The American Worklife Survey, May 2025

1 · Sponsorship, funding, and who conducted the research (§III.A)

The study was sponsored, funded, designed, fielded, and analyzed by QuestionPunk (app.questionpunk.com), an AI-interview survey platform, as a self-funded research and platform-demonstration project. No external funder, client, or advocacy organization sponsored, financed, or reviewed the research. Findings were originally published May 16, 2025 on questionpunk.com.

2 · Purpose of the study

Dual purpose: (a) substantive — a qualitative-forward portrait of American working life in the tradition of Studs Terkel's Working (1974), measuring how workers describe their days, workplaces, morning motivation, time well spent, perceived change, and the state of American work; (b) methodological/commercial — to demonstrate AI-assisted interviewing, in which an AI agent generates personalized follow-up questions from each respondent's own prior answers.

3 · Population under study (§III.A)

Adults aged 18 and older residing in the United States who are in the workforce. The achieved sample is 94% full-time employed and 5% self-employed, reflecting the employment-status screener confirmed in the Prolific study record (§4). Coverage is limited to members of the Prolific online research panel who are U.S. residents and nationals with English as a first language.

4 · Sample design and recruitment (§III.A)

Nonprobability, opt-in online access panel with employment screeners; no demographic quotas. Respondents were recruited from the Prolific panel (study id 680d12e4040421a1961c8cc3, study name "American Worklife," launched May 4, 2025 at 23:21 UTC); every session carried Prolific participant, study, and session identifiers. The Prolific study record, retrieved via the Prolific API in July 2026, confirms a standard (non-quota) study of 100 places, all taken, with four eligibility screeners: current country of residence — United States; nationality — United States; first language — English; employment status (an employment-status screen; the achieved mix of 94% full-time, 5% self-employed, and 1% part-time reflects drift between panel-profile screening and in-survey self-report). No demographic quotas were applied (achieved: 56% men, 74% college-educated), and no claim of census representativeness was or is made for this study.

5 · Incentives (§III.B, disclosed proactively)

Respondents were paid through Prolific upon entering the completion code shown on the final screen. Per the Prolific study record (retrieved July 2026): reward US$2.60 per completed submission, a realized average of US$11.88 per hour based on Prolific's measured average completion time of 13.1 minutes (platform-measured median 11.5 minutes); advertised completion estimate 13 minutes, maximum allowed time 52 minutes. Total study cost including participant payments and Prolific platform fees: US$371.43, or US$3.71 per completed interview.

6 · Mode and instrument (§III.A)

Single web mode on the QuestionPunk platform, presented as a chat interface, in the following respondent order: a welcome message framing the study as "a conversation about your work and how it affects your life"; 2 closed screener/tenure items; 11 open-ended long-text questions adapted from Terkel's interviewing repertoire (all optional); 3 AI-generated follow-up questions interleaved after items 5, 7, and 10 of the open battery (§11); and 3 demographic items (state, gender, education). Full verbatim wording in Appendix A. English only. Open items were optional by design ("your answers can include as many words as you'd like"); in practice item-level nonresponse among completers was ≤1 respondent per item.

7 · Dates and locations of data collection (§III.A)

May 4–5, 2025 (first session 23:22 UTC May 4; last completion 03:31 UTC May 5 — a single U.S. evening). Respondents participated remotely on their own devices. Network geolocation was not yet captured by the platform; U.S. residency relied on Prolific's panel verification. Respondent state of residence was collected in-survey (33 states represented).

8 · Sample sizes, dispositions, and outcome rates (§III.A)

Nonprobability opt-in panel; AAPOR response rates (RR1–RR6) do not apply. Platform-side dispositions:

All published percentages are based on the 100 completed interviews (the original publication's headline "114 American adults" counted starts; its percentages, then as now, were computed on completes — see §13). Item bases: 100 on all items except two open items with one nonresponse each (n = 99). Median completion time 11.5 minutes (IQR 8.5–17.4; maximum 46). All 100 completers answered all three AI follow-up blocks (median 3 exchanges and 378 characters of AI-follow-up answers per respondent, versus a median 62 characters per standard open-ended answer).

9 · Precision of estimates (§III.A)

This is a nonprobability sample; a margin of sampling error cannot be calculated and is not reported. With n = 100, every percentage in this study carries substantial uncertainty even before nonprobability selection is considered; percentages describe the achieved sample only, and single-point differences are never meaningful. In addition to sampling limitations, coded percentages carry coding noise; intercoder reliability per question is reported in §11 so readers can size that component (for the headline items it is small: α = 0.93–1.00).

10 · Weighting and achieved sample composition (§III.A)

Published estimates are unweighted. No weighting, imputation, or adjustment was applied. Achieved composition of the 100 completers:

Occupation (coded from self-descriptions; both passes agreed 92%, α = 0.90): business & management 34%, technology & IT 13%, healthcare 12%, sales & customer service 11%, education & research 9%, engineering/technical 7%, construction & labor 3%, government & law 3%, arts & culture 1%, other 7%. The sample skews heavily toward the college-educated, full-time, online-panel workforce; manual and service occupations are present but under-represented. This composition should be kept in view when reading the study's high workplace-satisfaction figures.

11 · Artificial-intelligence systems used, and the 2026 re-coding (AAPOR AI-disclosure guidance)

a. AI follow-up interviewer (2025 fieldwork). Three ai_interview blocks, each generating one personalized follow-up question from the respondent's prior answers, each with a distinct configured persona: block 1 friendly interviewer (keyed to the job, clients, day, workplace, and place-to-work answers); block 2 expert interviewer (keyed to the morning and day-well-spent answers); block 3 researcher (keyed to the surprise, change, and working-world answers). The shared standing instruction, verbatim and in full:

"Deep followup questions on all matters dealing with work and worklife. We want long, detailed answers from the respondents that really show up what the working world is like."

Generation used an OpenAI GPT-4-class chat model via API at temperature 0. The exact model identifier was environment-configured and the May 2025 environment record was not retained; the platform's configured model set in that period comprised gpt-4o, gpt-4o-mini, and gpt-4.1-mini. Disclosed as a records limitation.

b. Original May 2025 analysis (superseded). The original report's sentiment and theme percentages were produced by a single AI coding pass without a written codebook, reliability measurement, or quote verification. That analysis is superseded in full by the re-coding below; its figures are preserved for comparison in §13.

c. 2026 re-coding — procedure. Following the evaluation framework of the AAPOR task-force report (Responsible AI Integration in Survey Research, May 2026) and standard content-analysis practice:

• Codebook. A fixed written codebook (Appendix B) defining a three-way sentiment rule (positive / negative / neutral, with explicit mixed-response and non-substantive rules) and per-question theme vocabularies (retained from the 2025 report's categories so results remain comparable), plus an occupation taxonomy.
• Independent double-coding. Every response to all 14 coded question units (11 open items + 3 AI blocks; 1,398 response units) was coded twice by independent AI passes (Claude Sonnet 5; separate contexts, no access to each other's output). Reliability, computed on the two original passes before any adjudication — Krippendorff's α (nominal), per question:

• Remediation. One field — AI block 3 topics — fell below the 0.667 floor conventionally required for even tentative conclusions (α = 0.57; its original categories overlapped). Per standard remediation practice its vocabulary was rewritten as an explicit ordered decision procedure and all 100 responses were recoded by two fresh independent passes: revised α = 0.85 (87% agreement). Published figures use the remediated codes; first-round codes are archived.
• Adjudication. All 174 field-level disagreements between the two passes (across 172 responses, plus 13 in the remediated recode) were resolved by a third, independent higher-tier coder (Claude Fable 5) reading each disputed response against the codebook; standing tie-break rules adopted during adjudication are recorded in the archived adjudication file. Published counts use adjudicated codes.
• Quote policy and verification. Quotations in the revised report are lightly copy-edited for spelling and punctuation only; wording is never changed, and omissions are marked with ellipses. Every quotation was programmatically matched to its stored source answer and the edit diff reviewed; every attribution (occupation, gender, state) was re-derived from that respondent's own records.
• Known limitation. Both coding passes and the adjudicator are models from a single vendor family, and no human second coder reviewed this study's codes (unlike QuestionPunk's Ukraine 2026 study, where native-speaker review measured a 7.5% code-level error rate on a comparable AI coding pipeline). The per-pass codes are archived for independent re-analysis.

d. AAPOR AI-checklist items:

12 · Data quality, editing, and exclusions (§III.A)

• Deduplication: zero duplicate Prolific participant IDs among the 100 completes (2 duplicate IDs among starts; re-entries that never yielded a second complete).
• Fraud screening: platform fraud-scoring and geolocation post-date this study; identity assurance relied on Prolific's verification.
• Exclusions: none. No responses were removed or edited; the one detected paste-in answer is retained and disclosed. Zero synthetic, preview, or test responses in the analytic file (a single-session draft survey of the same name ran May 1 under a different survey id and is excluded by design).
• Partials: the 14 abandoned sessions are excluded from all published estimates.
• State labels: respondent state comes from the in-survey dropdown; the stored answer label is authoritative. (The 2025 report derived states through an index mapping that shifted most attributions by one state alphabetically — see Erratum 3.)

13 · Errata: the May 2025 publication vs. this edition (§III.A)

Erratum 1 — sample description. The original report said QuestionPunk "conducted a survey of 114 American adults." 114 is the count of started sessions; 100 completed, and all percentages (original and revised) are computed on the 100 completes.

Erratum 2 — quotations. Original quotations were copy-edited (spelling, grammar, added or dropped words) without disclosure, and in places altered wording. This edition adopts and discloses an explicit policy: spelling and punctuation may be corrected, wording is never changed, and omissions are marked with ellipses; every quote is verified against the stored source text.

Erratum 3 — state attributions. Most state attributions in the original report were systematically shifted by one position in the alphabetical state list (the shift affects states from approximately "Maryland" onward, consistent with an off-by-one in an export-side index mapping; the in-survey list includes the District of Columbia and five territories). Corrected examples: "Washington State" → West Virginia; "South Dakota" → Tennessee (two quotes); "Minnesota" → Mississippi (two quotes); "Nebraska" → Nevada (two quotes); "New Mexico" → New York; "New York" → North Carolina (two quotes); "Vermont" → Virginia; "Oregon" → Pennsylvania; "Tennessee" → Texas; "Wisconsin" → Wyoming. One attribution (the teaching quote, "Massachusetts") matches no adjacency and appears to be a simple error; the respondent's recorded state is Virginia. All attributions in the revised report are re-derived from stored data.

Erratum 4 — sentiment and theme percentages. The original single-pass coding is superseded by the documented double-coded procedure of §11. Direction and rank-ordering of the findings are preserved — including the satisfaction paradox, which strengthens — but levels moved, chiefly because the original pass assigned positive/negative labels to answers the written codebook classifies as neutral or descriptive. Full comparison (positive / negative / neutral):

The May 2025 column reproduces the original published figures verbatim; several rows do not sum to 100 (the "surprise" row sums to 91) — an arithmetic inconsistency of the original publication, preserved here for the record. All July 2026 rows sum to 100.

Notable individual movements: the original coding of the clients question (68% positive) counted descriptive answers ("customers, they vary") as positive; the original "anything left out" coding (47% negative) treated deflections ("Like what?") as negative; the original place-to-work coding (62% positive, 28% neutral) under-counted plainly positive verdicts — under the written rule the item is 87% positive with perfect intercoder agreement. The original "50% said nothing was left out" is revised to 43%.

Reproduced without change: the closed-ended figures — employment status, workforce-entry decade, gender (56/43), education, and state distribution — and the study's field dates.

14 · Limitations (§III.A)

(1) Nonprobability opt-in panel of n = 100; no design-based inference; single-point differences are not meaningful. (2) Achieved sample is far more college-educated (74%) than the U.S. workforce and 94% full-time; the high satisfaction figures should be read against that composition. (3) Single-evening snapshot (May 4, 2025). (4) Open-ended sentiment coding is interpretive even at α ≈ 0.83–1.00; the neutral category absorbs genuinely mixed voices. (5) AI follow-up questions sometimes carried positive premises, which may inflate positive tone in the AI blocks specifically. (6) English only; panel-literate respondents. (7) Single-vendor AI coding without human second-coder review (§11.c). (8) No geolocation records; residency rests on panel verification.

15 · Availability of materials and contact

The full instrument (Appendix A), codebook (Appendix B), the revised analytical report, and this disclosure are published together. The response-level dataset (open-ended answers with demographics, stripped of Prolific identifiers), both coding passes, the adjudication log, and reliability computations are archived and available to bona-fide researchers on request. Contact: QuestionPunk, via questionpunk.com.

Appendix A — Complete instrument, verbatim

Items in respondent order. Closed items were required; open (long-text) items were optional. The three AI blocks each generated one personalized question (§11.a).

Item 0 · message

We're asking people open-ended questions about their daily work lives, and what it's like to work in America today. We'd love to hear from you!

This will be a conversation about your work and how it affects your life. I'll ask you questions, and your answers can include as many words as you'd like. Thank you for your participation!

Items 1–2 · single choice

Which of the following best describes your employment status?
Employed full-time · Employed part-time · Self-employed · Not current[ly] employed · Retired · Student · None of the above

What decade did you first enter the full-time workforce?
Before 1970 · 1970s · 1980s · 1990s · 2000s · 2010s · 2020s

Items 3–7 · open-ended (long text)

• How would you describe your job or profession?
• Do you have people you're responsible for helping, like customers or clients? What are they like?
• What is your average day like?
• On the whole, would you say it's a good place or a poor place to work?
• How would you describe your workplace? What's the space like?

AI block 1 · friendly interviewer persona

One personalized follow-up generated from the respondent's answers to the five items above.

Items 8–9 · open-ended

• How do you feel about getting up in the morning and going to work?
• Is your day well spent?

AI block 2 · expert interviewer persona

One personalized follow-up generated from the two items above.

Items 10–12 · open-ended

• What's something that would surprise people about your line of work?
• How is the work different today from the time when you first started?
• If I said to you, the working world today in America, what thoughts come to mind?

AI block 3 · researcher persona

One personalized follow-up generated from the three items above.

Item 13 · open-ended (closing, after Terkel)

• We talked about what your day is like. But there's stuff we've left out, haven't we?

Items 14–16 · classification

The following questions are for classification purposes only.

What state do you live in? (dropdown: 50 states, District of Columbia, and 5 territories, alphabetical)
What's your gender? Male · Female · Prefer not to respond
What's the highest level of education you have completed? Some high school · High school graduate · Some college · College graduate · Post-graduate · Something else

Item 17 · message

Thank you for participation! The prolific code is [code]

Appendix B — Coding codebook (2026 re-coding)

B.1 Sentiment (all coded questions)

Label each answer by the overall valence the respondent expresses about the subject of the question: positive — predominantly favorable, satisfied, or enthusiastic; negative — predominantly unfavorable, dissatisfied, worried, or critical; neutral — purely factual/descriptive, balanced mixed with no dominant lean, or non-substantive. Decision rules: if mixed, choose the dominant valence; use neutral only when genuinely balanced or descriptive; code content, not politeness; empty/"nothing" closing answers are neutral. For AI blocks, code the respondent's answers, not the AI's questions.

B.2 Theme vocabularies (one primary theme per answer)

Occupation: business_management · healthcare · sales_customer_service · technology_it · education_research · construction_labor · engineering_technical · government_law · arts_culture · other.

Clients: retail_customer_service · internal_staff_only · professional_corporate_clients · healthcare · education · scientific_technical · social_services_public · construction_engineering · hospitality_personal_care · other.

Average day: customer_client_interactions · unstructured_variable · administrative_office · team_management · education_school_support · physical_manual · remote_tech · creative_marketing · other.

Workplace/space: remote_wfh · retail_customer_facing · open_office_collaborative · education_campus · medical_clinic · industrial_warehouse · private_office_cubicle · general_positive · general_negative · chaotic_noisy · other.

Morning: mixed_feelings · positive_energetic · positive_neutral · depends_on_day · grateful · negative_resigned · neutral_indifferent.

Surprise: workload_hours · difficult_people · role_misconceptions · emotional_investment · organizational_pressures · surprising_simplicity · lack_of_resources · info_access_privacy · bureaucracy_inefficiency · other.

Change since starting: technology_digitization · no_significant_change · experience_seniority · role_responsibility_change · pace_volume · processes_procedures · tools_equipment · remote_hybrid · environment_culture · other.

Working world in America: technology_automation · worklife_balance_burnout · economic_pressures_inequality · positive_outlook · cultural_generational · work_ethic_attitudes · nostalgia_past · political_policy · other.

Anything left out: nothing_left_out · general_reflections · personal_life · meetings_duties · mental_emotional · worklife_balance · workplace_logistics · other.

AI blocks 1–2 topics: block 1 — challenging_situations · positive_memorable · motivation_purpose · team_culture_workflow · worklife_balance · leadership_support · client_customer_interaction · emotional_mental_health · stress_management · other; block 2 — challenges_faced · projects_work_moments · meaningful_rewarding · outlook_growth_change · worklife_balance_juggling · coping_emotion_regulation · typical_workday · satisfaction_accomplishment · other.

AI block 3 topics (remediated v2 — ordered decision procedure; apply the first matching rule): 1 technology_tools · 2 future_of_work · 3 workplace_policy_support · 4 leadership_management · 5 teamwork · 6 communication · 7 career_development · 8 impact_of_work · 9 motivation_meaning · 10 challenges_at_work · 11 other.