Connecting AI literacy, ecological repair and the question of who gets to decide. We investigate how informed participation, accountable institutions and shared benefits can support a more livable world.
equilibri is a live social research project centered on AI literacy and ecological repair. We connect questions about how technology works with questions about the world it depends on. What should people understand, what should institutions change, and what would meaningful improvement look like?
For this studio, AI literacy includes the ability to evaluate an answer, question a decision and recognize when assistance is inappropriate. Ecological repair means pursuing recovery for damaged places alongside reducing further harm. Both require evidence, public participation and institutions willing to be accountable.
RESEARCHResearch documents both useful capabilities and persistent uncertainty about reliability and internal mechanisms. Workers report benefits alongside pressures around intensity, data collection and inequality. The same systems depend on physical infrastructure with local energy and water demands. [01] [14] [27] [03] [04]
INTERPRETATIONOur inquiry connects literacy with the power to participate, and ecological repair with institutions capable of requiring it. Understanding a system is only part of the task. People also need a say in its direction and a route to challenge harm. Relief from tedious work matters within that larger purpose.
This inquiry remains open. Leading research and independent journalism inform the position; proposals are revised as evidence changes. The central question is what we should ask of these systems before building more of our lives around them.
Language models learn statistical relationships from data. Predicting subsequent tokens, the pieces into which text is divided, is a central training task for many models. Further training shapes how an assistant follows instructions. The resulting fluency does not, by itself, establish factual accuracy. [13] [01]
The mechanism can support more than copying familiar phrases. Anthropic’s 2025 interpretability work identified planning and intermediate reasoning in selected tasks, alongside misleading explanations. Its methods captured only part of the computation in one model. Both capability and incomplete understanding belong in the account. [14]
INTERPRETATIONAI literacy is a form of practical agency. It includes understanding capabilities and limits, protecting private information and recognizing who has authority over a system. The aim is better judgment, including the confidence to ask for evidence or choose another approach.
A “language calculator” is one limited metaphor for a language model. It points toward computation, but offers none of an ordinary calculator’s assurance of exact results. A helpful explanation must leave room for both real capabilities and mistakes.
State what a successful answer would accomplish and whether a simpler method could do it.
Decide what evidence, error tolerance and human expertise the task requires.
Inspect original sources, test calculations and compare results with independent evidence.
Keep drafting, recommending and acting separate. An agent connected to software needs explicit permissions and a route to stop.
These are studio proposals for informed practice. Early organizational accounts of agents are useful evidence, but do not establish a settled model of safe delegation. [02]
UNESCO’s 2021 Recommendation links human rights, environmental protection, public participation and human accountability. The OECD’s 2025 framework for AI in government combines oversight with citizen engagement and proportionate safeguards. These are normative and policy frameworks, not proof that a particular governance arrangement succeeds. [28] [30]
INTERPRETATION · THE STUDIO’S POSITIONAI should be governed through democratically accountable public institutions, informed by independent expertise and shaped by the people affected. Developers and deploying organizations retain responsibility for their systems. Authority should be distributed, with explicit duties and a clear route to remedy.
Legislatures should establish public-interest rules; properly resourced regulators should enforce them within clear mandates. Courts and independent complaint bodies should provide routes to challenge decisions and seek remedy. Public agencies deploying AI must face scrutiny themselves.
Researchers, auditors and civil society need appropriate access to evaluate safety, discrimination and environmental effects. Protect privacy and legitimate security information, disclose funding and conflicts, and publish findings that people can understand. Expertise should inform public decisions.
Workers and their representatives should negotiate workplace conditions. Communities should influence infrastructure choices before commitments are made. Accessible, paid participation and representative deliberation can bring in people without technical credentials. Decision-makers should explain how that input changed the outcome.
Require developers and deploying organizations to evaluate risks, disclose limitations, report serious incidents and correct failures. Keep responsibility traceable across suppliers and customers. Voluntary commitments and external audits should support enforceable duties, not replace them.
International cooperation should support shared evidence, compatible reporting and participation by countries with fewer resources. The UN Advisory Body’s 2024 report proposed an inclusive, distributed governance architecture. We draw on that direction without treating the UN as a global AI licensing authority. [29]
Our proposed safeguards include published reasons, conflict-of-interest rules, independent review and accessible appeals. Consultation must have a defined influence on decisions. Public authority needs checks against surveillance and political abuse, just as private authority needs checks against capture and concentration.
Requirements should reflect the stakes, capability and deployment context. A consequential public decision needs stronger scrutiny than low-stakes drafting. Clear standards and public testing resources can support smaller developers and open research without weakening basic protections. Review rules as evidence changes. [28] [30]
Ada Lovelace Institute’s review finds public attitudes vary by context, with support for some benefits alongside demands for control and human judgment. Its evidence is predominantly British and predates much recent adoption. MIT Media Lab and OpenAI’s early studies likewise find social and emotional outcomes depend on how people engage and their circumstances. Neither supports a simple story of connection or isolation. [11] [12]
INTERPRETATIONSomeone may find in AI an accessibility aid or creative collaborator; another may see surveillance, extraction or unwanted mediation. A disagreement about a meeting summary can concern consent and trust as much as accuracy. Living together does not require one shared enthusiasm.
The studio proposes AI optionality where practical, disclosure when AI materially mediates a relationship, consent before processing someone else’s private contribution, and access to a person for consequential or sensitive interactions. These are proposed conditions for common life, not universal rights already guaranteed.
The IEA’s 2025 analysis estimated that coal and gas supplied more than half of data-centre electricity globally in 2024. Supply varies substantially by region. Its 2026 update describes an emerging push toward onsite gas generation in the United States as grid connections lag. [19] [03]
A renewable-energy contract is not the same account as the electricity physically serving a facility at every hour. The relevant questions include location, timing, additional generation and what happens when wind or sunshine falls. [19] [20]
Water, hardware and extraction also matter. A lower-carbon electricity supply does not settle the whole ecological account. [04]
PROPOSAL · PATHWAYS TOWARD REPAIRPlan wind and solar alongside storage, transmission and suitable firm low-carbon generation. Assess existing nuclear and geothermal where feasible. Compare local impacts, cost and delivery dates; a promised future reactor cannot power today’s demand. [19] [20]
Locate capacity where electricity and water conditions can support it. Shift flexible computation to less constrained hours. Examine hourly supply alongside annual purchasing claims, and test reliability before expansion. [20] [23]
Compare smaller models, conventional software and changes in process. Choose the least resource-intensive approach that meets a verified need. Track total demand as well as efficiency so savings are not silently absorbed by growth.
Publish electricity, water and emissions data with clear boundaries. Agree who pays for additional infrastructure and protect other customers from avoidable cost shifts. Include affected communities in decisions about siting and expansion. [03] [04]
Require operators to prevent damage first, then address the land and water impacts they create through locally agreed restoration and remediation plans. Establish a baseline, fund long-term monitoring and publish results. Count avoided harm and ecological recovery separately; a distant offset should not stand in for repairing a damaged place.
Reuters reported Microsoft’s January 2026 commitments to cover data-centre power costs and disclose regional water use. These are company pledges, not verified outcomes. The studio calls for enforceable terms and independent scrutiny. [25]
Danish evidence found no significant average effect on recorded earnings or hours in the early period studied. The ILO measures occupational exposure, not a forecast of job losses. These findings leave room for substantial differences between people, occupations and firms. [15] [16]
Reuters’ June 2026 reporting on ECB research described muted aggregate US wage and employment effects alongside vulnerability among junior workers in highly exposed sectors. An average can conceal unequal experiences. [26]
OECD research identifies social dialogue, skills and accountability as parts of the response. Training alone cannot settle questions of job quality or who receives productivity gains. Market concentration and conditions in outsourced labor also belong in the account. [17] [24] [18] [22]
PROPOSAL · PRACTICAL RESPONSESEmployers should negotiate the purpose, data rules, workload expectations and review process with workers or their representatives before rollout. Fund training during paid time and plan redeployment and income support where roles change. Public policy must also reach people outside stable employment.
Measure savings after verification, training and error correction. Agree how any gain supports better pay, shorter paid hours or improved staffing. Protect contractors from being paid less simply because a task becomes faster. Share results by role and seniority, not only as an organizational average.
Public buyers should require accessible service, labor standards and transparent total costs. Compare vendors and retain a viable exit. The return should include better provision for people, not only reduced expenditure or a higher volume of automated transactions.
These are institutional choices to negotiate and evaluate. They cannot guarantee that every role survives or that benefits will be evenly distributed. That is why income, security and bargaining power must be part of the brief from the beginning.
In a study of 5,172 customer-support agents, AI assistance increased issues resolved per hour by 15% on average, with different effects across workers. That is a task-productivity finding, not a finding about shorter days. [06]
A separate six-month study of 2,896 employees across 141 organizations found improved well-being after work reorganization and a four-day week at maintained pay, compared with 12 control companies. It was not an AI intervention, and participating organizations were not a random sample of all employers. [07]
INTERPRETATIONA faster task can become another assignment, a staffing reduction, higher income, a shorter day or time for learning and care. Contracts, ownership and bargaining power shape that choice. Time remains occupied if someone must keep watching for an alert.
In 1930, Keynes imagined productivity creating a problem beyond scarcity: how to inhabit greater leisure. His forecast is a historical provocation, not a schedule for AI. It invites a question industrial economies still need to answer. What is economic progress supposed to give people time to do? [08]
For equilibri, contemplation, play, civic participation, art, conversation and doing nothing do not need to improve tomorrow’s output to matter. Nor should leisure become another performance metric. Time without an assignment is a legitimate outcome.
A CHI study of 319 knowledge workers found that greater confidence in generative AI was associated with less self-reported critical-thinking effort. Participants described more emphasis on verification and integrating responses. The survey does not establish that AI causes lasting cognitive decline. [09]
Human–AI interaction research gives practical guidance on expectations, correction and control. Those controls matter most when a system is wrong, not just when it is convenient. [10]
INTERPRETATIONCognitive offloading can make room for thought; it can also remove occasions to practise. Our design question is where assistance helps someone learn, and where it conceals the reasoning they need to challenge an answer. Judgment, questioning, synthesis, taste and ethical interpretation deserve deliberate practice.
Separate permission to read, recommend and act. Preserve review time and a route to stop.
Batch handovers. A stream of tiny approvals can occupy a day even when each takes seconds.
Name who resolves errors. Human oversight needs authority, knowledge and time, not a ceremonial click.
Keep a workable human route where stakes or consent require it. Refusal should not secretly become a penalty.
These are studio design commitments, not a claim that an interface alone can secure them.
Judge a change by paid hours, workload and freedom from interruption, alongside the quality of what it produces.
Make permissions clear, preserve the right to challenge decisions and provide a workable alternative when consent or stakes require it.
Name the energy, water, materials and infrastructure involved. Identify who bears each burden and who is responsible for reducing it.
Measure absolute resource use as well as efficiency. Choose a smaller system, a simpler method or no computation when sufficient.
Negotiate pay, time, security and public benefit. Include workers and communities whose contribution is easy to leave out of the account.
Give affected people a meaningful voice, keep institutions accountable and preserve the right to challenge. Participation must allow room for disagreement.
These briefs turn the pathways into responsibilities and checks. They are proposals for partners to evaluate; no completed trial or proven outcome is claimed.
Educators, employers and civic organizations can teach people how language models generate responses, where they fail and how to verify an answer. Include practice with misleading outputs and a simpler-tool comparison. Participation should not require disclosing private material.
Evaluate whether people become better at identifying errors, checking sources and deciding when to decline assistance. Confidence and frequency of adoption are insufficient measures. Include both experienced users and people who prefer a human route.
Purchasers, operators and public authorities should require a credible supply plan before contracting. Compare cleaner generation, storage, grid capacity and flexible demand against delivery dates, local impacts and responsibility for infrastructure costs.
Publish measured, estimated and undisclosed values separately. Examine absolute demand, seasonal water stress and who pays. A company pledge or efficiency ratio alone is insufficient evidence of improvement.
Employers and worker representatives should agree the baseline, paid training, review time and distribution of any verified gain before a pilot begins.
Compare pay, workload, staffing, job security and discretionary time by role. Revise or stop if improvements depend on unpaid work, intensified targets or shifting burdens to less powerful workers.
Before deployment, organizations should name who authorizes the system, who independently reviews it and who can suspend it. Define permissions to read, recommend and act. Agree an accessible complaint process with the people affected.
Track errors, complaints, remedies and unequal effects. Publish reasons for consequential decisions and provide human review with real authority. Examine whether participation changes outcomes and whether people who decline face penalties or hidden extra work.
Researchers can compare AI-assisted disclosure analysis with a human-reviewed reference set and simpler tools. Publish source trails, resource use and uncertainty.
Measure missed issues, false alarms and verification effort. Keep responsibility with the analyst. Continue only where the assistance improves the quality or cost of scrutiny.
Agree outcomes and comparison methods in advance. Protect consent, publish uneven effects and report harm or no improvement. Keep income, time, ecological costs and social connection visible as separate outcomes. Review the decision when conditions change.
Primary studies, institutional research, model-developer research and independent journalism. Updated on . This is a curated inquiry, not a systematic review.
Numbers beside claims lead to the source notes below. Dates and evidence types distinguish findings, forecasts, reporting and organizational claims. Journalism provides context and scrutiny; a reported pledge is not a verified result. Model-developer studies are identified as such and read within their methodological limits. Sources inform the studio; they do not endorse its interpretations. New evidence should change the argument.
Capability, adoption and uneven performance. Benchmarks do not establish dependable autonomy in every setting.
Early organizational evidence. Interviews describe emerging practice, not representative adoption rates.
Observed electricity demand and conditional projections. Total data-centre demand includes non-AI workloads.
Lifecycle pressures and environmental monitoring. The studio does not adopt its older per-query energy estimates.
Forecasting, energy operations and innovation. Potential savings are not a measured net ecological benefit for AI.
Study of 5,172 customer-support agents. Task productivity is distinct from wages or shorter hours.
Six-month organizational intervention. Evidence for deliberate time reduction, not AI-created leisure.
A philosophical and economic provocation about abundance and leisure, not an empirical forecast adopted here.
319 knowledge workers and self-reported experiences. Associations do not establish permanent skill loss.
Guidance on expectations, correction and control. Interface design alone cannot guarantee institutional accountability.
Context-dependent attitudes, predominantly UK evidence. This is not a current global opinion poll.
Observational and controlled studies with mixed, context-dependent outcomes. Includes participation by the product developer.
Explains next-token training and tests a multi-token alternative. Training objectives do not fully describe a model’s learned capabilities or guarantee accuracy.
Selected experiments in Claude 3.5 Haiku reveal planning, reasoning and unfaithful explanations. Partial mechanistic evidence, not a complete account of all models.
Danish survey and administrative evidence. Early earnings and hours findings are bounded by the observation period.
Task-based exposure estimates. Exposure measures technical potential, not observed job losses.
Worker voice and negotiated transitions. Institutional context matters; no single distribution mechanism is prescribed.
Pay, conditions, contracts, management and representation in platform work. A framework for supply-chain inquiry.
Distinguishes physical electricity supply from contractual accounting. Historical shares and regional projections are dated, not real-time measurements.
Supply options, storage, grid constraints and hourly matching. Technology pathways involve costs and lead times; none removes the need to assess local conditions.
Early experiments across audit institutions. Does not validate autonomous environmental auditing.
Market concentration and barriers across AI supply chains. Grounds the question of who can capture value.
Demand flexibility and efficiency as complementary tools. System-level findings do not establish savings for every AI workload.
Published June 2026; latest underlying data reach late 2024. Training must sit alongside worker voice, privacy and accountability.
Reports corporate commitments amid local concerns. It does not establish delivery or independently measured environmental improvement.
Context on aggregate outcomes and junior-worker vulnerability. Reporting is distinguished here from the primary studies cited elsewhere.
Worker-reported benefits and concerns, including intensity, data use and inequality. Experiences vary; these are not universal effects.
Human rights, environmental sustainability, participation and accountability. Normative guidance; not a self-executing global enforcement system.
Proposes inclusive, distributed international coordination. Cited as a governance proposal, not a description of present regulatory authority.
Government AI governance, oversight, proportionate measures and citizen deliberation. Wider recommendations here are the studio’s interpretation, not measured effects or OECD endorsement.
equilibri connects AI literacy with ecological repair. The purpose is informed participation, accountable power and better conditions for the people and places that make technological life possible.
