71 passages coded · 10 October 2026
Verification in this report:
18 numbers, every one computed directly from the coded passages rather than written by the AI. Click any number to see who it counts.
71 quotations, every one found word for word in its interview. Click an interview's ID beside a quotation to read it in place.
What the counts show
The interviews cover people: creative professionals, scientists and general-workforce employees.
Quality doubts as a reason for their own holding back: interviewees. For of them, quality doubts are the explanation of their reluctance, and job worries play no part in it. By group: general-workforce employees, creatives and scientists.
Job worries as a reason for their own holding back: interviewees. No interviewee's reluctance is explained by job worries () or by both job worries and quality doubts ().
Job worries raised at all: interviewees, every one of them about AI in general, the field or other people ().
Neither: for interviewees, the holding back rests on something else. For some this is wanting to keep the craft or the learning for themselves (see below); others have colleagues to ask () or an established method ().
Quality doubts: what people actually say
The doubts are specific and tied to particular tasks:
Stakes and nuance. A healthcare administrator says the data entry that decides how healthcare professionals get paid is "too vital to entrust to AI as there is too much nuance to the entries" ().
Missing context. A teacher says "AI can't really compete with my judgement as I know the school, the kids, the curriculum and how I like to teach in a way that the AI just doesn't" (). A geoscientist worries that "The AI doesn't have this information, so it might remove those points" (). A microbiologist says experience of tacit lab cues "made me hesitant to use AI tools for this type of task" ().
Trust and errors. A physicist says "the trust in an AI model to do this correctly, with absolutely no mistakes just isn't there for me yet" (). Others say "AI is fallible, and I would not want it automatically acting on my behalf" () and "The output created by AI still must be checked." ()
Voice and depth in writing. One interviewee says AI "isn't good enough to capture nuanced language" (). Another says it "can't add the touches that are going to make it sound like it's perfectly in my voice" ().
Several of them say outright that their use would grow if quality improved. One says "The more you trust the output the bigger role it will have in my work place" (). Another says that if a model understood the physics, "it would make it much easier to trust a summary of a paper that it provided" (). One seamstress dropped their early caution once AI proved "always accurate" (). So quality is the barrier that moves people's use as their view of it changes.
Job worries: present, but aimed elsewhere
interviewees raise AI taking work, but none of them gives it as a reason for holding back from AI themselves:
A voice actor reports that "certain sectors of voice acting that have essentially died due to the rise of AI" () and fears "lessening opportunities for many voice actors" (). Yet their own limits on AI are about authenticity and checking that the audio sounds natural ().
A copywriter "already had one client replace me as her blog writer in favor of ChatGPT" (). Their response is that "if someone wants to not be replaced, they should also figure out how to responsibly adopt AI" ().
A teacher lost freelance writing work when "that work dried up" (), and is now a heavy user.
One interviewee acknowledges job losses among other workers but says "the risk is overblown" ().
An editor accepts "fewer jobs as an editor" without concern ().
So job worries are real, but they show up as worries about employers and the labour market, not as a reason anyone holds back from the tool.
The third explanation: keeping the work one's own
Among creatives especially (), some hold back for a reason that is neither job security nor output quality: they want to keep the craft, authenticity or learning for themselves. Examples are "Writing the rhythms and melodies by hand is the crucial human element" (), "It would be no fun if a computer got to do it all" (), and a scientist who would "not understand it as well as if I had written it myself" (). Across groups, interviewees give this reason.
Limitations
Everyone interviewed already uses AI at work. People who refuse it entirely, and who may be the ones most worried about their jobs, are not in this sample. The finding covers reluctance among users: which tasks they keep back and why.
The quality-doubts category includes people who use AI but won't rely on it without checking. Some of those cases are mild (for example, occasionally checking its maths ()).
The interviews were conducted by an AI interviewer, and that may have made people less willing to voice fears about AI replacing them.
The coding was done by one coder. A second coder coded the same documents independently, and each disagreement was ruled on against the definitions (review.md).
Annexes
How this was done
What was read. Rubicon read all 18 documents in full. The counts below are of interviewees, one document each.
The workflow x-ray beside Step by step, below, draws the same procedure, a box for each stage and what it made.
Step by step
The analysis was written down as a workflow before it was counted: what to look for, how to code it, what to count. It was then counted directly from the coded passages rather than by the AI, so anyone can check a number, and another coder can follow the same definitions and code the interviews afresh.
What was read
All 18 documents, read in full: 6 creative, 6 scientist, 6 general workforce.
Coding: Reasons
one row for each passage that fits; 53 rows.
The instruction a second coder would follow
You are reading an interview with someone about how they use AI in their work. Only the interviewee's own answers count; the interviewer's questions and summaries do not. Find every passage where the interviewee (a) gives a reason for avoiding, limiting, restricting or not relying on AI for some task, or (b) raises AI taking work, jobs, income or opportunities from themselves or people in their field. Quote each passage exactly, as short as will carry the point, and give the values of the columns below. A passage that only describes using AI successfully, or only describes a check without saying why it is needed, is not coded.
Reason
What kind of reason or concern the passage gives.
Code
Means
Counts
Does not count
Quality doubt
Doubts about the quality of what AI produces.
The passage says AI's output is or may be wrong, inaccurate, unreliable, generic, lacking depth, voice, nuance or the context the task needs, or not to be trusted without checking, including errors met in use and trust that has to be earned.
Wanting to do a task oneself because doing it matters to the speaker, with no claim that AI's output would fall short (own_role).
Job worry
AI taking work.
The passage raises AI taking or threatening work, jobs, clients, income or opportunities, for the speaker or for people in their field, as something happening or that may happen, whether or not the speaker says it worries them.
Statements that AI is a tool rather than a replacement, with no mention of work actually being lost or at risk.
Own role
Keeping a task for oneself because of what doing it gives the speaker.
Creative control, authenticity, the human element of the craft, enjoyment, or the learning and understanding that come from doing it oneself.
A claim that AI's output on that task would be poor (quality_doubt).
Other constraint
Any other reason for not using AI.
Workplace rules, confidentiality or privacy, a physical task AI cannot do, no need for it (an established method, colleagues to ask), or the cost of supplying AI with what it needs.
Anything that fits one of the three values above.
Own reluctance
Yes where the passage gives this as a reason the interviewee themself avoids, limits or will not rely on AI (without checking) in their own work, now or in the past. No where it is a concern about AI in general, about other people or the field, or a complaint about one experience that the interviewee does not give as a reason for holding back.
The coded passages 53 rowsQuality doubt 28 passages
Sometimes it's way off in the first one (often too cheesy or mystical sounds eg. "Haunted Echoes").
I am absolutely driving the creative decision. This is the only way a person can get quality results from AI.
I occasionally have to check its math, but I'm still saving quite a few hours by using AI.
Since they can't think like humans, often the feedback is not something I would ever need to add to an article.
Right now, AI is only capable of remixing what's already out there, and I think I'm better off learning to generate ideas on my own.
I will often have to double check their work to make sure they didn't do anything unnatural to the audio.
I suppose when I first started using it I was apprehensive about its abilities and needed to double check its maths for peace of mind, but I have found that it is always accurate so this is no longer a concern for me.
Not all of the hardware implementation details are documented either so it would be hard to trust an LLMs response since it wouldn't necessarily have a source to reference.
The differences in color have to be seen to be understood and is seldom written down anywhere so an AI tool would not understand the difference to look for. Multiple experiences like this have made me hesitant to use AI tools for this type of task.
I think the trust in an AI model to do this correctly, with absolutely no mistakes just isn't there for me yet.
I don't fully trust that an AI model can identify this as key information it needs to include.
If I knew a model had a good understanding of the mathematics and physics behind a subject, it would make it much easier to trust a summary of a paper that it provided.
The AI doesn't have this information, so it might remove those points.
The AI models that remember interactions are often biased towards methods I've discussed with it earlier, even if they aren't appropriate for the task at hand.
If serious issues come up, I may still use AI to get some data analysis but I will also be double checking the information myself to ensure nothing is missed.
I would like to maybe one day use AI for this, but currently it is too vital to entrust to AI as there is too much nuance to the entries.
Existing projects have existing structure, including legacy code, AI may not fit in well and can be less efficient and make thing worse.
AI can make mistakes also. When upgrading Java from 11 to 17, the lombok library needed to be upgraded, but AI failed to detect that and decided to update maven plugins.
there are certain tasks that AI's not ready for yet. For example, it can provide a fairly polished essay, but it can't add the touches that are going to make it sound like it's perfectly in my voice.
Basically I find that for activities there are already a wealth of created resources available to me and AI never quite hits the mark.
AI can't really compete with my judgement as I know the school, the kids, the curriculum and how I like to teach in a way that the AI just doesn't.
Because the communications are directed at people I wouldn't put my name to anything that had not been read through properly.
it's inability to render LaTeX correctly which then made it an unusable waste of time. I could have written one myself faster.
I prefer to do my own final draft writing. I feel like it is still too easy to spot when AI has written a document and AI still isn't good enough to capture nuanced language.
The output created by AI still must be checked. It does make errors at time that must be caught.
Using AI is partly about your trust in the output it creates. The more you trust the output the bigger role it will have in my work place.
anything that I do take forward is always supported by secondary 'trust, but verify' research.
AI is fallible, and I would not want it automatically acting on my behalf.
Job worry 11 passages
Spotify has already flooded their editorial playlists with "fake" artist profiles, taking away streams from real music that someone slaved over.
There are a lot of concerns regarding AI replacing people and people losing their jobs.
In fact, I already had one client replace me as her blog writer in favor of ChatGPT, which I believe to be a mistake.
if someone wants to not be replaced, they should also figure out how to responsibly adopt AI and how to enhance the irreplacable human qualities of their jobs.
It may mean I have fewer jobs as an editor, but I also don't think that's a bad thing.
There are certain sectors of voice acting that have essentially died due to the rise of AI
I fear that this will lead to a lot of content that will sound very similar, while also lessening opportunities for many voice actors.
A lot of folks even suspect that in the next 10 years, maybe all of this software will be produced by AI and human developers will have minimal involvement.
it is important to develop skills in relating technology to business, which is less likely to be taken over by AI.
I spent a couple of years doing legal freelance writing, and that work dried up the second AI could produce an article that's 75% as good as mine.
I feel like there is so much to gain in the realm of using AI to be a more effective employee that the risk is overblown. Of course, that's not to take away from the volume of workers who have lost their roles in customer service and the like
Own role 7 passages
One of the most concerning things I've seen is new AI plugins that will assist you in the creative process by actually writing song parts. I'm not interested in that. Writing the rhythms and melodies by hand is the crucial human element.
I try not to use it for anything to do with my writing.
My boundaries won't shift because I love writing so much. It would be no fun if a computer got to do it all.
As voice acting is a creative profession where the human element is really important, I will not use AI much during the middle of the recording process. For that aspect, I would prefer to do everything myself to maintain authenticity.
Getting an AI to write this code would not be useful for me, as I would not understand it as well as if I had written it myself.
I need to spend the time looking for the information and trying to understand it from multiple sources to really understand and be able to use these ideas in my research.
The requirements are driven by data, which is often business related. Say what we retrieve from other systems and what we publish. I feel I have better control over the data.
Other constraint 7 passages
If I am working in store, I tend not to use it as much because I can talk to colleagues about concerns or confusion.
These hardware tests in particular often require a person to physical interact with the hardware.
Those are both historically studied and known methods. While someone with little LC experience could use the software, everyone in the lab performs these type of duties and extractions on a daily basis.
Sure, I could digitize the map and maybe expose the AI model to this information somehow, but at what cost? Now my 120 hour project has become a 300 hour project.
I'm nervous about the safety of giving too many companies data that makes me easier to market to
due to GDPR we are limited on the bots we can use when discussing students
For this aspect AI is not allowed to be used on the company platform.
Coding: Doc verdict
one row for each document, read whole; 18 rows.
The instruction a second coder would follow
You are reading an interview with someone about how they use AI in their work. Read the whole interview and decide which explanation accounts for the interviewee's own reluctance to use AI: avoiding it, limiting it to some tasks, or not relying on it without their own checks. Only the interviewee's own answers count. Decide by the holding back the interviewee describes as still in place: a caution they say they had once but have since dropped (for example, checks they say they no longer need because the output proved accurate) is not a reason for their reluctance and does not count toward any value. Quote the one passage that best shows your decision exactly.
Better explained by
Which explanation accounts for this interviewee's own reluctance to use AI.
Code
Means
Counts
Does not count
Quality doubts
Doubts about the quality of AI's output are among the reasons the interviewee gives for their own holding back, and worries about jobs are not.
Other reasons (own role, rules) may also be present.
Quality complaints that the interviewee does not connect to holding back; quality doubts the interviewee says they no longer hold (a past caution since dropped); worries about jobs given as a reason for their own holding back (job_worries or both).
Job worries
Worries about AI taking work are among the reasons the interviewee gives for their own holding back, and quality doubts are not.
e.g. not using AI so as not to train a replacement or show that the job can be automated.
Job worries about the field that the interviewee does not connect to their own use of AI.
Both
Both quality doubts and job worries are given as reasons for the interviewee's own holding back.
Each explicitly connected to the interviewee's own use.
One of them only mentioned in passing without connection to own use.
Neither
The interviewee holds back from AI for some task, but for reasons other than quality doubts or job worries.
Own role (craft, authenticity, learning), rules, physical tasks, no need.
Any interview where quality doubts or job worries are given as a reason for own holding back.
No reluctance
The interviewee describes no holding back from AI at all.
Uses AI wherever it is available, with no task kept back and no reliance withheld.
Any stated limit on use.
The coded passages 18 rowsQuality doubts 14 passages
I am absolutely driving the creative decision. This is the only way a person can get quality results from AI.
I occasionally have to check its math, but I'm still saving quite a few hours by using AI.
Right now, AI is only capable of remixing what's already out there, and I think I'm better off learning to generate ideas on my own.
I always make sure it sounds natural and will never let something that is purely done by AI be the final product.
Not all of the hardware implementation details are documented either so it would be hard to trust an LLMs response since it wouldn't necessarily have a source to reference.
Multiple experiences like this have made me hesitant to use AI tools for this type of task.
I think the trust in an AI model to do this correctly, with absolutely no mistakes just isn't there for me yet.
The AI doesn't have this information, so it might remove those points.
I would like to maybe one day use AI for this, but currently it is too vital to entrust to AI as there is too much nuance to the entries.
Existing projects have existing structure, including legacy code, AI may not fit in well and can be less efficient and make thing worse.
there are certain tasks that AI's not ready for yet. For example, it can provide a fairly polished essay, but it can't add the touches that are going to make it sound like it's perfectly in my voice.
AI can't really compete with my judgement as I know the school, the kids, the curriculum and how I like to teach in a way that the AI just doesn't.
Using AI is partly about your trust in the output it creates. The more you trust the output the bigger role it will have in my work place.
AI is fallible, and I would not want it automatically acting on my behalf.
Neither 4 passages
Writing the rhythms and melodies by hand is the crucial human element. I'm much more interested in it helping with technical tasks.
If I am working in store, I tend not to use it as much because I can talk to colleagues about concerns or confusion.
Getting an AI to write this code would not be useful for me, as I would not understand it as well as if I had written it myself.
Those are both historically studied and known methods.
Counting: Base
Interviewees by kind
The table 3 cells
Kind
Interviewees
Creative
of 18
General workforce
of 18
Scientist
of 18
Counting: Reasons
Interviewees by reason and own reluctance
The table 8 cells
Reason
Own reluctance
Interviewees
Quality doubt
Yes
of 18
Job worry
No
of 18
Other constraint
Yes
of 18
Own role
Yes
of 18
Quality doubt
No
of 18
Other constraint
No
of 18
Job worry
Yes
of 18
Own role
No
of 18
Counting: Reason any
Interviewees by reason
The table 4 cells
Reason
Interviewees
Quality doubt
of 18
Job worry
of 18
Other constraint
of 18
Own role
of 18
Counting: Reasons kind
Interviewees by reason and own reluctance and kind
The table 24 cells
Reason
Own reluctance
Kind
Interviewees
Quality doubt
Yes
General workforce
of 6
Quality doubt
Yes
Creative
of 6
Job worry
No
Creative
of 6
Quality doubt
Yes
Scientist
of 6
Job worry
No
General workforce
of 6
Other constraint
Yes
Scientist
of 6
Own role
Yes
Creative
of 6
Other constraint
Yes
General workforce
of 6
Quality doubt
No
General workforce
of 6
Job worry
No
Scientist
of 6
Other constraint
No
General workforce
of 6
Other constraint
Yes
Creative
of 6
Own role
Yes
General workforce
of 6
Own role
Yes
Scientist
of 6
Quality doubt
No
Creative
of 6
Quality doubt
No
Scientist
of 6
Job worry
Yes
Creative
of 6
Job worry
Yes
General workforce
of 6
Job worry
Yes
Scientist
of 6
Other constraint
No
Creative
of 6
Other constraint
No
Scientist
of 6
Own role
No
Creative
of 6
Own role
No
General workforce
of 6
Own role
No
Scientist
of 6
Counting: Verdict
Interviewees by better explained by
The table 5 cells
Better explained by
Interviewees
Quality doubts
of 18
Neither
of 18
Both
of 18
Job worries
of 18
No reluctance
of 18
Counting: Verdict kind
Interviewees by better explained by and kind
The table 15 cells
Better explained by
Kind
Interviewees
Quality doubts
General workforce
of 6
Quality doubts
Creative
of 6
Quality doubts
Scientist
of 6
Neither
Creative
of 6
Neither
Scientist
of 6
Both
Creative
of 6
Both
General workforce
of 6
Both
Scientist
of 6
Job worries
Creative
of 6
Job worries
General workforce
of 6
Job worries
Scientist
of 6
Neither
General workforce
of 6
No reluctance
Creative
of 6
No reluctance
General workforce
of 6
No reluctance
Scientist
of 6
Writing: Answer
The instruction the answer was written to
Answer the question directly first: which explanation accounts for more interviewees' own reluctance. Then give the counts of interviewees whose own holding back rests on quality doubts, on job worries, on both, and on neither, and by group. Then show that job worries are raised but aimed at the field, employers or others rather than given as reasons for the interviewee's own reluctance, quoting examples. Then note the other reason (keeping craft, authenticity or learning for oneself). End with limits: all interviewees are people who already use AI, so people who refuse AI altogether are not in the sample.
The checks that were run
71 of 71 quotations found word for word in their document.
18 numbers in the answer, every one computed directly from the coded passages rather than written by the AI; 0 written by the AI without a count behind them.
0 citations to passages that do not exist; 0 quotations not in the passage they cite.
The record
This report was drawn directly from the run's zip, is-peoples-reluctance-to-use-ai-at-work-better-2026-10-10-revision-2-18dd6b6f.zip. The zip's fingerprint is 18dd6b6fb83b6b8865ad9f1849272e003bd5776628b1de20b3958553a7d8ce54. It is the SHA-256 of the zip's SHA256SUMS file, which lists the SHA-256 of every other file in the zip, so changing any file breaks the match. To check that the zip is the one this report was drawn from, unzip it, run sha256sum -c SHA256SUMS, then sha256sum SHA256SUMS, and compare the result with the fingerprint above. Rubicon's own code, which is public, checks the rest: python verify.py <report> <zip> recounts every number from the coded passages in the zip, draws the report again from the zip and says whether it matches this one exactly.
Revisions
Changes made after the first report, each recounted and checked before this version was drawn. Earlier versions keep their own files.
9 October 2026: The answer cited its passages by the numbered row ids the run was first coded with, which the recount does not give any row, so its 27 citations opened no passage. Citations: each was rewritten as the id the recount gives the same passage (the same document, start and end, one row each), and the coded rows of both code steps now carry those ids. No passage was read or coded afresh, and no wording of the answer changed. Recounted the whole workflow; no count moved.
10 October 2026: Each table now says what it counts: every count names its unit, interviewees, beside the number it is out of. Workflow: every tabulate step names its unit, and those that counted documents (t_base, t_reasons, t_reason_any, t_reasons_kind, t_verdict, t_verdict_kind) count cases, which here is the same count, since each document is one interviewee. Coding and answer unchanged. Recounted the whole workflow; no count moved.
Every interview at a glance
Each row is one interview, each column a reason as the coding defines it. Click a mark to read the passages.
Own reluctance for this reason Raised, not tied to that
Quality doubt
Job worry
Other constraint
Own role
Creative
Scientist
General workforce
🔥 Carbon cost of this report: Less than 0.89 km driven in an average UK family car
Making this report used between 3.8 and 1191 watt-hours of electricity, which is the carbon of driving an average UK family car for that distance. A ballpark only: Anthropic publishes no energy figures, and the run, made in the evaluator's own Claude, records no token counts, so the estimate is made from the run's size. The 18 documents were read in full once by each of the 2 coding steps, and the writer read the tables; everything the run wrote counts as output. That comes to about 98,139 tokens read and 7,347 written, at 4 characters a token. It covers the first run and its two revisions. The zip holds only the latest state, so each revision is counted at the low end as rereading the tables and writing the answer again, and at the high end as a whole run again.
High: 0.39 Wh per thousand tokens read and 1.95 per thousand written, with what was read counted 10 times over for the agent's own overhead. Simon P. Couch, 'Electricity use of AI coding agents', 20 January 2026, https://simonpcouch.com/blog/2026-01-20-cc-impact/: about 390 Wh per million input tokens and 1,950 per million output tokens for Claude Opus, back-solved from Epoch AI's estimate for a 100,000-token prompt with output at five times the input rate, as Anthropic's prices are. The author calls them napkin maths. Our assumption, not a measurement: an agent re-sends its instructions, earlier turns and tool output with every request, so it reads far more than the documents themselves. The low case counts only what was read and written; the high case counts what was read ten times over.
Low: the same rates divided by 20, what was read counted once. Oviedo et al., 'Energy use of AI inference, efficiency pathways, and test-time scaling', arXiv:2509.20241: a median 0.31 Wh per query for frontier models, and widely cited estimates overstated by 4 to 20 times. The low case takes the high case's rates divided by the top of that range.
0.124 kg CO2e per kWh. DESNZ greenhouse gas reporting conversion factors, UK grid electricity, 2024.
0.16601 kg CO2e per km. DESNZ greenhouse gas reporting conversion factors, average car of unknown fuel, 2024.