Research on advanced AI, human agency and ecological responsibility. Examining the conditions for safe development and constructive coexistence.
equilibri is an editorial research project examining advancing machine intelligence, human agency and the living world. It brings published studies, institutional assessments and independent journalism into an inquiry about safety, public authority and constructive coexistence.
The inquiry connects present pressures on electricity, water, livelihoods and public services with the possibility of much more capable systems. Scientific discovery and useful assistance deserve serious attention. So do misuse, failures of oversight and the concentration of power. A desirable relationship between people and AI must be established through evidence, institutions and practical protections.
CURATED RESEARCHEnergy analysis documents rising data-centre demand. Workplace studies report useful assistance alongside concerns about intensity, data use and unequal outcomes. Model capabilities remain uneven. Benefits and burdens depend on the task, the institution and the conditions of deployment. [03] [27] [01]
INTERPRETATIONConcern is a legitimate response to consequential uncertainty. This inquiry takes warnings seriously, tests promises against evidence and remains open to beneficial forms of cooperation. Human judgment, the ability to refuse or challenge a use, and responsibility for consequences remain central to that relationship.
Research findings, interpretation and proposals are labelled throughout. The inquiry begins with ecological responsibilities and governance, then considers literacy, working life, frontier risks and the conditions for cooperation. Earth, the interactive data visualization below, brings the infrastructure behind these questions into view.
The IEA estimated that coal and gas supplied more than half of global data-centre electricity in 2024. These totals include workloads beyond AI, and supply differs by region. Its 2026 update describes an emerging push toward onsite gas in the United States as grid connections lag. [19] [03]
Annual renewable-energy purchasing and the electricity physically available each hour are different accounts. Both need clear boundaries. Location, timing, additional generation and backup power determine what a clean-power claim can establish. [19] [20]
The footprint also includes water, mineral extraction, manufacturing and electronic waste. UNCTAD’s wider digital-economy research describes unequal environmental burdens and access to benefits across countries. An AI-specific assessment needs to account for its share of that supply chain. [04] [35]
PROPOSAL · PRIORITIES FOR REPAIRAssess new wind and solar, storage, transmission and feasible firm low-carbon supply, including existing nuclear and suitable geothermal. Test costs and delivery dates against new demand and other public needs. Future technology should enter a supply plan only when its delivery is credible. [19] [20]
Assess grid capacity, seasonal water stress and cumulative impacts with the people affected. Shift flexible workloads when it genuinely helps the grid, while respecting service needs. A cleaner national mix alone does not justify a new facility or establish community agreement. [20] [23]
Compare models, conventional software and changes in process against the same task and quality standard. Measure resource use in practice; model size alone is insufficient. Track total demand as well as efficiency, including whether easier or cheaper use increases consumption.
Require comparable reporting on electricity, water and emissions, with measured and estimated values distinguished. Agree who funds grid upgrades and local services before expansion. Give communities access to independent expertise, enforceable conditions and a route to challenge non-delivery.
Prevent damage first. Where harm occurs, assign responsibility for remediation and locally agreed ecological restoration, including supply-chain impacts. Establish a baseline, fund long-term monitoring and publish results. Repair cannot erase emissions or guarantee recovery from irreversible loss.
Reuters reported Microsoft’s January 2026 commitments on data-centre power costs and water disclosure. They provide something concrete to follow: what is delivered, who pays and whether affected people can verify the result. [25]
George Packer’s reporting in The Atlantic follows residents in Clinton, Iowa, who opposed a data-centre proposal developed through confidential planning. Their differing views of AI converged on a shared demand for influence over local decisions. [38] Opposition to data-centre expansion has been widely documented internationally, including in Ireland, the Netherlands, Chile, the UK and Australia. Across these settings, communities have challenged pressure on water, electricity, land and everyday life, with concerns shaped by local conditions. [39] [40]
UNESCO’s 2021 Recommendation connects human rights, environmental protection, participation and accountability. The OECD’s 2025 framework for AI in government combines institutional capacity, oversight and public engagement. These frameworks guide decisions; they do not establish that a particular arrangement works. [28] [30]
INTERPRETATION · THE STUDIO’S POSITIONAI governance should combine democratic authority, independent scrutiny and the knowledge of affected people. Developers and organizations deploying systems carry responsibilities that consultation cannot transfer to the public. Clear duties and accessible remedies make that division of power practical.
Legislatures should define public-interest obligations. Regulators need resources and clear mandates to enforce them; courts and independent complaint bodies need to offer effective remedies. Agencies using AI must explain and answer for their own decisions.
Researchers, auditors and civil society need access appropriate to the risks they evaluate. Protect personal data and legitimate security interests while disclosing funding, conflicts and material findings. An audit should inform action, with clear responsibility for fixing what it finds.
Workers and their representatives should negotiate deployment conditions. Communities should shape infrastructure decisions before commitments are made. Pay for participation, make it accessible and explain what it changed. Consultation should not be presented as consent, especially where people cannot freely refuse.
Require risk assessment, documented limits, serious-incident reporting and correction of failures. Keep duties traceable across suppliers and customers. Independent review can test compliance; enforceable obligations are needed to act on its findings.
The UN Advisory Body’s 2024 report proposes international coordination that includes countries underrepresented in AI governance. Shared evidence and resources for independent research can make participation more substantive. The report is a proposal for cooperation, not a global licensing authority. [29]
Frontier safety requires scrutiny of evaluation itself. The 2026 safety assessment identifies difficulties detecting dangerous behavior and limited evidence on the effectiveness of risk management. These limits strengthen the case for independent testing, protected reporting and authority to respond; they do not establish that any one oversight model is sufficient. [41]
The studio proposes published reasons, conflict-of-interest rules, independent review and accessible appeals. Public authority needs checks against surveillance and political abuse; private influence needs checks against capture of public decisions. People should be able to see who has authority and how to contest its use.
Requirements should reflect a system’s capabilities, purpose and consequences. Decisions affecting livelihoods or access to services need stronger scrutiny than routine drafting. Clear standards and shared testing resources can support smaller organizations and open research. Some uses may require restriction; others can proceed with safeguards and review. [28] [30]
Ada Lovelace Institute’s review finds attitudes vary with the application and its context. Its evidence is predominantly British and predates much recent adoption. Early MIT Media Lab and OpenAI studies find social and emotional outcomes vary with usage and personal circumstances. Neither represents everyone’s experience. [11] [12]
INTERPRETATIONPeople who welcome new assistance, people affected by an unfair decision and people who need an accessible service all have knowledge this inquiry needs. Agreement should grow from tangible protections and benefits, with space for disagreement about a particular use.
We propose clear notice when AI materially mediates an interaction, appropriate permission to use private contributions and a workable human route for consequential services. Alternatives should support people who rely on assistive tools as well as those who prefer to decline AI. These proposals need to be assessed alongside the rights and duties that apply in each setting.
Language models learn patterns from training data. Predicting the next token, a piece of text, is a central training objective for many models; further training shapes instruction-following. That description helps explain how responses are produced, but does not capture everything a trained system can do. Fluency alone does not establish accuracy. [13] [01]
Anthropic’s 2025 interpretability experiments traced planning and intermediate steps in selected tasks, alongside explanations that did not reflect the underlying computation. The methods covered only part of one model. A model’s explanation of its reasoning is not independent verification of its answer. [14]
INTERPRETATIONAI literacy means being able to assess assistance, understand its limits and know where a decision can be challenged. It includes privacy, evidence and the difference between a suggestion and authorized action. Technical confidence is only one part of that competence.
Providers, employers and public institutions should make learning and safe use possible through paid time, accessible information and appropriate support. People should not need expert knowledge to receive fair treatment or get help when a system fails.
PROPOSAL · FOUR HABITS OF INFORMED PRACTICEName the need and the outcome that would count as success. Compare assistance with a simpler approach.
Agree the quality standard, privacy boundaries and expertise required before using the system.
Use original sources and independent checks suited to the stakes. An answer can sound certain and still need correction.
Distinguish drafting, recommending and acting. Check what an agent can access, change or commit, and who can stop it.
Humlum and Vestergaard’s Danish research found no detectable average change in earnings or recorded hours during the early period studied, while documenting task reorganization. The ILO’s global index measures occupational exposure to AI, not observed job losses. Adoption, exposure and economic outcomes are different measures. [15] [16]
Reuters’ June 2026 reporting described muted aggregate US wage and employment effects alongside risks for junior workers. Stanford’s August payroll update found younger workers falling behind in more AI-exposed occupations, mainly through reduced hiring. The authors caution that these patterns do not establish how much AI caused them. [26] [37]
The human work behind AI also belongs in the account. Fairwork’s 2025 study assesses 16 online labor platforms, including data-annotation services, against pay, conditions, contracts, management and representation. It reports improvements alongside persistent gaps in worker protection. [18]
INTERPRETATIONA more productive task does not by itself mean a better wage, a secure job or a stronger public service. Training matters, but bargaining power, market structure and social protection shape who gains. Workers should have a say in how integration changes their work. [17] [22] [24]
PROPOSAL · PRACTICAL RESPONSESNegotiate purpose, data rules, workload and review with workers or their representatives. Fund learning during paid time. Where roles change, plan income protection, redeployment and viable routes into work. Public support must also reach contractors and people without stable employment.
Count training, verification and error correction before claiming a saving. Negotiate better pay, staffing, security or shorter hours at maintained pay. Include contracted and outsourced work, and preserve routes for new entrants to learn. Report outcomes by role and seniority.
Public buyers should assess accessibility, labor conditions, privacy and total costs over a contract’s life. Retain the capacity to change supplier or stop a system. Judge success by service quality and fair access, with complaints and human review that people can actually use.
Economic policy should support people through changes in demand for their work. Access to benefits should not depend on being an early adopter or having a job that is easy to automate.
The International AI Safety Report 2026 examines emerging risks including misuse, unreliable behavior and possible loss of control. Researchers disagree about the likelihood of the most severe outcomes. The assessment also identifies limits in evaluation and evidence about the effectiveness of safeguards. [41]
INTERPRETATIONHere, superintelligence refers to a possible system that substantially exceeds human capabilities across a broad range of cognitive work. It is a scenario for analysis, with contested timelines and consequences. Benchmarks, forecasts and deployment evidence answer different questions; none provides a complete account of what comes next.
Uncertainty about probability does not remove the need to examine severe consequences. A plausible danger does not establish its inevitability. The research task is to identify mechanisms of harm, assess the evidence for them and ask which safeguards remain credible as capabilities and autonomy increase.
Competence at a task does not confer the right to set its purpose. Decisions about access, permissions and deployment need legitimate human institutions, independent scrutiny and avenues for challenge. [42]
Require evaluations suited to the deployment, independent access appropriate to the risk, incident reporting and explicit conditions for restriction or suspension. Record what remains untested and who can act when evidence changes.
Symbiosis is a research question here: under what conditions can human and machine capabilities complement one another without creating coercive dependence or eroding agency? Its value must be assessed through outcomes, correction, meaningful choice and ecological costs. The term does not establish machine consciousness, aligned interests or a guaranteed future.
A CHI survey of 319 knowledge workers found greater confidence in generative AI was associated with less self-reported critical-thinking effort. Participants also described work shifting toward verification and integrating responses. The study does not establish lasting cognitive decline or a causal effect. [09]
Human–AI interaction guidance emphasizes clear expectations, correction and control. W3C’s cognitive-accessibility guidance adds the importance of understandable content and involving people with disabilities in design. These approaches inform support that people can use in practice. [10] [36]
The OECD’s 2026 practitioner interviews examine emerging organizational use and governance of agents. They describe early practice, rather than a settled model for delegating consequential work. [02]
INTERPRETATIONAssistance can open possibilities for people who face barriers to reading, writing, communication or routine tasks. Agency depends on whether people can shape that help, correct it and get support. Effective oversight also requires institutions to give reviewers authority, time and the ability to change a decision.
Separate permission to read, recommend and act. Make commitments visible and reversible where possible.
Limit interruptions and batch handovers. Protect time to assess what needs review.
Name who resolves errors. Reviewers need authority, knowledge and enough time to act.
Support people who rely on assistance and people who decline it. Provide a usable route to a person when the stakes require one.
Reduce ecological pressures, assign responsibility for damage and fund restoration that can be independently assessed.
Measure total demand as well as efficiency. Compare approaches against the same need and quality standard.
Give affected people influence before commitments are made. Publish reasons, disclose conflicts and explain what changed.
Keep duties clear, decisions open to challenge and remedies accessible. Oversight needs the authority to act.
Provide appropriate support and workable alternatives. Count workload and interruptions when judging a benefit.
Negotiate pay, security, time and public benefit. Include the workers and places that make the system possible.
These briefs turn the inquiry into decisions for public authorities, operators, employers and educators. They are editorial proposals for evaluation, not completed interventions or findings of effectiveness.
Before contracting or expansion, public authorities, purchasers and operators should examine additional electricity supply, grid capacity, seasonal water stress and cumulative local impacts. Assess delivery dates, public-service needs and who will pay. Give affected communities independent support to participate.
Publish measured, estimated and undisclosed values separately. Set enforceable conditions, review points and responsibility for repair. A clean-power contract alone cannot establish the ecological acceptability of a project.
Before deployment, name who authorizes, independently reviews and can suspend the system. Involve affected people while alternatives remain open, and explain how their input changes the decision. Agree permissions, complaint handling and remedies.
Track errors, unequal effects and whether challenges produce correction. Fund human review with real authority. Check whether an alternative service is usable, including for people with disabilities, and whether declining assistance creates hidden penalties.
Employers, educators and civic organizations should provide time, accessible materials and supported practice. Teach privacy boundaries, how outputs are produced, how to check important claims and where to get help. Compare AI with simpler methods without requiring private information as practice material.
Evaluate whether people can identify errors, choose suitable assistance and challenge a decision. Include varied languages, abilities and experience. Institutions remain responsible for safe deployment; literacy must not become a condition for receiving fair treatment.
Researchers and public bodies can compare AI-assisted analysis of environmental disclosures with a reviewed reference set and simpler tools. Publish the source trail, uncertainties and resource costs of the work.
Measure missed issues, false alarms and verification effort. Keep findings open to challenge and responsibility with the institution. Continue only when the comparison shows a worthwhile contribution; monitoring alone is not evidence of ecological repair.
Employers and worker representatives should agree a baseline, paid learning, review time and distribution of any verified gain before a trial. Include contractors, outsourced workers and routes into the occupation.
Compare pay, workload, staffing, security and discretionary time by role. Revise or stop if a benefit depends on unpaid work, intensified targets or transferring burdens to others. Where gains are real, make the agreement on sharing them enforceable.
Agree the baseline, comparison, review points and stopping conditions before a trial. Report harms, costs and lack of improvement alongside benefits. Assess rights, ecological effects, income and time separately, with affected people involved in deciding what a worthwhile outcome means.
Explore public data on computing infrastructure and electricity. The globe brings three sources into one view, so their patterns, gaps and differences can inform better questions about the places behind AI.
National generation shares for the same year. This compares one part of the energy context; it does not rank countries or establish where new demand should go.
This globe combines third-party data. Directory pins do not verify operating status, AI activity or electricity demand. Country colours show national generation, not facility supply. Reported CFE means carbon-free energy, including nuclear, with the operator’s matching method; it is one metric, not a sustainability rating.
Data centres © Ringmast4r · Electricity: OWID / Ember / EI · CFE: Google Cloud · Map: Natural Earth.
Start with a place and compare the layers. They have different coverage, scales and reporting methods. Water, land, labor conditions and community priorities require further evidence.
Renewables and nuclear appear separately in each country’s mix. Low-carbon generation still has environmental impacts. Grid capacity, water, biodiversity, materials, affordability and local rights remain part of an infrastructure decision. Nearby markers do not establish an electricity connection.
“A question to investigate” is an editorial prompt based on the selected national generation share. Its bands are for learning, not engineering or policy thresholds. They do not assess a project’s feasibility, establish community agreement or recommend relocation.
Download the displayed dataset · Latest source check · Directory data licence · Source notes [31] [32] [33] [34]
Primary studies, institutional reports, design guidance and independent journalism. Editorially revised . This is a curated inquiry, not a systematic review or original empirical study.
Selection favors traceable evidence relevant to ecological and ethical decisions. Notes identify methods, dates and limits; forecasts, observed outcomes, policy frameworks and corporate commitments serve different purposes. Developer involvement is named. English-language and institutional sources are prominent, so this collection cannot stand for every place or experience. The studio’s interpretations and proposals are labelled, and source inclusion does not imply endorsement. Revisions should follow better evidence, including findings that challenge the position.
A synthesis of capability, adoption and responsible-AI indicators. Benchmark results vary by task and do not establish reliability in a particular deployment.
Practitioner accounts of developing, deploying and governing agents. Useful for emerging practice, not representative adoption rates or proof of safe delegation.
Energy-sector estimates and conditional projections. Data-centre totals include non-AI workloads; uncertainty about future demand and supply remains material.
An institutional overview of environmental pressures across the AI lifecycle. Its older per-query estimates are not used as current measurements here.
Applications in forecasting, operations and innovation. Estimated potential does not establish a net ecological benefit after deployment costs and wider demand are counted.
A staggered rollout to 5,172 customer-support agents, with uneven productivity effects. Findings in this setting do not establish gains in pay or discretionary time.
A six-month work-reorganization and time-reduction intervention. Participating firms were not randomly sampled, and the intervention did not involve AI.
A historical essay about productivity and leisure. Used to frame a question about economic purpose, rather than as a forecast for contemporary AI.
Self-reports from 319 knowledge workers. Associations between confidence and critical-thinking effort do not establish causation or enduring loss of ability.
Research-derived interaction guidance on expectations, correction and control. Useful design principles still require institutional duties and usable remedies.
A rapid review of context-dependent attitudes, predominantly from the UK. It predates much recent adoption and is not a global opinion poll.
A summary of observational and controlled studies of ChatGPT use and well-being. Results depend on context; the product developer participated in the research.
Model-training research comparing next-token and multi-token prediction. A training objective helps explain a system without fully describing its learned capabilities.
Selected mechanistic experiments in Claude 3.5 Haiku, conducted by its developer. Partial traces reveal capabilities and misleading explanations without fully explaining the model.
Danish surveys linked to administrative records. Early average earnings and hours findings coexist with changes in tasks; country and observation period limit generalization.
Task-based estimates of technical exposure, with differences by occupation and country. Exposure is not a measure of adoption, displacement or inevitable job loss.
Analysis of worker voice and collective bargaining during AI integration. Institutional context shapes the options; the studio’s proposed agreements are its own.
An assessment of 16 online labor platforms, including data-annotation services, against five fair-work principles. It finds improvements and continuing protection gaps; coverage is not the whole AI labor supply chain.
Physical electricity supply, contractual purchasing and conditional scenarios. Historical generation shares refer to 2024, not current facility telemetry.
Energy supply, storage, grid constraints and matching methods. Options differ in cost, lead time and local suitability; projections are conditional.
Comparative research on AI experiments in public audit institutions. It informs a proposal to evaluate assistance, rather than validating autonomous environmental oversight.
Competition analysis of AI markets, concentration and barriers to entry. It informs questions about ownership and distribution, not a forecast of household gains.
System-level analysis of flexible demand and efficiency. A specific workload needs its own assessment of feasibility, service requirements and verified savings.
Published June 2026, drawing on earlier studies and data reaching late 2024. Skills evidence is interpreted alongside working conditions, voice and institutional responsibility.
Reporting on company commitments amid local concerns. Delivery, costs and ecological outcomes need subsequent evidence.
Journalism on ECB research about aggregate US outcomes and junior-worker vulnerability. Kept distinct from the primary studies in this collection.
Policy research on worker-reported benefits and concerns, including workload, data use and inequality. Context and differences between workers matter.
An intergovernmental ethical framework connecting rights, sustainability and accountability. It guides policy but does not itself enforce obligations in every jurisdiction.
An advisory proposal for inclusive international coordination. It is not a description of an existing global regulator or licensing system.
A framework for institutional capacity, oversight and public engagement in government AI. The studio’s broader proposals are interpretations, not established effects or OECD endorsement.
Earth plots 4,760 screened listings from 18,110 source records. Location precision and coverage vary; screening does not verify operating status, computing capacity or AI activity.
Annual national generation shares with varying country coverage. Low-carbon combines renewables and nuclear. This layer does not establish facility supply or account for electricity imports.
Operator-reported hourly CFE matching averaged over the year, plus grid operational emissions intensity. Earth maps 46 reviewed regional anchors, not buildings; CFE includes attributed clean generation.
Simplified public-domain geography for orientation and a coarse coordinate screen. Coastlines and small territories lose detail; the display does not resolve disputed boundaries.
Examines material extraction, digital infrastructure and waste, including unequal burdens across countries. It covers the digital economy broadly, rather than isolating AI’s footprint.
Supplemental accessibility guidance on understandable content and involving users in design. It is not an AI-specific trial or a claim of WCAG conformance.
US ADP payroll patterns through mid-2026 differ by age and AI exposure. The authors identify limitations in causality and generalization; these are not definitive estimates of AI-caused job loss.
Reporting on infrastructure planning and residents’ participation in Clinton, Iowa. A local account, not a representative study or environmental-impact measurement.
Documents community opposition in several countries, including Ireland, the Netherlands, Chile and the UK. A synthesis of reported cases, not a global survey of public opinion.
Reports the withdrawal of a proposed Lane Cove facility following community opposition and changing planning conditions. A specific Australian case, not a measure of worldwide opposition.
Reviews emerging capabilities, misuse, reliability, loss-of-control risks and risk management. Expert disagreement and evaluation limits remain material. It informs scrutiny of frontier risks, not a precise prediction of superintelligence or endorsement of the studio’s proposals.
Calls for safeguards that preserve human agency and oversight throughout the AI lifecycle. Provides a normative basis for examining coexistence and institutional responsibility; it is not empirical evidence that a particular safeguard works.
equilibri brings evidence into a conversation about the choices already being made. The possibilities of advanced intelligence and the severity of its risks belong in the same inquiry. Its contribution depends on how systems are developed, governed and used, and on whose lives and places count in the decision. A livable future requires safety that can be examined, institutions that remain answerable and cooperation that protects human agency and the living world.
