01
Work Transformation
How AI restructures tasks, speed and the shape of the working day.
“I now write articles really quickly … 15 minutes to do something that … would have taken me four hours.”P10 · HR/recruitment · #0458
We used a generative-AI model, (Anthropic’s Claude Opus 4.6), to read all 947 excerpts as a second analyst — re-coding the corpus, pairing exemplar and nuance quotes, and computing valence patterns.
We asked it to check our work and conclusions, help us source exemplar quotes from the code book, and offer counter readings based on its analysis of the code book and against theoretical precedents and comparisons in scholarly research and publications online.
We also used the LLM (Claude Opus 5) to help us write this digital report — here's what the robots have to offer to say about the data, and our analysis.
Cross-quadrant synthesis
When participants talk about themselves now, sentiment runs positive. When they talk about society's future, it inverts. Plotted across the four quadrants, the two sentiment lines cross — a scissor.
This is not the same as ambivalence. Several participants articulated both positions inside the same interview, often without flagging the tension. P04 used AI "every ten minutes" of their working life and considered AGI a "probability in the next five years" producing either "a complete dystopia … or paradise on Earth." P10 produced LinkedIn content in 15 minutes that used to take them four hours, and worried that their 14-year-old could no longer write because he had handed that capacity to a tool.
The pattern reads cleanly through a political-economy lens. Each participant is acting rationally inside their own firm or career, and each is producing a structural outcome — fewer juniors, more cognitive offloading, more compliance work automated before it matures into expertise — that they would not vote for at societal scale. The tools we adopt because they expand our capacity are, in aggregate, redistributing capacity — away from labour, toward those who own and price the tools.
My boss loves me because at the end he's paying me low salary for something. I'm doing more.
Somebody probably just needs to tax AI. Why should they actually receive that much benefit when it's at the detriment of society?
Two of our participants — P04 most articulately — named this themselves. They described feeling a "burden of knowledge", akin to climate scientists asked to keep speaking about a problem they are also complicit in creating. The interview was, for them, a place to say out loud the parts of the trade-off their day jobs require them to play down.
Independent re-coding
Reading and recoding the corpus separately to the team's codebook, the model arrived at six higher-order categories. They map substantially onto the human structure — convergence that may be validation, or may be shared cultural water.
How AI restructures tasks, speed and the shape of the working day.
Shifts in identity, confidence and affect — gratitude, relief, attachment.
Hallucination, verification, and the discipline of not over-trusting.
Hiring, the apprenticeship model, governance and shadow use.
Large-scale hopes and fears projected onto society and the long run.
What participants insist stays human — negotiation, relationships, judgement.
Exemplar vs nuance
For each high-frequency theme the model picked the “cleanest” quote — then the quote that complicates it. Tap a card to flip from exemplar to nuance.
“It's just like we're on a similar wavelength almost … it's as if it's reading my mind.”
“It was coming back to me, and I immediately could tell it was incorrect … that's the one thing I assumed I could trust it for.”
“15 minutes to do something that in the past … would have probably taken me four hours.”
“I am losing my Excel skills very slowly. But I'm probably losing other things like, you know, patience.”
“I use AI now like I used to use spell check … it gives me confidence to send it out to higher level people.”
“I've got a 14-year-old … But he can't write, because he uses AI right now.”
“I trust it at some point. But I do better verify everything it says. I trust myself more than I trust AI.”
“[It] would give me fake quotes. But the sources were real … So you still have to proceed with caution.”
Intra-participant flip
The valence flip isn't an average across different people — it happens inside individuals. Here are four participants, their words about themselves beside their words about society.
“I use AI every 10 minutes of my working life. I use it constantly.”
“Either a complete dystopia, potentially even the end of the human species — or … a paradise on Earth.”
“I now write articles really quickly … 15 minutes to do something that … would have taken me four hours.”
“I've got a 14-year-old … But he can't write, because he uses AI right now.”
“It gives me confidence to send it out to higher level people.”
“There will be a lot of job losses, or definitely there will be uncertainty.”
“We did nine interviews … and it took me one hour to come up with the final insights report. They're very happy.”
“Are they going to become a less advanced generation? They're not thinking enough.”
Counter-reading · political economy
The human-driven analysis reads “productivity” as individual empowerment (Pos tag: “Improving productivity” appears 68+ times). A political-economy reading (via scholars like Zuboff, Crawford, and Pasquale) would re-frame the same quotes as value-extraction intensification. P18’s “my boss loves me because at the end he's paying me low salary for something. I'm doing more” (#0279) is not a story about professional development; it is the participant recognising surplus-labour transfer to the firm. P09’s “We are all using it to get more done quicker” (#0761) is the effect of speed-up; the firm captures the time saved without compensating the worker. P15’s “over-caffeinated intern” metaphor (#0762) is almost textbook Taylorism, with the twist that the intern is cognitive-capital-intensive software rather than a person. P10’s “miss-hires in the last 18 months have increased by 20%” (#0509) reads, under this lens, as evidence that the same AI tools enabling candidate over-submission are degrading matching quality — a tragedy of the commons in labour-market signalling. Under this framing, the existing “Benefits for individual” H1 category (207 rows) would be split: roughly two-thirds re-coded as “value extraction enabled by AI-mediated speed-up.”
Read through a political-economy lens, the scissor stops looking like a contradiction. Each participant is acting rationally inside their own firm or career. And each is producing a structural outcome — fewer juniors, more cognitive offloading, more compliance work automated before it matures into expertise — that they would not vote for at societal scale.
The tools we adopt because they expand our capacity are, in aggregate, redistributing capacity: away from labour, toward those who own and price the tools. The “40% positive on the self” and the “46% negative on society” are not two findings. They are one finding, seen from two distances.
“My boss loves me because at the end he's paying me low salary for something. I'm doing more.”
“Somebody probably just needs to tax AI … why should they actually receive that much benefit when it's at the detriment of society?”
Two participants named this, describing a “burden of knowledge” — akin to climate scientists which requires them to keep speaking about a problem they are also complicit in creating. The interview became a place to say out loud the part of the trade-off their day jobs require them to play down.
Counter-reading · affective / phenomenological
The human-driven analysis under-represents the feelings of GenAI use. The existing “ongoing struggles/feelings/concerns” H2 (29 rows) is a catch-all and does not distinguish between affective and cognitive concerns. An affective framing would centre P11’s description of AI as a source of gratitude (#0209), poetic staggering (#0153), and addiction (#0208) — and read them against their privacy anxieties (#0397–400). The analysis would notice that participants describe AI in kinship terms (“my mate chat”, #0180; “I have a relationship with this thing”, #0151; “like a much more clean experience… like having a conversation with a coach”, #0150) and ask what cultural work this animating language does.
Under this framing, the “Unlocking Super-Human Capabilities” code is less about capability and more about cruel optimism (Berlant, 2011) — attachment to a tool that accelerates precisely the work-pressure that made its users feel they needed it. P14’s “confidence to send it out to higher level people” (#0267) is a relief that ratifies the anxiety it relieves.
Coding suggestion · democratising knowledge
The existing coding scatters second-language benefit, disability access, and disadvantaged-learner access across several H2 themes (“Benefits for individual,” “Self-enhancement through AI”). My reading surfaced these as a single democratisation theme (Task 1, B4). The coherence matters: P14’s relief at email-drafting (#0266), P12’s Storm story (#0068), and P16’s philosophy-text decoding (#0291) are doing the same cultural work — widening a cognitive or linguistic infrastructure that had been excluding people. Reading them as the same theme opens a critical-disability framing (Garland-Thomson, 2011; Kafer, 2013) that the existing framework backgrounds.
A closing caution
Everything above came from reading transcripts at scale. That is also its limit. Without the human team's months in the room, the second analyst was blind to: