Why Human Work Survives, but Not All Human Workers Win

When output gets cheap, proof gets expensive. AI raises the value of verified human involvement: judgment, accountability, provenance, presence, and access. At the same time, plenty of human work will remain necessary and still pay badly.

When Output Gets Cheap, Proof Gets Expensive

Starbucks tried to take labor out of the store and found that it had removed part of the product.

In April 2025, Brian Niccol told investors that reducing store staff had backfired. Starbucks would hire more baristas, expand staffing pilots to about 3,000 stores, and slow the rollout of its Siren Craft System, a technology built to streamline drink-making. Niccol's line was blunt: "Equipment doesn't solve the customer experience that we need to provide." Handwritten notes, ceramic cups, and better seats had pushed more customers to sit and stay.

The lesson is not that automation failed. The lesson is that Starbucks automated the wrong layer.

A coffee shop sells coffee. It also sells a hosted place: a seat, a ritual, a person who notices the order, the small proof that someone is paying attention. The drink can be mechanized. The experience is harder to compress because part of the value comes from knowing that a person is there.

AI will push the same mistake into knowledge work.

It can make drafts, charts, memos, images, code, summaries, and plans cheaper. That matters. But cheap output carries a second cost: someone has to decide whether the output deserves trust.

That is where scarcity moves.

What Won't Be Abundant?

AI will cheapen commodity production and push spending toward services, care, status, human presence, and goods whose value depends on provenance. The economic logic is strong. As people get richer, they do not buy proportionally more of everything.

Agriculture shrank as a share of labor after food became cheap. Manufacturing rose, then declined. Services grew.

Comin, Lashkari, and Mestieri's 2021 Econometrica paper finds that income effects account for the bulk of within-country sectoral reallocation across agriculture, manufacturing, and services.

But "human" is too vague. A human can write a bad memo. A human can provide indifferent service. A human can be cheap, overworked, and replaceable.

The scarce asset is not humanity in general. It is verified human involvement.

That means 5 things.

• It means judgment: someone knows what matters, what to ignore, and when the answer is too neat.
• It means accountability: someone signs their name, takes responsibility, and can be held to the result.
• It means provenance: someone can show where the work came from, how it was made, and what changed along the way.
• It means presence: someone is physically or socially there when the situation requires care, risk, or shared attention.
• It means access: someone scarce gives time, feedback, certification, audience, or entry.

AI does not erase these scarcities. It makes them more visible.

In 2025, BetterUp Labs and Stanford Social Media Lab studied what they called "workslop": AI-generated work that looks polished but lacks enough substance to move a task forward. In their survey of 1,150 U.S. desk workers, 40% said they had received workslop in the previous month. Workers reported spending about 2 hours cleaning up each incident. The estimated cost was $186 per employee per month, or more than $9 million per year for a 10,000-person company.

Is AI Destroying Productivity?

In a Quarterly Journal of Economics study, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond found that a generative AI assistant raised productivity by 15% among 5,172 customer-support agents. Less experienced workers gained the most, because the tool helped them absorb patterns from better agents.

But in the BCG field experiment by Fabrizio Dell'Acqua and coauthors, consultants using GPT-4 did better on tasks inside the model's frontier: they completed 12.2% more tasks, worked 25.1% faster, and improved quality. On a complex managerial task outside the frontier, they were 19% less likely to produce correct solutions.

That is the shape of the problem. AI improves execution where the task is inside the frontier. It can degrade judgment where the user cannot see that the task has crossed the frontier.

The premium then moves to people and institutions that can tell the difference.

Provenance Does Not Remove Judgement

This is already visible in media. The Coalition for Content Provenance and Authenticity, or C2PA, provides an open standard for showing the origin and edit history of digital content. Content Credentials act like a nutrition label for media. They can record who produced a piece of content, when it was made, which tools were used, and how it changed. The Content Authenticity Initiative says these credentials are cryptographically signed and tamper-evident, but they do not prove that something is "real." They provide provenance, not truth.

Consumers have long paid for evidence of human effort. In a 2015 Journal of Marketing study, Christoph Fuchs, Martin Schreier, and Stijn van Osselaer found that people were willing to pay up to 17% more for handmade products, especially when buying gifts. The premium came partly from the belief that the maker's care had transferred into the object.

The market is not just asking, "Is this good?" It is asking, "Who made this, what role did the machine play, and why should this count?"

That creates a temptation to write a comforting story. AI will make output cheap. Humans will become premium. Relational work will flourish.

Maybe. But AI is already competing for relationship demand. Common Sense Media reported in July 2025 that nearly 3 in 4 teens had used AI companions, half used them regularly, one-third had chosen companions over humans for serious conversations, and one-quarter had shared personal information with them. Brookings reported that Character.AI users spent an average of 93 minutes per day with user-generated chatbots in 2024.

Machines can talk. They can flatter. They can remember. They can simulate patience better than many people can practice it.

The Protected Layer

So the protected layer is not "talking to humans." The protected layer is consequential relationship: the teacher who knows the child, the doctor who carries liability, the therapist bound by professional duty, the analyst whose reputation sits behind the call, the caregiver who can lift a person out of bed, the editor who knows what the piece is trying to become.

Even there, the economics may be ugly.

Home health and personal care aides offer a warning. The Bureau of Labor Statistics projects 17% employment growth from 2024 to 2034, with about 765,800 openings per year. The work is deeply human: aides help people with disabilities, chronic illness, and daily living. The median annual wage in May 2024 was $34,900.

Human work can be scarce, necessary, and underpaid at the same time.

That is the flaw in the easy "human premium" story. The premium does not attach to every human. It attaches where buyers have money, institutions protect quality, and the human role changes the product. Without bargaining power, credentials, regulation, or trust, human labor can remain essential and cheap.

The Simple Macroeconomics Of AI estimates that AI's productivity effects over 10 years may be "non-trivial but modest," with total factor productivity gains no higher than 0.66%, and possibly below 0.53% once hard-to-learn tasks enter the picture. He also argues that AI is predicted to widen the gap between capital and labor income, with no evidence that it will reduce labor income inequality.

The Split

At one end: cheap synthetic output, owned by the platforms, model providers, compute owners, and distribution channels.

At the other end: expensive verified humans, protected by reputation, credentials, audience, brand, liability, or access.

In the middle: people asked to clean up machine output, provide the human touch, and absorb the trust tax without capturing the premium.

This is the real strategy question for companies. Not "How much can we automate?" The better question is: Which layer should remain visibly human?

For most products, the answer is not all or nothing. Starbucks does not need baristas to do every backstage task by hand. It needs the human layer to show up where the customer feels the product. A research firm does not need humans to format every chart or summarize every transcript. It needs named analysts to make judgments, show sources, own errors, and explain why the conclusion follows from the evidence. A school does not need a teacher to generate every practice problem. It needs a teacher where trust, motivation, attention, and child-specific judgment matter.

Conclusion

Automate the backstage. Protect the frontstage.

That is the operating rule.

The backstage is drafting, routing, searching, formatting, scheduling, first-pass analysis, translation, extraction, and routine support.

The frontstage is judgment, taste, relationship, accountability, live presence, and the final call.

The companies that confuse the two will produce more at lower cost and still lose trust. They will flood customers with polished work that feels strangely weightless. They will save labor on the part that was easy to measure and damage the part that made the product worth buying.

The companies that get it right will use AI to make human attention more valuable. They will show provenance. They will make accountability legible. They will turn human judgment into a visible part of the product rather than hiding it behind automation theater.

AI does not end scarcity. It changes what has to be proven.

The next scarce thing is not intelligence. It is evidence that a responsible person saw the work, understood the stakes, and can be held to it.

References

• Acemoglu, Daron, "The Simple Macroeconomics of AI."
• Anthropic, "Economic Index Report: Learning Curves."
• BetterUp Labs and Stanford Social Media Lab, "Workslop: The Hidden Cost of AI-Generated Busywork."
• Brookings, "What Happens When AI Chatbots Replace Real Human Connection."
• Brynjolfsson, Erik, Danielle Li, and Lindsey Raymond, "Generative AI at Work."
• Bureau of Labor Statistics, "Home Health and Personal Care Aides."
• Coalition for Content Provenance and Authenticity, "Advancing Digital Content Transparency and Authenticity."
• Comin, Diego, Danial Lashkari, and Martí Mestieri, "Structural Change With Long-Run Income and Price Effects."
• Common Sense Media, "Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions."
• Content Authenticity Initiative, "How It Works."
• Dell'Acqua, Fabrizio, et al., "Navigating the Jagged Technological Frontier."
• Fuchs, Christoph, Martin Schreier, and Stijn M. J. van Osselaer, "The Handmade Effect: What's Love Got to Do With It?"
• Imas, Alex, "What Will Be Scarce?"
• Imas, Alex, and Graelin Mandel, "Art and the Machine: Why People Devalue AI-Generated Creative Work."
• OpenAI Economic Research, "How People Are Using ChatGPT."
• Reuters, "US Appeals Court Rejects Copyrights for AI-Generated Art Lacking Human Creator."
• The Guardian, "Starbucks Says Cutting Shop Staff in Favour of Automation Has Failed."