AI for environmental protection and sustainable development

Executive summary

International experience shows that AI delivers its greatest and most credible environmental benefits when the technology is tied to a clear physical chain of decisions or actions: satellite detection that leads to an inspection, a forecast that changes energy production, an optimisation model that controls pumps or traffic signals, or a sensor that provides earlier warning of fires and emissions. AI systems that only produce analyses, visualisations or general recommendations have a considerably weaker documented environmental impact.

The most promising solutions for Sweden are:

• Multimodal remote sensing and machine learning for forests, land use, wetlands, coastal environments, biodiversity and methane emissions.
• Forecasting and optimisation algorithms for electricity systems, district heating, water supply, wastewater treatment and flexibility in buildings.
• Digital twins for flooding, heat, stormwater, water networks, transport and physical planning – provided they are built on standardised data and used for concrete decisions.
• Edge AI and distributed sensors for forest fires, water quality, species protection and supervision where connectivity is limited.
• Computer vision and robotics for waste sorting, materials recycling and quality control in circular flows.

The conditions are relatively favourable. Swedish environmental monitoring has long time series distributed across several agencies, while Boverket and Lantmäteriet are already working towards more up-to-date, comparable and standardised planning and geospatial data. The challenge is therefore less about creating yet more isolated AI pilots and more about making data interoperable, linking systems to the operational processes of public authorities, and funding independent evaluation and long-term management.

The main recommendation is that Sweden establish a multi-year national programme for AI for the environment and sustainable development, with shared environmental data infrastructure, six to ten problem-oriented pilot environments and conditional scaling-up. Funding beyond the pilot phase should only be provided when a solution demonstrates measurable additional environmental benefit, an acceptable total life-cycle impact, robustness beyond its training data, and a clear chain of responsibility for errors and appeals.

AI should not be treated as environmentally cost-free. Data centres use electricity, water, land, network capacity and hardware; the IEA reported that data-centre electricity use increased by 17 per cent during 2025, while the OECD and UNEP are calling for more life-cycle-based measurement. Governance should therefore measure net benefit: avoided emissions and resource losses minus the system's electricity, water, material and rebound effects.

Method and assessment framework

The report is based primarily on material from public authorities, multilateral organisations, research institutions and peer-reviewed articles. Company data have been included when they describe real-world implementations, but are identified as self-reported or modelled where independent verification is lacking. The state of the evidence was assessed as at 3 August 2026.

The case studies cover energy, transport, water, waste, agriculture, forestry, biodiversity, climate and weather modelling, and environmental monitoring, as well as several technical categories: supervised machine learning, deep neural networks, remote sensing, edge AI, optimisation algorithms, digital twins, robotics and multimodal sensor systems.

Four levels of evidence

• A – strong: a peer-reviewed study or independent evaluation with a relevant comparison, validation or counterfactual design.
• B – operational: officially reported results from actual operation, but without a complete independent impact evaluation.
• C – self-reported or modelled: effects reported by suppliers, preliminary effects or effects calculated through models.
• D – not yet quantified: the technical function has been demonstrated, but the environmental outcome has not been quantified.

High accuracy in a classification model is not the same thing as environmental benefit. The assessment therefore distinguishes between model performance (precision, recall, forecast error), operational impact (shorter processing times, earlier detection), environmental outcomes (reduced emissions, reduced water loss, conservation status) and net impact, including the AI system's energy use, hardware, data collection and rebound effects.

Selected international case studies

Google DeepMind – cooling data centres (USA/global)

Reinforcement learning and predictive optimisation in energy and building operations, operational at internal scale. Thousands of sensors for temperature, power, pumps and cooling systems require calibration, stable telemetry and strict safety limits. Google reported up to 40 per cent lower energy use for cooling; the system calculates actions approximately every five minutes under safety constraints. Evidence B/C: real-world operation, but reported by the company. Scalability in Sweden: high – relevant to data centres, hospitals, larger properties, water and wastewater facilities and district heating, combined with heat recovery and reporting of electricity and water use.

Machine learning for wind-power forecasting (USA and France)

Deep neural networks for production and market value in energy, in operation or pilot use in commercial environments at organisations including Google, DeepMind and ENGIE. The models are based on historical wind production, weather forecasts, turbine status and market data. Google stated that day-ahead forecasts increased the value of wind energy by around 20 per cent through better commitments to the electricity market – primarily an economic and system-operational outcome; direct emissions reductions have not been isolated. Evidence C. Models can degrade when weather, turbines or market rules change, and forecasts must express uncertainty and be compared with physics-based models. Scalability: very high – greatest benefit through coordination between SMHI, Svenska kraftnät, grid companies and producers.

GraphCast and GenCast – global weather forecasting (UK/global)

Graph neural networks and generative ensemble models for weather, climate-related risk and energy forecasting, in advanced research and pre-operational use. Training is based on global reanalysis data and numerical weather fields, including 0.25-degree resolution. GraphCast produces ten-day forecasts and GenCast generates 15-day ensembles in minutes; GenCast outperformed ECMWF's ensemble on 97.2 per cent of 1,320 evaluated targets. Evidence A for model performance, but the environmental benefit depends on how the forecasts are used. There is a risk of systematic errors under extreme or novel climate conditions, and of over-interpreting deterministic results. Scalability: high as a complement – not a replacement – to SMHI's physics-based models, particularly for wind, cloudbursts, fire risk, flooding and emergency preparedness.

Global Forest Watch GLAD – tropical forests worldwide

Satellite-based change detection and machine learning for forests and enforcement, operational at global scale through the World Resources Institute, the University of Maryland, Google Earth Engine and public authorities and organisations. Landsat and Sentinel imagery at approximately 30 and 10 metres' resolution respectively requires frequent cloud-free observations, local validation and clear definitions of forest and disturbance. The system provides near-real-time warnings of changes in tree cover, but the warnings indicate disturbance – not automatically illegal logging – and should not be used as a direct measure of annual deforestation area. Evidence A/B for detection; the environmental impact varies according to official action. False positives may be caused by fire, storms, forestry or seasonal variation, and field verification is required before sanctions. Scalability: very high, but Swedish models must be trained for boreal forests and Swedish forestry practices.

Global Fishing Watch – the world's oceans

Multimodal remote sensing, neural networks and tracking analysis for fishing, marine biodiversity and monitoring, operational at global scale. The platform combines AIS positions, satellite radar, optical imagery, night lights and vessel registers; AIS data are biased towards larger vessels and vessels that report, while so-called dark vessels require supplementary sensors. The system has created a global map of industrial fishing and can identify behaviour consistent with fishing, AIS deactivation and vessels missing from normal registers. Evidence A/B; the ecological impact depends on enforcement and management. Risks include incorrect vessel attribution and unevenly distributed monitoring capacity. Scalability: high – relevant to the Baltic Sea, the North Sea, marine protected areas, bottom trawling and lost gear, through cooperation between the Swedish Agency for Marine and Water Management, the Swedish Coast Guard, the Swedish Maritime Administration and neighbouring countries.

Wildlife Insights – camera traps worldwide

Computer vision for species identification and image filtering in biodiversity monitoring, an operational platform at scale driven by organisations including Conservation International, the Smithsonian, WCS, WWF and Google. The models were trained on tens of millions of images and are reported to be able to filter out around 88 per cent of empty images with less than 2 per cent error, sharply reducing manual work. The conservation impact is indirect and depends on the data changing management practices. Evidence B/C. Cameras may capture people, so the platform has controls for sensitive locations, and species locations may need to be protected from hunting and disturbance. Scalability: high for county administrative boards, SLU, nature reserves and infrastructure projects, provided there are shared Swedish taxonomies, validation datasets and rules for images of people.

Rainforest Connection and Arbimon – bioacoustics in tropical forests

Edge sensors, cloud-based audio analysis and machine learning for biodiversity and illegal logging, operational in several projects but not comprehensive. Continuous recordings from solar-powered devices are compared with species libraries and acoustic signatures for phenomena such as chainsaws; quality is affected by wind, rain, microphones and a lack of local reference libraries. Arbimon reports monitoring thousands of species, including hundreds of threatened species, and studies show that automated bioacoustics can track recovery in tropical forests. Evidence A/B. Risks include recording people, sensitive species locations and high communication costs; local involvement and manual checking of alerts are crucial. Scalability: high for targeted environments – Sweden already has relevant bioacoustic data, including a Gothenburg dataset with 10,691 recording hours and more than 239,000 bird occurrences from 30 sites.

UNEP's Methane Alert and Response System (global)

Satellite remote sensing, AI-based signal filtering and atmospheric modelling for methane emissions and environmental enforcement, operational at global scale through UNEP's International Methane Emissions Observatory. Observations from more than 30 satellite instruments are combined with wind and atmospheric data, facility registers and operator feedback, requiring correct plume attribution and ground-based verification. MARS is the first global satellite-based system for detecting and notifying major methane emissions; UNEP reported more than 1.3 million observations analysed since 2023 and documented cases in which notifications were followed by verified action. Evidence B. False attribution can have economic and legal consequences – a warning must be distinguished from legal proof. Scalability: very high for landfills, wastewater facilities, biogas, agriculture and energy infrastructure, where Sweden can combine MARS and Copernicus with drones and ground measurements.

ALERTCalifornia – California, USA

Computer vision on a camera network with edge and cloud AI for wildfires and civil preparedness, operational at state scale through UC San Diego, CAL FIRE and emergency services. Panoramic cameras are combined with topography, weather, previous fire data and operator verification; performance is affected by fog, clouds, sunlight and vegetation. During the first large-scale season, the system was used for more than 1,200 fires and was reported to detect fires before the first 911 call in more than 30 per cent of cases, with reported detection 10–30 minutes earlier. The exact avoided burned area is unspecified. Evidence B/C. Success factors include operators in the alert chain, geographical redundancy and integration into ordinary control centres. Scalability: very high – it should be tested in fire-prone parts of Götaland, Svealand and the coastal areas of Norrland, integrated with SMHI's fire-risk models and municipal emergency services.

ZenRobotics – Finland and Spain

Computer vision, machine learning and industrial robotics for waste and the circular economy, commercially mature. A demonstrator project near Barcelona received approximately EUR 488,000 in EIT funding and roughly the same amount in co-financing. Cameras and sensors above conveyor belts require labelled object data and continuous retraining because waste is heterogeneous and packaging changes. The system is reported to achieve around 2,000 picks per hour per gripper and up to 98 per cent purity in certain fractions, but results vary considerably between material streams. Evidence B/C. The working environment may improve when hazardous manual sorting is reduced, while vendor lock-in and ownership of training data are central procurement issues. Scalability: high for construction and demolition waste, plastics, textiles and municipal recycling facilities – with Swedish test fractions and life-cycle analysis.

PlantVillage Nuru – Africa

Offline computer vision on mobile phones, that is, edge AI, for agriculture and crop protection, distributed as a field-based public service by Penn State, FAO, IITA, CIMMYT and local advisory organisations. Diagnoses are based on mobile images of plant leaves and locally labelled disease cases; results depend on image quality, number of leaves, disease stage and local prevalence. Field studies show approximately 74–88 per cent diagnostic accuracy when several leaves are used, but net reductions in pesticides, emissions or land use are unspecified. Evidence A for diagnostics, D for environmental outcomes. Unequal access to phones and unclear advice on pesticides can cause harm. Scalability: medium – Swedish applications exist in plant diseases, invasive species and forest damage, but advice must explicitly promote integrated pest management.

PUB's intelligent water optimisation – Singapore

AI, scenario modelling, predictive analysis and digital twins for water supply, wastewater and flooding, operational at scale within the water utility PUB together with GovTech. Sensors, demand forecasts, weather, water levels and operating constraints are used; a catchment system handles more than 350,000 data points per day and filters anomalous measurements. Daily production planning fell from up to two hours to approximately 15 minutes, and lowering the desired starting level in service reservoirs by 5 per cent reduced the energy required for treatment and pumping without affecting supply; planning time fell by 64 working days per year. Exact energy and emissions reductions are unspecified. Evidence B. Critical infrastructure requires cybersecurity, redundancy, manual backup procedures and clear limits on autonomous control. Scalability: very high for Swedish water and wastewater organisations in leakage, cloudbursts, infiltration and inflow, and energy, through regional testbeds and shared data models.

Maritime digital twin – Singapore

Digital twin, simulation, real-time data and route optimisation for shipping, ports and coastal development, at early operational scale at the Maritime and Port Authority of Singapore and GovTech. Vessel positions, port operations, shipping lanes, weather and infrastructure require shared identifiers, low latency and regulated access to security-sensitive data. The platform was launched in 2025 and is intended to support a shared operating picture, simulation and optimisation, but publicly verified reductions in fuel use, waiting time or emissions have not yet been specified. Evidence D for environmental outcomes. Risks include cyber threats, commercially sensitive data and the twin becoming an expensive visualisation with no decision-making power. Scalability: medium to high for the Port of Gothenburg, Lake Mälaren, coastal planning and electrified shipping – starting with defined questions such as berth-call optimisation and power demand.

Google Project Green Light – traffic signals in several cities

Machine learning and optimisation based on aggregated mobility patterns in urban transport, operating in at least 18 cities in 2025. Google reports a modelled potential of up to 30 per cent fewer stops and 10 per cent lower emissions at junctions, but the results are partly based on model calculations. Evidence C. Car-centred optimisation can worsen access and safety for pedestrians, cyclists and public transport, or create induced traffic; the objective must be multimodal. Scalability: medium, and benefits should be measured against bus priority, cycling measures, congestion management and changed land use.

A recurring evidence gap runs through the material: many projects report high technical accuracy, the number of data points analysed or working hours saved, but fewer measure long-term changes in emissions, ecosystem status or resource use. Where quantified environmental benefit is absent, it should not be assumed merely because the technology is advanced.

Cross-cutting lessons and transferability to Sweden

AI works best when the action chain is short. Data-centre cooling and water optimisation affect physical systems directly and continuously. Fire and methane warnings have a somewhat longer chain but can be linked to identified operators, response times and actions. Biodiversity platforms and digital twins, by contrast, risk stopping at improved knowledge unless they are integrated into permits, enforcement, area protection, maintenance or investment decisions.

Remote sensing is the clearest low-hanging fruit for Sweden. Sweden has large geographical areas, extensive forest and aquatic environments, access to Copernicus data, national geospatial data and established environmental monitoring. The cost of analysing additional areas can become relatively low once models and data platforms are established.

Swedish environmental data are valuable but organisationally fragmented. Long time series, geospatial data and municipal operational data are held by many actors, but metadata, currency, licences, identifiers and APIs vary. A Swedish AI initiative should therefore fund data quality and information management at least as systematically as model development.

Digital twins should be decision products, not three-dimensional prestige projects. A Swedish municipality rarely needs a complete real-time replica of the entire city; a bounded twin for stormwater, urban heat islands, power demand or construction logistics can deliver greater benefit at lower cost.

Edge AI is particularly interesting where data are sensitive or connectivity is weak. Analysis in the camera, sensor or phone can reduce bandwidth, latency and the need to transfer raw personal or environmental data. Swedish fire cameras, acoustic sensors and water stations should therefore transmit only alerts, uncertainty measures and necessary extracts, while raw data are stored selectively.

Open models are not enough without open evaluation data. Swedish programmes should create reference datasets for, for example, forest damage, methane plumes, waste fractions, traffic scenarios and bioacoustics, with geographically separated test sets and clear rules for sensitive information.

Social objectives must extend beyond efficiency. Traffic optimisation that reduces cars' stops can increase traffic volumes. Forest analysis that maximises timber production can conflict with biodiversity, carbon storage and reindeer husbandry. Digital governance must work with multiple objectives and report trade-offs.

A unified Swedish priority list

• Introduce and scale early: satellite-based environmental monitoring, methane detection, fire warnings, energy forecasting, and water and pump optimisation. Relatively mature technologies, clear operational users and measurable outcomes.
• Pilot with strict impact evaluation: digital twins, AI-controlled traffic signals, robotic waste sorting and bioacoustics at scale. High potential but substantial variation in cost and actual environmental effects.
• Restrict to experiments and support processes: generative AI for environmental decisions, fully autonomous enforcement, automated permit decisions and comprehensive city twins. Weaker evidence, greater rule-of-law risks and the risk of major indirect costs.

Policy recommendations

The short term refers to 2026–2027, the medium term to 2028–2030 and the long term to 2031 onwards. The cost levels are analytical orders-of-magnitude estimates, not finished budget calculations.

National mission programme for AI for the environment and sustainable development

Short term, high cost. The programme should start from societal problems and fund data, organisational change, the model, procurement and evaluation as one whole. Main actors: the Government, the Ministry of Climate and Enterprise, the Ministry of Rural Affairs and Infrastructure, the Swedish Environmental Protection Agency, Vinnova, Formas, the Swedish Energy Agency and AI Sweden. Indicators: at least six cross-sectoral pilots by 2027, 100 per cent with a pre-registered baseline, at least three solutions qualified for scaling in 2028, and reported net benefit per krona invested.

Federated Swedish environmental and planning data space

Short to medium term, high cost. Not one central database, but shared standards, catalogues, identifiers, APIs, access rules and quality statements. Main actors: DIGG, Lantmäteriet, the Swedish Environmental Protection Agency, Boverket, SMHI, SLU, the Swedish Agency for Marine and Water Management, the Swedish Transport Administration, Statistics Sweden and municipalities. Indicators: the proportion of priority datasets with machine-readable metadata, documented currency and uncertainty, API availability, the number of data-sharing municipalities and reduced time spent preparing data.

National system for AI-assisted nature and forest monitoring

Short to medium term, medium–high cost. Combine Sentinel/Landsat, aerial imagery, laser scanning, bioacoustics, camera traps and field inventories. The system should complement, not replace, statistically designed environmental monitoring. Main actors: the Swedish Environmental Protection Agency, the Swedish Forest Agency, SLU, county administrative boards, Lantmäteriet, the Swedish National Space Agency, and landowners and conservation organisations. Indicators: area and species groups with recurring coverage, precision and recall by habitat type, time from change to verified alert, and cost per verified observation.

National programme for methane and point-source emissions monitoring

Short term, medium cost. Integrate UNEP MARS, Copernicus, commercial satellites, drones and ground measurements, and create a standardised alert and reporting process. Main actors: the Swedish Environmental Protection Agency, the Swedish Energy Agency, municipal environmental offices, Avfall Sverige, Svenskt Vatten, the Swedish Board of Agriculture and operators. Indicators: the number of detected and verified emissions, median response time, the proportion of alerts receiving an operator response, tonnes of methane avoided, false alarms and cost per reduced tonne of carbon-dioxide equivalent.

AI-supported national system for early forest-fire detection

Short to medium term, medium–high cost. Combine cameras, satellites, lightning data, fire-risk indices and local edge models. Operators must always verify alerts before action. Main actors: MSB, municipal emergency services, county administrative boards, the Swedish Forest Agency, forest owners, universities and RISE. Indicators: median minutes gained through earlier detection, detection rate, false alarms per camera day, the proportion of alerts verified within five minutes and burned area compared with the historical baseline.

Scale up AI for energy forecasting and flexibility

Short to medium term, medium–high cost. Shared forecasting interfaces should be developed for wind, solar, load, heat demand and grid constraints. Main actors: Svenska kraftnät, the Swedish Energy Agency, SMHI, grid companies, district-heating companies, property owners and municipalities. Indicators: forecast error, reduced renewable generation curtailed in MWh, freed-up peak capacity, reduced peak generation, energy savings and carbon-dioxide equivalents, as well as the number of flexible buildings.

Modular digital twins for water, flooding, heat and planning

Medium term, high cost. The state should fund reusable components rather than having each municipality procure a proprietary all-in-one platform. Main actors: Boverket, Lantmäteriet, SMHI, MSB, Svenskt Vatten, municipalities, regions and the Swedish Transport Administration. Indicators: reduced water loss, kWh per cubic metre of water, the number of overflow events, improved forecasting time for cloudbursts, the number of planning decisions in which scenarios have been documented, and reuse of components between municipalities.

AI-based traffic management with multimodal and climate-policy objectives

Short to medium term, medium cost. The optimisation should prioritise public transport, walking, cycling, road safety, air quality and emissions – not merely cars' journey times. Main actors: the Swedish Transport Administration, the Swedish Transport Agency, municipalities, regions, VTI and public transport authorities. Indicators: carbon dioxide and nitrogen oxides, the number of stops, bus reliability, waiting times for walking and cycling, accidents and conflicts, traffic volume, and effects by neighbourhood and socioeconomic group.

Performance-based procurement for circular material flows

Medium term, high cost. Procurement should reward verified material recycling and quality, not the number of robots or the quantity of AI functionality. Main actors: the Swedish Environmental Protection Agency, the National Agency for Public Procurement, Avfall Sverige, municipalities, the construction sector and producers covered by extended producer responsibility. Indicators: sorting purity, material yield, contamination, tonnes of virgin material replaced, life-cycle emissions, energy use per tonne and occupational safety incidents.

DIGG already has a central role in supporting lawful and effective use of AI in the public sector and recommends that public organisations integrate an AI policy into their ordinary governance system. The proposed environmental AI assessment should build on these structures and be coordinated with IMY, rather than developed as a parallel regulatory framework.

Action plan and funding

Timetable 2026–2032

• 2026 H2: government commission and national programme management; inventory of data, models and ongoing pilots; shared rules for impact assessment and procurement; baselines for energy, water, emissions and biodiversity.
• 2027: first version of the environmental data space; six to ten pilots in energy, water, nature, methane, fire and transport; reference datasets and independent evaluators contracted; skills programme for public authorities, municipalities and water and wastewater organisations.
• 2028: first impact evaluation and public results report; discontinuation or redesign of pilots without verified benefit; scaling of at least three solutions to several regions; shared framework agreements and open technical components.
• 2029–2030: national operation of selected monitoring and warning systems; regional digital twins for water, cloudbursts and energy systems; integration into enforcement, environmental monitoring and physical planning; results-based funding and long-term management budgets.
• 2031–2032: full life-cycle evaluation and international comparison; revision of data and ethical frameworks; continued scaling only where net benefit is positive; planned decommissioning of obsolete models and platforms.

The pilot phase should begin with problems where both the baseline and the decision chain are clear. A fire project should define in advance how earlier detection will be measured, who verifies the alert and how avoided damage will be estimated. A water project should have historical data on leakage, energy, overflow events and operational disruptions. A biodiversity project should define which management action can follow from an AI observation.

Each pilot should use a staged process: problem and mandate, followed by data review, legal and ethical assessment, limited shadow operation, operational pilot with human control, independent impact evaluation and finally a decision on scaling, redesign or discontinuation. During shadow operation, the system produces recommendations or alerts without controlling operations.

Funding

• The state funds shared data infrastructure, standards, reference datasets, security reviews and independent evaluation.
• Operational lead organisations co-finance pilots to ensure ownership within the relevant service and take over management costs after verified benefit.
• Results-based contracts are used where outcomes can be measured robustly, such as reduced leakage or energy demand. Payment should not be based on the supplier's own model calculation.
• EU funding is used for research, demonstration and cross-border solutions. Horizon Europe has an indicative budget of EUR 93.5 billion for 2021–2027; LIFE is the EU's dedicated funding instrument for the environment and climate.

As the current programme period for several EU programmes approaches its end, Swedish actors should both apply for remaining calls in 2026–2027 and influence the design of the next multiannual EU budget. EU funding should not be used to postpone decisions on national operational funding.

The evaluation should preregister indicators and preferably use control areas, phased roll-out or before-and-after comparisons adjusted for weather, economic conditions and other simultaneous measures. Results should be reported both as absolute values and relative change. A ten per cent reduction is inadequate information without a baseline, uncertainty interval and disclosure of whether activity increased at the same time.

Risks, governance and ethical frameworks

The EU AI Act is based on a risk-based model and aims to ensure safe and reliable systems while protecting fundamental rights. Not all environmental AI systems are automatically high-risk systems, but use within critical infrastructure or for decisions affecting people's rights may entail stricter requirements. Classification must therefore be carried out for the specific use case, not for the technology in general.

Cameras, microphones and sensors in nature create privacy risks for people, as well as risks that sensitive information about the locations of threatened species may be disseminated. Openness should be the general rule, but with exceptions where openness itself would harm what is meant to be protected.

The EU's Data Governance Act and Data Act improve the conditions for data sharing and common data products. However, they do not automatically resolve questions about semantics, data quality, the principle of public access to official documents, confidentiality or responsibility for incorrect AI conclusions. These matters need to be operationalised in Swedish sector-specific data contracts.

A Swedish environmental AI impact assessment

• Purpose and additionality: which environmental problem is being solved, what action follows, and why is a simpler method not sufficient?
• Data: who owns the data, how were they collected, which groups and environments are missing, and how long may they be stored?
• Model: what baseline is the model compared with, which errors are most common and how is uncertainty expressed?
• Environmental footprint: how much electricity, water, network capacity and hardware are required for training, operation and sensors?
• Human control: who may amend or reject the result, who is responsible when things go wrong and how can an appeal be made?
• Security: how are manipulation, data drift, intrusion and incorrect updates detected?
• Distributional effects: which areas, organisations or groups receive benefits and which bear costs?
• Decommissioning: how are data exported, how is the system safely shut down and how is the public authority's expertise preserved?

Conclusion

The most important conclusion is that Sweden should prioritise specialised, relatively narrow AI over general and resource-intensive AI platforms. The international cases with the clearest benefits do not primarily use large language models. They use satellite imagery, time series, sensor fusion, optimisation and domain-specific models integrated with a defined physical operation.

Remote sensing and multimodal environmental monitoring have the greatest overall potential. The technology is scalable across large areas, builds on Swedish and European data resources and can strengthen both routine environmental monitoring and rapid enforcement. Here, AI should be used to prioritise where people should investigate, not to automatically establish legal responsibility.

Energy and water optimisation are likely to deliver the fastest measurable resource benefits. Sweden has extensive district heating, electrification, wind power, pumping and treatment systems, as well as a growing need for flexibility. Technically conservative models with safety constraints can deliver greater net benefit than more spectacular generative systems.

Digital twins are strategically important for climate adaptation and sustainable development, but should be developed modularly. Highest priority should go to water, cloudbursts, flooding, heat, energy and transport nodes where scenarios can actually change investment or operations.

Edge AI for fire and biodiversity is promising where time, connectivity and privacy are critical. Swedish pilot projects should keep people in the alert chain and avoid permanent camera and audio networks without a clear purpose, retention period and independent oversight.

Robotic sorting is commercially mature, but should only receive public funding when the quality and life-cycle benefits of materials recycling can be verified. Generative AI is most appropriate as a limited support layer for documentation, search and communication – not for environmental monitoring, critical control, permitting or legal enforcement.

Sweden's overall success factor will not be access to a single algorithm. It will be the ability to combine high-quality data, public-sector domain expertise, clear objectives, long-term management, the rule of law and independent impact measurement.

Source areas

• Swedish environmental monitoring: the Swedish Environmental Protection Agency on environmental monitoring, remote sensing and future research needs.
• Swedish AI governance: DIGG's support for public-sector AI, AI policies and guidance together with IMY.
• Planning and geospatial data: Boverket on digital planning processes and Lantmäteriet on geospatial standards and Geotorget.
• EU regulation: the EU AI Act, Data Governance Act and Data Act, as well as current information on implementation.
• AI's own environmental footprint: the IEA on data-centre electricity use, the OECD on measuring AI's environmental impact and UNEP's life-cycle perspective.
• Energy and weather: DeepMind on data-centre cooling, wind forecasting, GraphCast and GenCast.
• Forests and oceans: Global Forest Watch and Global Fishing Watch, together with related scientific studies.
• Biodiversity: Wildlife Insights, Rainforest Connection, Arbimon and scientific research on bioacoustics.
• Methane and fire: UNEP's Methane Alert and Response System, together with ALERTCalifornia and CAL FIRE.