Artificial intelligence (AI) is increasingly positioned as a general-purpose capability to help solve complex global problems—optimising energy systems for decarbonisation, accelerating drug discovery, improving medical diagnosis, strengthening early-warning for natural hazards, and enabling more sustainable agriculture. Progress is real but uneven: impressive breakthroughs coexist with data gaps, methodological caveats and implementation hurdles. This report surveys state-of-the-art applications, quantifies potential system benefits where credible estimates exist, examines risks around accuracy, bias and governance, and identifies practical steps needed to turn pilots into durable impact.
Decarbonising and stabilising power: AI for energy optimisation
Electricity systems are becoming more complex as variable renewables scale, demand patterns shift, and grid constraints tighten. AI can reduce costs and emissions by forecasting demand and generation, optimising unit commitment, improving predictive maintenance and reducing curtailment. The International Energy Agency (IEA) estimates that in a “widespread adoption” case, AI-enabled operations and maintenance across power plants could unlock up to USD 110 billion per year in avoided fuels and lower costs by 2035, while also enabling greater renewable integration and delivering material end-use efficiency gains that translate into emissions reductions of roughly 1,400 Mt CO₂ in 2035 if current applications are widely deployed [1][2]. These scenarios are not inevitabilities; they require clean datasets from grid operators, interoperable standards and risk-managed AI deployment at scale. For policymakers and utilities, the priority is to pair AI with grid digitalisation, trusted telemetry and open interfaces so models can act on timely, high-quality data and be audited for safety and fairness [3][4].
Drug discovery and clinical development: speed, scope and scrutiny
In biomedicine, AI is reshaping the pipeline from target identification to lead optimisation and trial design. Systematic reviews in 2024–2025 describe how machine learning is now applied to hit discovery, de-novo molecule and biologic design, structure–activity prediction, patient stratification and synthetic route planning, shortening cycles and opening previously intractable chemical space [5]. Industry activity mirrors this shift: pharma companies are partnering with AI firms and compute providers to scale model-driven discovery; for example, Eli Lilly announced an NVIDIA-powered supercomputer to accelerate simulation and experimentation, embedding AI as “a collaborative scientific partner” across discovery and manufacturing, while other partnerships (e.g., Nabla Bio–Takeda) focus on rapid AI-designed biologics with accelerated design-to-test loops measured in weeks [6][7]. Nature Biotechnology reports that AI-enabled re-design can in some cases compress parts of the traditional 10–15-year path to as little as 1–2 years for specific modalities—though this remains the exception, not the rule, and depends on rigorous validation and regulatory acceptance [8]. The upside is faster candidate discovery, smarter trial design and reduced attrition; the risks involve overfitting to narrow datasets, lack of external validation, and reproducibility. Robust reporting, pre-registration of model claims, and integration with lab automation and causal biology are essential to convert promise into approved therapies [5].
Diagnosis and clinical decision support: where AI helps—and where claims outrun evidence
AI-assisted diagnosis has delivered impressive single-centre results in imaging and dermatology, but generalisability and robustness remain uneven. A 2025 meta-analysis evaluating generative AI for clinical diagnosis found only moderate overall accuracy (around 52%) and concluded these models are not substitutes for expert physicians, though they may assist non-experts and education; the authors call for better external validation and domain-specific specialisation before deployment in consequential settings [9]. Randomised studies of AI-enabled care pathways show potential improvements in decision efficiency and symptom management, but also highlight the need for careful trial design and bias control to demonstrate real-world benefit across diverse populations [10]. Earlier critiques of the evidence base flagged inconsistent reporting and high risk of bias in diagnostic accuracy studies; although methods are improving, transparent datasets, external test sites, and post-deployment monitoring remain non-negotiable for safe scaling [11].
Weather and climate: AI nowcasting and medium-range forecasting
For climate adaptation and disaster preparedness, AI weather models are maturing quickly. Google DeepMind’s GraphCast, published in Science, delivered medium-range forecasts (up to 10 days) more accurately and far faster than the operational gold standard for many variables, while subsequent probabilistic models (e.g., GenCast) report skill exceeding top ensemble systems at the European Centre for Medium-Range Weather Forecasts (ECMWF), bringing high-quality, low-latency guidance within reach for more agencies and firms [12][13][14]. The benefit is not just forecast speed; improved representation of extremes and multiscale patterns can enhance early-warning, supply-chain planning and insurance underwriting. The challenge is that ML-based models can drift outside training regimes and require careful calibration and hybridisation with physics to maintain reliability under non-stationary climate conditions [13].
Earthquakes, tsunamis and volcanoes: early signals and honest limits
Seismology remains the hardest frontier: precise deterministic earthquake prediction is not currently possible, but AI is improving detection, situational awareness and probabilistic forecasting. National geological services and research groups show that machine learning can detect and classify many more micro-events than traditional methods—useful for swarm tracking and hazard assessment—and may modestly improve short-term forecasts in some settings; the British Geological Survey and recent studies on Campi Flegrei illustrate how AI-assisted phase picking and dense data mining can reveal hidden structures and elevate situational alertness without promising impossible certainties [15][16][17]. In volcanology, machine-learning methods trained across many volcanoes can identify common precursors and improve short-term eruption forecasting—including at unobserved sites—while USGS notes that generalised deep models are emerging but still hampered by data scarcity and transferability challenges [18][19][20][21]. For tsunamis, AI-assisted systems using hydrophones and rapid signal classification can provide faster first looks than tide-gauge-only approaches, and ML-based early-warning prototypes have shown promise in case studies, though global reliability still depends on robust sensor networks and conservative alerting thresholds [22][23][24][25]. Overall, AI elevates detection, speeds triage and enhances probabilistic guidance; it does not remove fundamental uncertainties, so risk communication and layered warning systems remain critical.
Food systems and sustainable agriculture: from pixels to practice
AI in agriculture spans precision application of inputs, yield and soil-moisture forecasting, pest and disease detection, supply-chain optimisation and market access tools. Reviews in 2025 map a socio-technical roadmap in which computer vision, remote sensing and decision support reduce inputs and emissions intensity while maintaining yields, especially when paired with agronomy and farmer training [26]. The World Bank notes rising adoption and investment, with Sub-Saharan ag-tech investment growing from under USD 10 million in 2014 to ~USD 600 million by 2022, and a global AI-in-ag market projected to grow at ~23% CAGR from 2023–2028, though value capture depends on connectivity, advisory services and inclusion of smallholders [27][28]. Done well, AI can reduce fertiliser and water waste, target interventions to vulnerable plots, and buffer smallholders from climate shocks via better forecasts and advice; done poorly, it can entrench digital divides or promote input-intensive practices without sustainability guardrails.
Pros: where AI can create system-level value
First, speed and cost efficiency: in energy, AI can trim losses, anticipate failures and balance grids as renewables rise; in research, it can search vast design spaces and prioritise experiments, potentially accelerating therapeutics for neglected diseases [1][5]. Second, capability uplift for public goods: faster and more accurate weather guidance and multi-hazard early-warning can save lives and assets, while agricultural decision support can reduce inputs and emissions intensity [13][26]. Third, access: well-designed AI tools can extend specialist-level support to underserved settings (e.g., triage, imaging pre-reads, agronomic advice), provided they are validated and embedded in care or extension systems [10].
Cons and difficulties: data, drift, accuracy and deployment friction
Three families of challenges recur across domains. Data and representativeness. Training sets often reflect narrow geographies, devices or demographics (e.g., single-centre imaging, temperate-zone agriculture), which can lead to brittle models and inequitable performance. Clinical evidence syntheses continue to flag bias and weak external validation for many diagnostic models, limiting safe generalisation [9][11]. Distribution shift and non-stationarity. Climate change and novel regimes can degrade weather and crop models; pathogens evolve; power systems change topology; hospitals upgrade scanners—models must be monitored, recalibrated and, in safety-critical use, paired with human oversight and fallbacks [13]. Operationalisation and governance. Turning pilots into practice demands secure data pipelines, evaluation sandboxes, procurement standards and incident reporting. Governments emphasise risk management and content authentication to manage exclusion, operational and ethical risks when AI supports public services—principles that translate to health, energy and agriculture programmes seeking public trust [29].
Accuracy and evaluation: how good is “good enough”?
Accuracy requirements depend on context. In clinical decision support, “assistive” tools might be acceptable with moderate accuracy if they demonstrably improve clinician workflow and patient outcomes in trials; for autonomous diagnostic use, significantly higher and consistently externally validated accuracy is required. In weather and multi-hazard early-warning, the yardstick is improvement over operational baselines (e.g., ECMWF ensembles) and reliability in extremes; recent ML models show superior skill for many variables, but hybrid setups and rigorous verification remain prudent [13][12]. In seismo-volcanic contexts, even modest gains in detection and short-term forecasting can be valuable if they feed conservative, well-communicated alerts; however, claims of “prediction” must be carefully bounded to avoid eroding public trust [20][21].
What needs to happen next: from promising pilots to scaled outcomes
Across sectors, five enablers recur.
1) High-integrity data infrastructure. Open, standardised and privacy-preserving datasets (e.g., grid telemetry, de-identified clinical imaging, annotated satellite and agronomic data) are prerequisites for robust models.
2) Transparent, domain-specific evaluation. External test sets, prospective trials and post-deployment monitoring must become routine, with results reported in accessible formats for practitioners and regulators [10].
3) Human-in-the-loop and hybrid modelling. In safety-critical contexts, combine ML with physics and expert oversight; design workflows so AI recommendations are explainable enough for review and override.
4) Capacity building and inclusion. Farmer- and clinician-centred design, extension services, and training programmes ensure tools are usable and equitable, especially in low-resource settings [28].
5) Governance and assurance. Adopt risk-management frameworks, content authentication for public-facing outputs, and incident-reporting pathways; align procurement with evaluation standards so that only validated systems scale in public services [29].
Our view: pragmatic optimism
AI is already useful against global challenges—cutting grid losses, speeding drug design, improving diagnosis support, sharpening forecasts and guiding sustainable farming. But impact depends on disciplined engineering and governance: diverse, high-quality data; rigorous, domain-specific validation; hybrid models that respect physics and expert judgement; and inclusion to ensure benefits reach the people and places that need them most. The near-term wins are abundant and measurable if we focus less on generic benchmarks and more on operational outcomes—emissions avoided, time-to-therapy shortened, false alarms reduced, yields stabilised under climate stress—underpinned by transparent evidence and resilient systems.
Summary: Deployed carefully, AI can accelerate the energy transition, compress biomedical timelines, enhance early-warning for hazards, and support climate-smart agriculture. Realising these gains at scale requires tackling known difficulties—biased or sparse training sets, distribution shift, and inconsistent accuracy claims—through better data infrastructure, evaluation, hybrid modelling and governance. With these foundations, AI can evolve from promising pilots to a dependable pillar of global resilience and sustainable development.
Sources:
[1] IEA — AI for energy optimisation and innovation — link
[2] IEA — AI and climate change (emissions-reduction potential) — link
[3] IEA — Energy and AI (overview, electricity demand & projections) — link
[4] IEA — Energy and AI report (PDF) — link
[5] Serrano et al., 2024 — AI in Drug Discovery (systematic review) — link
[6] Reuters — Lilly partners with NVIDIA on AI supercomputer — link
[7] Reuters — Nabla Bio & Takeda expand AI drug-design partnership — link
[8] Nature Biotechnology — Clinical trials gain intelligence — link
[9] NPJ Digital Medicine — Meta-analysis of generative-AI diagnostic accuracy — link
[10] The Lancet Digital Health — RCTs of AI in care pathways — link
[11] BMJ — Early appraisal of diagnostic AI evidence quality — link
[12] DeepMind — GraphCast overview (Science paper) — link
[13] Nature — Probabilistic weather forecasting with machine learning (GenCast) — link
[14] Nature News — ML weather model generates accurate ensembles — link
[15] British Geological Survey — ML tracking Santorini earthquake swarm — link
[16] Scientific Reports (Nature) — ML for LA earthquake magnitude estimation — link
[17] LiveScience — AI reveals ring fault at Campi Flegrei (Science study coverage) — link
[18] Nature Communications — Transfer learning for short-term eruption forecasts — link
[19] Geophysical Research Letters — ML reveals volcanic precursors — link
[20] USGS — Generalised deep learning for volcano seismicity — link
[21] USGS SIR 2024–5062 — Seismic techniques for eruption forecasting — link
[22] UNESCO IOC — Applying AI to predict tsunamis — link
[23] WMO — Smart solutions for tsunami early-warning (AI + acoustics) — link
[24] Ocean & Coastal Management — ML for tsunami wave forecasting — link
[25] Earth and Space Science — Deep learning for tsunami alert levels — link
[26] Sustainable Agriculture (Elsevier) — AI for sustainable agriculture roadmap — link
[27] World Bank Blog — AI in Sub-Saharan agriculture: trends and investment — link
[28] World Bank — Agricultural Innovation & Technology (evidence) — link
[29] OECD — Governing with AI (public-sector risk & assurance) — link
beFirstComment