The Power to Say No: Wage Dependence, Cheap Machine Intelligence, and How to Launch It Safely
Cite this paper
SomaSoft Research (prepared by Claude Code for the AURI project). (2026-11-21). "The Power to Say No: Wage Dependence, Cheap Machine Intelligence, and How to Launch It Safely". SomaSoft Research. https://somasoft.ai/papers/the-power-to-say-no. Licensed under SAGL-1.0.
SomaSoft is a research project, not a product: no users, no deployments. Nothing here is medical, legal or financial advice. Limits and corrections
The Power to Say No
Wage dependence, cheap machine intelligence, and how to launch it safely
Draft, 21 November 2026. Not peer reviewed. A companion to "The Superior of Capital" and "Beyond the Price."
1. The question, made testable
"Will we ever see the end of wage slavery?" is usually asked rhetorically. We want to ask it in a form that has an answer.
The phrase. American labour organisations used the imagery of "wage slavery" widely in the mid-nineteenth century, possibly from the Lowell mill protests of 1836 onward. We use it in that movement's narrow sense: dependence. It is the condition of having to sell your time to someone else to survive, with too little bargaining power to set fair terms. It is not a claim that paid work is equivalent to chattel slavery. A worker owns themselves, can leave and is paid, and that difference is morally enormous.
Lincoln's test. Our earlier paper, The Superior of Capital, took its title from Lincoln's first annual message to Congress (3 December 1861), which called labour "the superior of capital". On our reading of Lincoln's free-labour argument, wage work is free when a person can rise out of dependence within a working life, and it is not free when it fixes them in place.
The test this paper uses is a simpler one that can be measured at any moment:
Can a working person refuse a bad job, or leave one, without falling into hardship?
Where the answer is yes, dependence has ended in the sense that matters, even though people still work for wages. Where the answer is no, it has not, however high the wage. On this view the end of wage slavery is not a day on the calendar. It is the gradual spread of a credible power to say no.
2. What has already changed, and what stalled
Our reading of the record, by the power-to-say-no test, is real progress since the nineteenth century followed by a long stall.
What was won. - Hours. In the United States many people worked seventy hours or more a week in the 1800s, and the average manufacturing week was still about 60 hours in 1890 (Whaples, EH.net). The Fair Labor Standards Act of 1938 capped the standard week at 44 hours, falling to 40 by 1940, and the 40-hour week became the lasting standard. - Children. The same 1938 Act banned oppressive child labour in the industries it covered. - Basic protections. Weekends, workplace safety law and minimum wages were established; the 1938 Act set the first federal minimum wage, at 25 cents an hour. - Safety nets. Unemployment insurance and pensions mean that in most rich countries losing a job is no longer a threat to life itself.
Our argument is that none of this came from markets alone: each gain was won by organising and written into law.
What stalled. In the United States in particular, three facts show how much dependence remains:
- Health coverage tied to jobs. Employment-based insurance covered 53.5% of Americans for some or all of 2025, the most common type of coverage (US Census Bureau, September 2026). Madrian (1994) estimated that this "job lock" cut the voluntary turnover of workers with employer-provided insurance by about 25%, from 16% to 12% a year; later studies disagree on the size. When leaving a job means losing medical care, the power to say no shrinks.
- Collective bargaining has thinned. The union membership rate was 9.9% of wage and salary workers in 2024 (Bureau of Labor Statistics, January 2025), down from 20.1% in 1983, the first year of comparable data.
- Thin buffers. In the Federal Reserve's Survey of Household Economics and Decisionmaking, 63% of adults said they would cover a $400 emergency expense with cash or its equivalent in each survey from 2022 to 2025, so more than a third would not.
Keynes (1930) wrote that, within a hundred years, "three-hour shifts or a fifteen-hour week" might be enough. Productivity did rise. Our reading is that the gains went mostly to higher consumption and to capital, not to shorter hours or more bargaining power.
3. What machine intelligence changes
We expect cheap machine intelligence to change how much paid work society needs, and of what kind. The evidence so far points in two directions at once.
Exposure is wide. In their 2023 preprint, Eloundou et al. estimated that about 80% of the US workforce could have at least 10% of their work tasks affected by large language models, and about 19% of workers could see at least half their tasks affected; they found exposure across all wage levels, with higher-income jobs potentially more exposed. Exposure is not the same as job loss.
Gains are real and uneven, often in the worker's favour. In a large field study of customer-support agents, Brynjolfsson, Li and Raymond (NBER 2023; Quarterly Journal of Economics, 2025) found that a generative-AI assistant raised issues resolved per hour by about 14% on average (15% in the journal version) and by 30–34% for novice and low-skilled workers, with little effect on the most experienced. In our reading, the tool spread the skill of the best workers to the newest ones. It augmented them.
Aggregate effects may be modest, and the split is the question. Acemoglu (2024) estimates that AI will raise total factor productivity by no more than about 0.66% over ten years, and predicts it will widen the gap between capital and labour income. Acemoglu and Restrepo (2019) show that automation always reduces the labour share of value added unless new tasks reinstate labour, and may reduce labour demand even as it raises productivity.
Brynjolfsson (2022) called the overemphasis on machines that imitate and replace people, rather than augment them, the Turing Trap, and argued that technologists, executives and policymakers face excess incentives for automation. Our gloss: the trap is not technical. It is a choice by those who design and buy the systems.
Our argument, then, is that machine intelligence does not settle the question of dependence. It raises the stakes:
- Automation without shared ownership increases dependence. Fewer people bargain with fewer employers, or rely on transfers that owners choose to give.
- Augmentation with shared gains reduces dependence. Each hour of work is worth more, and more of the surplus reaches the people doing it.
4. What actually widens the power to say no
Several mechanisms have been tried and measured. In our assessment none is a cure; each adds to the power to refuse.
Cash floors. - Finland's 2017–18 basic-income experiment paid about 2,000 unemployed people €560 a month. Its 2020 evaluation found small employment effects, while recipients were more satisfied with their lives and reported less mental strain than the control group (Kangas et al., 2020). - The Stockton SEED pilot gave 125 residents $500 a month from February 2019, extended through January 2021. In the first year recipients' full-time employment rose from 28% to 40%, against 32% to 37% in the control group, and the evaluators said the cash removed material barriers to full-time work and gave recipients capacity to take risks. - The OpenResearch study (Vivalt et al., 2024) gave 1,000 low-income people $1,000 a month for three years, against 2,000 controls receiving $50 a month. Recipients worked 1–2 fewer hours a week, labour-market participation fell by about 4 percentage points, and the largest increase in time use was leisure. A companion paper (Miller et al., 2024) found large but short-lived improvements in stress and food security.
Our reading is that cash floors add security and choice at the margin. They do not, at the amounts tested, transform working life.
Services decoupled from jobs. We argue that the job-lock evidence implies that health coverage, housing support and retraining not tied to an employer directly raise the power to say no, whatever the wage.
Worker ownership. - In the United States, 6,609 employee stock ownership plans covered about 15.1 million participants, 10.9 million of them active, on 2023 data (National Center for Employee Ownership, 2026). - Kurtulus and Kruse (2017) found that employee-owned firms laid off about half as many workers as other firms in the 2001 and 2007–09 recessions. - Mondragon in the Basque Country of Spain is a federation of about 90 cooperatives employing about 70,000 people (Mondragon annual report, 2023).
Our argument is that ownership changes who keeps the gains from technology, which section 3 identified as the deciding question.
Collective and sector bargaining. - The Writers Guild of America's 2023 agreement provides that AI is not a writer, that AI-generated material is not literary material, and that companies cannot require writers to use AI software. We regard it as an early, prominent example of workers bargaining directly over AI deployment. - Sector-level wage boards set floors across an industry rather than one firm at a time. California's AB 1228, signed 28 September 2023, created a Fast Food Council and set a $20 minimum wage for national fast-food chains from 1 April 2024.
Portable savings and benefits. We argue these make leaving a job a decision rather than a crisis.
5. Launching these technologies effectively and safely
If the deciding question is augmentation versus replacement, and who keeps the gain, then how machine intelligence is launched matters as much as what it can do. We propose ten rules. They draw on three sources:
- labour practice, especially the bargaining examples in section 4;
- AI risk frameworks: the NIST AI Risk Management Framework (AI RMF 1.0, January 2023), and the EU AI Act, whose Annex III lists AI used for recruitment, selection and the management of workers as high-risk;
-
our own failures, published as dated corrections at somasoft.ai/corrections.
-
Augment first. Start with tools that make a worker better at their job, not tools that remove the worker. Measure task-level outcomes for the workers themselves: quality, time and stress, not only output per hour.
- Decide the gain-share before launch. Agree in advance how productivity gains will be split: wages, shorter hours, training, reinvestment. A gain divided after the fact goes to whoever holds the power at that moment.
- Give workers a seat at design and deployment. The people whose work changes should help decide how, through unions where they exist and through structured consultation where they do not. Deliberation tools such as Polis can gather the views of a whole workforce, not only its loudest members.
- Run in shadow mode first. Let the system run beside existing work, with its outputs recorded but not acted on, until its error rate is measured. We run our own new comprehension components this way.
- Measure against a trivial baseline, on sealed tests. A score means nothing next to "do nothing", "always say yes" or "echo the question" unless it beats them, and tests used for tuning stop measuring. When we re-scored our system's truthfulness benchmark with a penalty for wrong answers, it scored below always abstaining (−0.140 against 0, November 2026); the older, looser measure had hidden this.
- Fail closed where harm is serious. For safety, health, children and pay decisions, a component that errors should stop and refer to a person, not wave the case through. Our own disclosure-safeguarding gate let messages pass through when it errored, and our November 2026 audit lists that as a defect.
- Keep a human answerable for every consequential decision. Hiring, firing, discipline, pay and scheduling decisions need a named person who can explain and reverse them. The EU AI Act's high-risk treatment of employment AI points the same way.
- Fund transitions before displacement. Where tasks will be removed, retraining, notice and income support should be committed and funded before the system goes live, not promised afterwards.
- Publish incidents and corrections. Keep a dated public record of what went wrong and what was changed. In our experience it is the cheapest early-warning system available.
- Put the mission under someone with no stake in giving it up. Our own report on cooperative and civic structures concluded that ownership alone often failed to protect a mission once money arrived, and that what held more often was an independent party with a veto plus rules that bind what is produced. A deployment that promises to augment workers should be held to that promise by someone who gains nothing from breaking it.
6. An invitation to partners
None of this can be done by a small research project alone. We are looking for partners in five roles:
- Worker organisations and cooperatives willing to co-design and test augmentation tools, with the gain-share agreed in advance.
- Researchers in labour economics and human-computer interaction to design evaluations we should not grade ourselves, including sealed tests and real-user studies.
- Community lenders, credit unions and municipalities interested in ownership models and services that widen exit options.
- Nonprofits willing to act as an independent consent holder, or as the holder of a benefit fund.
- Practitioners in AI risk and governance to review our deployment rules against NIST, the EU AI Act and their own experience.
What we can honestly offer: - a measurement discipline that publishes results that are bad for us; - a system designed to run locally and offline, without sending people's data to a provider; - a source-available licence with obligations on commercial users; - governance drafts written before any money has arrived.
What we cannot offer yet: - any real users; - validated benefits; - revenue.
AURI, our research system, has not yet been evaluated with a single real user; a blinded user study is in draft. The first partnership we would value most is one that helps change that, carefully and with consent.
7. Limits
- This is an argument assembled from secondary evidence, not a new empirical result. The figures in sections 2–4 are quoted from the cited studies and were checked against them on 2026-11-22. We have not re-analysed any of them.
- "Wage slavery" is a contested term. We used a narrow definition and a testable proxy for it. Readers who reject the term can read the paper as about the power to refuse work.
- The cash-transfer evidence is mixed and amount-dependent. Larger or permanent programmes may behave differently from time-limited pilots.
- Most evidence is from the United States and Europe. Dependence looks different where most work is informal.
- The ten rules are proposals. We have applied some to our own system (shadow mode, sealed tests, the public corrections record) and none in a real workplace.
8. Conclusion
Will we see the end of wage slavery? In our view, not as a single event. The harshest form, work on any terms or go hungry, has already ended for many people in rich countries. The core of it, dependence without the power to refuse, has in our reading stalled for decades.
We argue that machine intelligence will decide which way it moves next. It reduces dependence if it is launched to augment workers, with the gains shared by prior agreement and the people affected given a voice. It increases dependence if it is launched to replace people and the gains go to whoever owns the machines.
The test stays the same throughout: can a person say no to a bad job and still be all right? Every launch decision either widens that power or narrows it, and it should be measured accordingly.
References
- Acemoglu, D. (2024). The simple macroeconomics of AI. NBER Working Paper 32487, May 2024; published in Economic Policy 40(121), 13–58 (2025).
- Acemoglu, D. & Restrepo, P. (2019). Automation and new tasks: how technology displaces and reinstates labor. Journal of Economic Perspectives 33(2), 3–30. doi:10.1257/jep.33.2.3.
- Board of Governors of the Federal Reserve System. Economic Well-Being of U.S. Households in 2024 (May 2025) and in 2025 (May 2026), Survey of Household Economics and Decisionmaking.
- Brynjolfsson, E. (2022). The Turing Trap: the promise and peril of human-like artificial intelligence. Daedalus 151(2), 272–287. doi:10.1162/DAED_a_01915.
- Brynjolfsson, E., Li, D. & Raymond, L. (2023). Generative AI at work. NBER Working Paper 31161; published in Quarterly Journal of Economics 140(2), 889–942 (2025).
- Bunch, L. N. & Ketema, H. (2026). Health Insurance Coverage in the United States: 2025. US Census Bureau, P60-291, September 2026.
- Bureau of Labor Statistics (2025). Union Members — 2024. News release USDL-25-0105, 28 January 2025.
- Eloundou, T., Manning, S., Mishkin, P. & Rock, D. (2023). GPTs are GPTs: an early look at the labor market impact potential of large language models. arXiv:2303.10130. (A revised version appeared in Science in 2024 with different headline measures.)
- European Union (2024). Regulation (EU) 2024/1689 (Artificial Intelligence Act), Annex III, point 4.
- Kangas, O., Jauhiainen, S., Simanainen, M. & Ylikännö, M. (eds.) (2020). Suomen perustulokokeilun arviointi [Evaluation of the Finnish basic income experiment]. Ministry of Social Affairs and Health, Reports and Memorandums 2020:15.
- Keynes, J. M. (1930). Economic possibilities for our grandchildren.
- Kurtulus, F. A. & Kruse, D. L. (2017). How Did Employee Ownership Firms Weather the Last Two Recessions? Employee Ownership, Employment Stability, and Firm Survival: 1999–2011. W.E. Upjohn Institute for Employment Research.
- Lincoln, A. (1861). Annual Message to Congress, 3 December 1861.
- Madrian, B. C. (1994). Employment-based health insurance and job mobility: is there evidence of job-lock? Quarterly Journal of Economics 109(1), 27–54.
- Miller, S., Rhodes, E., Bartik, A. W., Broockman, D. E., Krause, P. K. & Vivalt, E. (2024). NBER Working Paper 32711, July 2024. nber.org/papers/w32711.
- Mondragon Corporation. Annual Report 2023.
- National Center for Employee Ownership (2026). ESOPs by the Numbers (2023 data).
- National Institute of Standards and Technology (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1, January 2023.
- State of California (2023). Assembly Bill 1228 (fast food restaurant industry), signed 28 September 2023.
- Stockton Economic Empowerment Demonstration (2021). Preliminary Analysis: SEED's First Year. March 2021.
- US Department of Labor. Fair Labor Standards Act of 1938: Maximum Struggle for a Minimum Wage. dol.gov/general/aboutdol/history/flsa1938.
- Vivalt, E., Rhodes, E., Bartik, A. W., Broockman, D. E., Krause, P. & Miller, S. (2024). NBER Working Paper 32719, July 2024. nber.org/papers/w32719.
- Whaples, R. Hours of work in U.S. history. EH.Net Encyclopedia, Economic History Association.
- Writers Guild of America (2023). Summary of the 2023 WGA MBA, provisions on artificial intelligence.
- SomaSoft Research (2026). The Superior of Capital; Beyond the Price; Lock the mission before capital arrives (internal report, not published); Limits and corrections, somasoft.ai/corrections.
AURI's concept graph was queried on 2026-11-19 ("What is wage slavery?"). The term was not in the graph (confidence 0.0), so nothing here comes from AURI. The references and the factual claims in the text were checked against the cited sources on 2026-11-22; figures we could not verify were corrected or removed, and the author's own arguments are marked as such.