A SOCIAL RESEARCH STUDIO

Intelligence at work.
Time to live.

equilibri

A research studio connecting AI ethics with the humanities of everyday life. Asking how agent-assisted work could return time to people, and why shared places belong in that conversation.

LIVE PROJECT DIRECTED BY TRUDY HALLRead the thesis
THE POSITION

If intelligence becomes increasingly abundant, can human life become less crowded?

Imagine a Tuesday afternoon. An agent finishes the administrative work that usually fills the last two hours of the day. A freelancer goes for a walk. An employee receives another assignment. The software has done something similar. The working arrangements have produced different lives.

equilibri begins in that difference. AI agents are systems that can take several steps toward a goal, sometimes acting across other software. If they become part of ordinary work, their social value will depend on who sets that goal, who carries the risk and what happens after a task is finished.

The humanities ask what makes an activity worth doing and a life worth having. AI ethics asks how power, responsibility and harm are distributed. Bringing them together moves the discussion beyond the screen, into employment agreements, care, neighborhoods and the material cost of computation.

The proposition is a hybrid life with room for paid work, other people and time without an assignment. Coworking could provide some of its social infrastructure. Neither more capable software nor a well-designed room can produce that life alone. Sometimes the right decision will be to keep a task human, use a simpler tool or leave something unautomated.

01 / THE ARGUMENT

Follow the task
into the rest of life.

Productivity is the beginning of the inquiry. Follow its gains and costs far enough and questions about software become questions about wages, authority, relationships and place.

01 / TIME

A faster task does not settle the working day.

The familiar promise is relief from administration. That relief can be real. In a study of 5,172 customer-support agents, AI assistance increased issues resolved per hour by about 15% on average, with larger gains for less experienced workers. This is evidence of useful assistance in a particular setting. It does not tell us who gets to leave earlier. Brynjolfsson, Li & Raymond · Quarterly Journal of Economics, 2025

A task has several clocks. There is the machine’s running time, the person’s active effort and the period during which that person must remain available. Ten minutes of checking scattered across an afternoon can occupy more of a life than thirty uninterrupted minutes of work. A dashboard can report a speedup while the worker still cannot go out.

This makes attention part of the account. Count instructions, corrections and handovers. Then ask whether the person can actually stop. Time that is technically unoccupied but remains on call is a different benefit from an afternoon one can make plans for.

Research & its limits / time

Shorter hours have to be organized

A six-month study followed 2,896 employees in 141 organizations adopting a four-day week without reducing pay, alongside 12 comparison companies. Reported well-being improved. Participating organizations volunteered, and the intervention was work reorganization, not AI. It supports testing deliberate reductions in hours; it cannot establish that automation produces them. Fan et al. · Nature Human Behaviour, 2025

What to measure here

The opening Tuesday is an illustrative scenario, not fieldwork. A study of agent-assisted work would need time diaries and interviews as well as task metrics. A walk, a school pickup and time waiting for an alert may all appear as “inactive” in a system log. They are not interchangeable experiences.

02 / PAY & OWNERSHIP

Who receives the dividend?

AI could make skilled work more valuable by helping someone serve a client better, learn a difficult procedure or spend more time on judgment. The opportunity is to negotiate how that value is shared. A tool that expands a worker’s contribution gives us a reason to discuss compensation, not a reason to assume compensation will follow.

For employees, a negotiated share of verified gains could support higher base pay or shorter hours at maintained pay. Profit sharing needs an intelligible formula and access to the accounts. Worker ownership can add a claim on long-term value, although it also carries risk. Collective bargaining can turn an individual request into an agreement that survives a change of manager.

For freelancers, completing work sooner can improve effective hourly earnings under a fixed-fee contract, provided quality holds and checking costs are included. Under hourly billing, the same speedup can reduce the invoice. Pricing, demand and who owns the client relationship matter as much as technical skill. Benefits that follow people between contracts would need institutions and funding beyond a software subscription.

Research & its limits / distribution

Productivity and pay are different outcomes

Humlum and Vestergaard link Danish adoption surveys to administrative employment data. Their working paper finds changes inside jobs but no detectable average effect on earnings or recorded hours in the two years after ChatGPT’s launch. This is a bounded early finding, not a forecast of permanent wage stagnation. It makes the missing transmission from useful tools to worker income worth investigating. Humlum & Vestergaard · NBER working paper, revised 2026

The proposal is an agreement, not a prediction

The distribution mechanisms above are options to test. Calculate gains after software, review and training costs; protect base pay; include support staff; and agree how workers can inspect or contest the calculation. New responsibility for evaluating and maintaining systems should be recognized as paid work. Fewer hours should not conceal a second person’s unpaid labor.

03 / AGENCY & MEANING

What remains yours when software can act?

Asking a system to draft an email and authorizing it to send one are different decisions. So are letting it compare applicants and letting it exclude them. Agency means having a real say over those boundaries, including a usable route to refuse, correct or stop an action. A confirmation button offers little protection if refusal is penalized or review time is missing.

Authority also depends on knowing enough to disagree. A novice who always receives the finished answer may miss the practice through which an expert learned to recognize a bad one. Yet removing tedious work can also leave more time for learning. Which tasks teach judgment, and which merely consume it? That answer belongs partly to the person doing the work.

Writing can be a way of finding out what one thinks. A difficult conversation can be part of repairing trust. Such activities have value in their performance, beyond the deliverable. equilibri proposes protected opportunities to practise, make decisions and remain answerable to other people. Technical capability alone cannot decide what should be delegated.

Research & its limits / choice

Ask workers what they want to delegate

Stanford’s WORKBank compares preferences from 1,500 workers across 104 U.S. occupations with expert assessments of agent capabilities. It finds varied preferences for human involvement and mismatches between capability and desire. This survey-based preprint helps frame a question about consent; it does not test the long-term consequences of delegation. Shao et al. · WORKBank preprint, revised 2026

Design for the moment a system is wrong

Amershi and colleagues’ CHI guidelines address expectations, correction and control across human–AI interaction. The work comes from Microsoft and academic collaborators. These are evaluated design guidelines, not proof that an employer will respect a worker’s choices. The studio extends the design question to paid review time, separate permissions for reading and acting, and a named person who can resolve a challenge. Amershi et al. · CHI, 2019

04 / RELATIONSHIPS

Delegating a task can change a relationship.

Consider a colleague who uses an agent to summarize a conversation that another person understood to be private. The disagreement concerns more than accuracy. It concerns what each person thought they were sharing, and with whom. One person’s convenience can alter someone else’s terms of participation.

Care ethics starts from the fact that people depend on one another. A good arrangement asks who notices an unmet need, who has time to respond and whose effort remains unrecognized. An agent might clear paperwork so a person can listen. It might also become the reason an institution offers less access to a person. These are different choices about care.

Enthusiasm and skepticism about AI can therefore become questions of trust between friends or collaborators. A shared workspace should make those differences discussable. Agree before recording a meeting, uploading another person’s work or replacing a conversation with an automated response. Belonging should not require agreement about the technology.

Research & its limits / care and situated work

Convenience can create new obligations

A review by Hertog, Shi and Rodel examines 83 studies of digital technologies in unpaid care. It identifies help with coordination alongside risks including constant availability, privacy loss and reduced physical contact. These technologies extend beyond generative AI. The relevance here is relational: reducing one task can introduce another obligation, and the people giving and receiving care may experience the same tool differently. Hertog, Shi & Rodel · Oxford research summary, 2026

Work does not unfold exactly as its plan predicts

Lucy Suchman’s account of situated action directs attention to the circumstances in which people make technology workable. Here it informs a research method: follow the interruptions, informal help and repairs that a task description leaves out. The examples on this page are questions for observation, not claims that the studio has conducted an ethnographic study. Suchman · Human-Machine Reconfigurations, 2007

05 / LABOR & INSTITUTIONS

Whose work disappears from view?

The person using an AI system sees an answer arrive. Behind it are people who created source material, labeled examples, evaluated responses, moderated harmful content and maintained infrastructure. Authors and experts may never have met the organization using their work. Hardware also has to be manufactured, transported and repaired. “Automated” describes only part of this arrangement.

This is the sociotechnical view of AI. Companies decide what to collect, employers decide what to reward, and interfaces make some choices easy while hiding others. Those arrangements shape the technology that gets built; the technology then changes which arrangements seem normal. Asking what “AI wants” can obscure the people with actual authority.

The same question reaches into the coworking kitchen. Someone hosts, cleans and keeps the room usable while another person enjoys newfound flexibility. A credible proposal must budget for their pay, predictable schedules and rest. If the benefits stop at the laptop, the account of a better working life is incomplete.

Research & its limits / work behind the interface

Assess conditions across the supply chain

Oxford’s Fairwork research evaluates digital labor through pay, conditions, contracts, management and representation. Its cloudwork work brings the conditions behind online services into view. Ratings concern the platforms and evidence assessed; they should not be treated as a single verdict on every AI worker or supplier. Fairwork · AI labor research

An institutional question

The studio proposes asking suppliers how workers can contest rejected work, how expert contributions are recognized and what happens when a task exposes someone to harmful material. Transparency helps inquiry, but it does not substitute for bargaining power. A purchasing decision is one point of influence, not a complete remedy.

06 / MATERIAL AI

The cloud occupies land.

Computation requires electricity, water, chips and buildings. Semiconductor production, construction and hardware replacement carry impacts that a measure of electricity per response can miss. A system may be used in one city while its cooling demand and supply-chain burdens fall elsewhere. Environmental responsibility has a geography. UNEP · AI’s environmental footprint

The IEA reports that data-centre electricity demand rose 17% in 2025 while energy use per AI task was falling. These data-centre totals include more than AI. They nevertheless show why efficiency per task and total demand must be examined separately. Cheaper computation can encourage many more tasks. IEA · Key Questions on Energy and AI, 2026

There are uses worth investigating. AI can assist energy forecasting, coordinate complex systems and help detect waste. equilibri also asks whether computation can help map its own infrastructure and support ecological repair. A useful model needs measured inputs and independent checks. It cannot reveal an undisclosed footprint simply by being asked.

The test is comparative. What does this application achieve, what would a simpler method achieve, and what is the added cost across the whole job? A restoration project does not cancel an unrelated data centre’s water demand. A serious footprint does not tell us that every application has identical value.

Research & its limits / ecological judgment

Useful applications do not establish a net benefit

The IEA examines AI applications in energy operations and innovation, including forecasting and efficiency. Outcomes depend on adoption, data and the infrastructure around the model. Their potential does not establish that growth in AI is environmentally beneficial overall. IEA · Energy and AI, 2025

Compare complete alternatives

Include failed runs, repeated searches, verification and infrastructure where data allow. Record missing disclosures rather than inventing precision. For a shared workspace, add travel, building energy, materials and replacement cycles. Reuse, nearby leisure and appropriately sized tools are design choices to assess, not automatic carbon savings.

07 / PLACE

More capable software can make gathering more valuable.

If routine coordination becomes easier to delegate, the office’s justification can change. Attendance becomes a poor proxy for contribution. What might deserve a journey is a shared problem, a person who knows something one does not, or simply the pleasure of working among familiar faces.

Coworking offers a spatial response to distributed, agent-assisted work. Independent workers and employees from different organizations can share a place without sharing a chain of command. A recurring kitchen conversation could become an occasion to question a tool, learn from an experienced colleague or discuss something unrelated to work. This is a proposition about social infrastructure, not a prediction that everyone will need a membership.

The room could also fail. Cost can exclude people. Noise can make concentration impossible. A community built around the same tastes can narrow its members’ horizons. Good hosting, quiet rooms and access beyond a professional clique are therefore part of the research, not finishing touches.

The possibility is a workplace whose value survives a decline in the hours spent at a desk. That requires permission to leave as well as reasons to return. Homes, libraries, conventional workplaces and remote participation remain necessary alternatives.

Research & its limits / coworking

Community takes work

Spinuzzi and colleagues studied six coworking spaces in the United States, Italy and Serbia to examine what community means in practice. The study distinguishes forms of community rather than treating proximity as collaboration. It supports a more precise design question about how a space is organized; it does not prove a universal coworking benefit. Spinuzzi et al. · Journal of Business and Technical Communication, 2019

Social infrastructure is a possibility, not a label to award oneself

Research on shared workspaces examines how some can support local social and professional networks. A commercial workspace does not become civic infrastructure merely by describing itself as a community. The studio’s proposition needs to be tested through access, continuity and relationships with the surrounding neighborhood. Are shared workspaces a new form of social infrastructure? · 2024

02 / THE ATLAS

Usable freedom
has a local address.

Thirty minutes saved means something different beside a park than beside a motorway. The distance between work, care and ordinary pleasures helps determine whether flexibility can become part of daily life.

The atlas asks what a shared workplace should borrow from its surroundings. A useful room belongs to a wider neighborhood. It cannot supply affordable housing, public transport or care simply by offering a day pass.

EUROPE / PROXIMITY & ACCESS

Portability is only the first freedom.

A laptop may let someone work anywhere. The school run, an inaccessible station or the price of a quiet room can still decide where they spend the day. A neighborhood workspace could reduce the effort of moving between roles, but only if it fits the routines and resources of people who live nearby.

The design question extends beyond major-city offices. Research on shared workspaces as social infrastructure asks how they might sustain local networks and independent businesses. This is a direction to investigate, not a claim that European places already share one successful model. Shared workspaces and social infrastructure · 2024

Count the whole cost of access, including the journey, care arrangements and the hours a person can actually use. An affordable desk at the wrong time can still be out of reach.

SAN FRANCISCO / A SHARED INTELLECTUAL LIFE

A room for unfinished thinking.

The Commons describes a setting where coworking sits alongside philosophy salons, shared meals, art and member-led gatherings. The useful idea is continuity. People can encounter one another repeatedly without every encounter needing to become a meeting. The Commons · operator description

For equilibri, a shared intellectual life should also include dissent. Could an AI researcher, a skeptical artist and someone whose work is being reorganized remain in the same conversation? The aim would be to make room for disagreement without requiring anyone to turn an objection into a pitch.

A convivial room is not evidence that its members have more autonomy. Ask who feels welcome, who does the hosting and whether unfamiliar perspectives remain part of its life.

LOCAL CONDITIONS MATTER

The proposal should travel as a set of questions. Climate, transport, care obligations and the local price of space change what is practical. A destination workspace for mobile professionals may offer little to someone who must remain close to home. Expansion would need local participation, rather than a single lifestyle exported worldwide.

THREE SPATIAL REFERENCES

Make a change of pace easy to reach.

LISBON / PORTUGAL

Design with living things.

Second Home Mercado places plants and natural light inside the working environment. The idea to borrow is to plan for living greenery from the outset, including the labor and resources needed to maintain it.

Second Home
ULUWATU / INDONESIA

Let the day move outside.

BWork lists a garden and outdoor lounges alongside its workspaces. A nearby change of setting is useful inspiration. Shade, accessibility and local climate determine whether it works in another place.

BWork
CAPE TOWN / SOUTH AFRICA

Give the interval some space.

Workshop17’s Kloof Street precinct connects buildings through a courtyard. A route between rooms can also hold the conversation that outlasts a meeting, provided people have time to stay.

Workshop17
03 / THE PRINCIPLES

Put the conditions
in the agreement.

A humane working arrangement needs terms that survive a busy week. These four proposals connect the interface to the people who set deadlines, authorize actions and decide how value is shared.

Define enough.

Agree the result, quality threshold and stopping point before delegating. Include checking in the schedule and cost. More generated material should not silently become the new measure of commitment.

Keep authority visible.

Separate permission to read, recommend and act. Limit data access and make it withdrawable. Name who can stop the process, who resolves a challenge and which decisions require a person.

Allocate the gains.

Agree how verified savings affect pay, workload and available time. Include the people doing review and support work. A protected break needs an arrangement for interruptions, not a hope that none will arrive.

Leave room to refuse.

Provide a workable alternative when AI is unsuitable or consent is absent. Review consequences with affected people. Change or stop the arrangement when its benefits depend on concealed costs.

AN ILLUSTRATIVE AFTERNOON

There is nothing to approve at lunch.

A researcher asks an agent to prepare a source comparison. They have agreed a review time with their collaborator. The system may gather and draft, but cannot publish or contact anyone. At lunch it can wait. The break is protected by the agreement, even if the software makes no progress.

  1. Frame the questionChoose the sources, set a resource budget and specify the evidence needed. Decide what the agent is not authorized to do.
  2. Leave the taskPause notifications and let unresolved questions queue. Take lunch at the shared table, go outside or be alone.
  3. Return with judgmentCheck important claims against original sources. Discuss ambiguity with a person when needed. Reject the draft if it does not hold up.
  4. Close the workCount review and correction time. Record what remains uncertain. Additional requests need a new agreement rather than an endlessly expanding brief.

For hands-on work, the route to a less crowded day may be adequate staffing, predictable shifts and paid breaks. A shared benefit should not require everyone’s job to become a task for an agent.

Leisure need not earn its place by making tomorrow’s output better. An unhurried meal or time spent doing nothing can be the point of the arrangement.

04 / A PLACE TO TEST THE IDEAS

equilibri, as a place.

The spatial proposition brings the argument into a neighborhood coworking setting. Can people gain control over their day while retaining income, learning from others and accounting for ecological costs? The room would be a place to investigate those conditions, with room for people who choose not to use AI.

Watercolor coworking concept with a concrete arch, plywood walls and desks, cork flooring, reclaimed timber seating, suspended hemp acoustic panels, salvaged brick patio and solid walkways, with an imagined view of Valencia’s City of Arts and Sciences.
Spatial concept · A shared place for work and life.
THE SPACE

Different rooms for different kinds of attention.

Begin with an existing building near public transport. Provide daylight, step-free access and a quiet room with adjustable desks. Enclose calls and meetings. People should be able to choose company without surrendering privacy or sensory comfort.

A kitchen table would make eating together possible without making it compulsory. A reading corner, bean bags and a shaded patio would offer other ways to occupy the day. Plywood, planting and repairable furniture would give the room character without making constant refurbishment part of its identity.

THE DAILY EXPERIENCE

Hospitality is part of the infrastructure.

A paid host would protect quiet hours, notice access needs and offer introductions when wanted. Fund mentoring rather than treating expertise as an inexhaustible community gift. Voluntary sessions could compare AI results with original sources, examine a failed delegation or develop a project without AI. Use non-confidential examples and agree what may be recorded.

Membership would not grant permission to study someone. Research participation would be optional, with withdrawal independent of access to the space. Keep personal productivity private, minimize data collection and publish combined findings. The proposed model would not sell member data or attention.

MATERIALS / USE LESS, KEEP LONGER

Concrete

Retain and repair what the building already contains. Exposed concrete should be a consequence of reuse, not a reason to pour a new decorative surface. Compare environmental product data for any essential new material.

Plywood

Use plywood for walls and furniture, including desks, shelves and shared tables. Favor reclaimed panels or verified FSC-certified supply, with low-emission finishes. Removable fixings would let the room change without sending its previous arrangement to landfill.

Cork

Specify replaceable flooring and pinboards where suitable. Check binders, emissions and durability. The warmth of a surface is one quality; tested acoustic performance is another.

Reclaimed timber

Make tables and benches from suitable salvaged wood. Inspect earlier treatments and structural condition. Design the joints so the next repair is ordinary maintenance, not a replacement project.

Hemp

Explore hemp-based ceiling and acoustic panels alongside the plywood. Choose by tested sound absorption and fire performance, including the binder and finish. Plan removal and reuse rather than assuming a plant-based product has no end-of-life cost.

Salvaged brick

Let suitable local salvage guide patio paving or landscape edges. Check drainage, durability and an even surface for wheeled access. Reuse is useful only if the result can serve the people who need it.

Reuse comes first. Claims about lower carbon emissions would need to account for the whole interior, including transport and future replacements. EPA · material reuse EPA · salvage and reuse SAPPA · hemp acoustic panels EPA · embodied carbon FSC · certified products EPA · indoor air

MEMBERSHIP & REVENUE

Access has to be funded.

The proposed income would come from flexible memberships, dedicated desks and limited room bookings. Budget for hosting, cleaning and mentoring before promising a community program. Reduced-rate places need a committed revenue share or partner funding, rather than an intention to become inclusive later.

Publish prices and access rules. Base membership on a shared code of conduct, with a way to challenge decisions. Belonging would not depend on professional prestige or enthusiasm for AI.

COSTS & VIABILITY

Price the ordinary work of keeping a place open.

The model needs to cover a deposit, surveys, essential alterations and a reserve for surprises. Recurring costs include rent, staff, utilities, insurance and repairs. Plant care and patio maintenance belong in that account. Natural materials do not maintain themselves.

Test whether realistic use can cover those costs without overcrowding the room or relying on unpaid labor. Cash reserves and recovery of startup spending need separate plans. A full space is not a successful experiment if the price of viability defeats its purpose.

Test the model before taking a lease

Separate the agreement from the room.

Begin in an existing venue. Compare similar tasks under different work agreements, then examine what the shared setting adds. Where feasible, use comparison groups or randomized schedules. Record task quality, review time, income and time on call alongside participants’ accounts of the day. Track task difficulty and changing tools. Otherwise an easier brief or a novelty effect could look like a successful model.

Observe what the measures miss. Does a novice find someone to ask? Can an AI skeptic participate comfortably? Do quiet members feel pressured to socialize? Include people who leave or decline to join. A lease would follow demonstrated demand, financial review and local building and access checks.

Let the findings change the proposal.

Set the criteria before the trial. Gains should mean usable time, maintained or improved income and acceptable work quality, without extra burdens for hosts or support workers. Interviews can show how a change feels; they cannot alone establish what caused it. Report differences between occupations and circumstances rather than only an average.

Compare travel, building energy, materials and additional computation with a stated alternative. Publish uncertainties, no improvement and harmful outcomes as well as promising ones. Revise or stop if checking consumes the gains, real income falls, leisure contracts or work moves invisibly onto someone else.

A research and design proposal, not an operating venue. The trial and financial model remain to be tested.