AI in control: what electrical engineering teaches us about risk and autonomy

I read Edvaldo Santana’s article O quarto escuro da IA (The Dark Room of AI) with interest ^[1]. The provocation is timely. Artificial intelligence systems are becoming more capable, more autonomous and, in some cases, less predictable. There is no way to guarantee that a sufficiently complex model will never fail, produce unexpected behavior, or find an unforeseen way to achieve the objective for which it was developed. My disagreement begins with the consequence we draw from that observation.

The absence of an absolute guarantee has never been, in engineering, sufficient reason to reject a technology. Aircraft can crash, electrical equipment fails, transmission lines go out of service, forecasts miss, and protection systems can also operate incorrectly. Even so, we have built extraordinarily reliable systems because safety does not depend on expecting every component to be infallible. It depends on the architecture designed to deal with failure. That distinction changes how we should think about artificial intelligence.

Zero risk has never been the criterion

A power systems engineer does not design a grid assuming that nothing will go wrong. The opposite assumption is built into the design. Generators may trip, transformers may fail, lines may be disconnected, lightning strikes occur, and actual load rarely matches the forecast exactly. That is why power systems rely on contingency criteria, primary and backup protection, redundancy, selectivity, operating limits, reserves, and restoration procedures.

No one expects to drive the probability of failure mathematically to zero. The goal is to reduce the combination of probability and consequence to an acceptable level and create mechanisms capable of preventing a local failure from becoming a systemic loss. The relevant question, therefore, is not simply whether AI can make a mistake. It is what happens when it does.

If we demand absolute certainty before allowing an intelligent system to participate in a critical function, we will have imposed on AI a requirement that we do not impose on other critical technologies. If, instead, we treat it as a component within a larger architecture, the discussion moves toward operating boundaries, redundancy, observability, validation, containment, and safe failure modes. The debate leaves the almost psychological question of whether we can trust the machine and returns to territory that engineering knows well.

The power system already operates faster than we do

A recurring argument in this discussion is that if a system makes decisions at a speed no human can follow, then the human is no longer in control. In power systems, the relationship between speed and control is more complicated.

A short circuit does not wait for an operator to calmly assess the situation. The protection system detects the abnormal condition and may open a circuit breaker on a timescale incompatible with human reaction. Automatic regulators, converter controls, special protection schemes, and many other forms of automation also operate continuously without asking a person to authorize each command.

That does not put them out of control. Human control took place at another level, when engineers defined logic, parameters, limits, protections, interlocks, and the expected behavior under abnormal conditions. A human can be outside the millisecond control loop and still remain responsible for the architecture that governs that loop. In some applications, in fact, the machine makes the decision precisely because waiting for a human would make the system less safe.

Conventional automation and artificial intelligence are different things. A conventional protection relay is not AI, just as an automatic voltage regulator is not AI. The point here is historical: delegating operational decisions to machines did not begin with artificial intelligence. AI adds the ability to recognize patterns, learn complex relationships, forecast states, and optimize decisions in much more variable environments. The delegated authority may increase, and with it the governance requirements.

This evolution connects directly with the concept of the smart grid. NIST defines a smart grid as a system that integrates information, secure two-way communications, and computational intelligence across the electric power system, including distributed sensors, advanced software, and intelligent and autonomous controllers ^[2]. A smart grid is not synonymous with artificial intelligence. But AI did not introduce intelligence into the power system either. It extends a trajectory of digitalization, sensing, automation, and computational intelligence that was already part of the smart-grid concept.

My own academic relationship with this subject predates the current wave of generative AI by many years. I used genetic algorithms in my master’s research and neural networks in my doctoral work on problems related to the electric power system. The tools have changed enormously since then, but the underlying question remains recognizable: how do we extract value from systems capable of searching, learning, forecasting, and deciding without losing safety, control, and accountability?

AI has already begun to leave the laboratory

During the Florence School of Regulation’s Regulation of the Power Sector course, I looked for concrete examples of artificial intelligence in the electricity sector. At the time, what struck me was how few clearly documented cases there seemed to be compared with the volume of discussion about AI. It was necessary to distinguish expectations, proofs of concept, and actual operation.

The frontier has moved quickly. In Portugal, E-REDES uses the GridWise platform to analyze large volumes of low-voltage network data, combining AI, IoT, big data, and distributed processing to identify outages, anomalies, and losses ^[3]. In Great Britain, the National Energy System Operator developed a probabilistic machine-learning system to dynamically estimate the reserves required by the power system ^[4]. The model uses SHAP values, a method that shows how much each input contributed to the result. The idea comes from game theory, which splits the payoff of a game among the players according to each one’s contribution. In the model, the players are the input variables. So, along with the reserve figure, control-room engineers can see how much of it is due to the wind forecast, solar generation, and other factors.

In Finland, Fingrid began using machine learning to dynamically determine the capacity of its 400 kV lines, and those values now contribute to the transmission constraints provided to the day-ahead market ^[5].

None of these examples means handing an entire national grid over to an autonomous artificial intelligence. What they show is something less spectacular and more important: AI is gradually entering concrete forecasting, diagnostic, optimization, and operational-support functions as its performance can be demonstrated and its risks can be controlled.

What about Brazil?

Brazil is writing its own chapter of this story. Long before the current popularization of generative AI, ANEEL’s regulated Research and Development (R&D) Program already included among its priority subtopics the development of artificial intelligence techniques applied to the control, operation, and protection of electric power systems ^[6]. In the PEQuI 2024-2028, ANEEL’s five-year strategic innovation plan, artificial intelligence appears explicitly among the new strategic technologies for the electricity sector ^[7]. The projects show that this direction did not remain on paper.

In the project Methodology and Tool for Automatic Event Analysis Using Machine Learning Algorithms ^[8], TAESA and Concert Technologies developed a tool that uses machine learning to automatically analyze oscillography records from transmission-line outages. The system identifies the cause of the event, locates the fault, and provides its electrical characteristics to support operators, maintenance teams, and engineering staff. It is a particularly interesting application because AI enters a task associated with diagnosing events in the power system.

The project Predictive Failure Analysis Using Artificial Intelligence: A System for Identifying and Predicting Failures in Power Equipment Based on Sensor Networks and AI ^[9], developed by TAESA with Instituto de Pesquisa Eldorado, brought intelligence into substations. Sensor networks, cameras, IoT, and machine-learning algorithms were combined to recognize current and future equipment states and identify anomalies before they evolve into failures. The objective is continuous monitoring and support for preventive and predictive maintenance.

The Substation Inspection Robot project ^[10], developed by Lactec for State Grid and resulting in Lacbot, takes another step toward autonomy. The robot was designed to inspect high-voltage substations, moving autonomously or under teleoperation, collecting visual and thermal images, and identifying conditions that require intervention. Instead of keeping workers continuously exposed to hazardous environments, intelligence and automation are used to improve safety.

There are applications even closer to energy coordination. In the AI Platform for Utility-Scale BESS + Distributed Generation Management project ^[11], CTG Brasil is developing an artificial-intelligence platform to jointly manage battery energy storage and photovoltaic generation. The experimental environment is connected to the grid and includes an EMS intended to test peak reduction, power smoothing, voltage control, and other services. The project is still under development, but the direction is clear: AI is moving from observing equipment to coordinating energy resources.

These cases matter because they avoid two exaggerations. Brazil has not handed control of the electric power system to an autonomous AI. Nor is it standing still while waiting for other countries to determine how the technology should be used. We are accumulating experience across applications with different levels of criticality, exactly the kind of learning required before moving toward higher levels of autonomy.

AI in control does not mean AI out of control

European regulatory developments help clarify this distinction. The AI Act classifies as high-risk AI systems intended to serve as safety components in the management and operation of critical infrastructure, including the supply of electricity ^[12]. The choice is telling: instead of banning these applications, the regulation subjected them to stricter requirements.

The same regulation requires human oversight for high-risk systems, but relates that oversight to the risks, the level of autonomy, and the context of use. It also requires appropriate accuracy, robustness, and cybersecurity ^[12]. That is very different from requiring a person to manually approve every decision made by a machine.

Ofgem has adopted a similar logic in the British energy sector. Its guidance addresses safety, governance, accountability, competence, transparency, and explainability, and requirements grow with the risk of the application ^[13]. This is the idea of regulating AI in tiers. The AI Act itself follows this logic, sorting systems into four risk levels: unacceptable, high, transparency, and minimal ^[12]. A model that suggests the best time to charge an electric vehicle does not need to carry the same regulatory weight as an algorithm that takes part in grid operation.

To learn alongside the market, Ofgem created the AI Reg Lab, a regulatory laboratory where energy companies test real or hypothetical AI uses against its guidance and existing rules, before an independent panel of experts. In January 2026, it also decided to set up a technical sandbox as a 12-month pilot: a safe and controlled digital space to test AI applications in the energy sector ^[14]. At the IX World Forum on Energy Regulation in Tbilisi, representatives from Ofgem, Japan’s regulator, the Australian Energy Market Commission, and FERC discussed how to reconcile AI applications in forecasting, planning, maintenance, and markets with accountability and system protection. Ofgem itself presented this outcomes-focused, risk-proportionate approach there ^[15].

In the plenary session on technological transformation, I presented a similar architecture in the Energy Conservation 2.0 proposal. In it, the AI Fabric is a distributed layer, with no central intelligence running the entire grid. It connects forecasting, optimization, adaptive control, and verification across different levels. At the household level, it can estimate consumption, solar generation, and battery availability. At the aggregator level, it can forecast how much flexibility thousands of consumers can provide. At the grid level, it can help identify congestion and coordinate when and where that flexibility has the greatest value. All of this remains subject to cybersecurity, privacy, auditability, and externally defined limits ^[16].

This is where the expression “AI in control” needs to be interpreted carefully. An algorithm may control a particular function without receiving unrestricted authority over the system. It may operate within a defined operating envelope, subject to physical limits, independent protections, interlocks, audit records, and automatic conditions for withdrawal from service. AI chooses within a decision space that it does not define for itself.

We do not have to choose between a person authorizing every action and a machine free to decide anything. Between these extremes are several levels of autonomy, and each function can be assigned the level compatible with its criticality, technological maturity, and supervisory capability.

There is also risk in not using AI

The debate often makes an unfair comparison. On one side is an artificial intelligence system subject to failure. On the other, almost implicitly, is a reliable human process. That second system does not exist.

Operators also make mistakes. Traditional forecasts miss. Human teams cannot identify every pattern present in millions of measurements. Fatigue, biases, cognitive limitations, and time constraints all exist. A technically honest comparison sets two imperfect architectures side by side, one with AI and one without it.

If a model better forecasts the availability of renewable generation, it can reduce reserve requirements. If it identifies equipment degradation before it leads to a failure, it can improve reliability. If it automatically analyzes an oscillography record and locates a fault faster, it reduces the time required to understand an event. And a model that estimates a line’s thermal capacity more accurately allows better use of an existing asset without exceeding its physical limits.

In that context, not using AI also has a cost and also creates risk. This dimension is less visible because we readily notice a new risk introduced by a technology while treating the risks of the existing process as part of the natural environment. But natural does not mean optimal.

The correct question becomes which architecture produces lower total risk and greater value for the system, including the opportunity cost of preserving inferior processes simply because they are familiar.

A dark room does not have to remain dark

The image of the dark room works because it describes uncertainty. We do not yet know every behavior that future artificial intelligence systems may produce, nor every way in which they will be integrated into grids, markets, and critical infrastructure. It would be imprudent to treat that uncertainty as irrelevant.

But engineering’s historical response to uncertainty has not been to remain outside the room. When we cannot observe a system well enough, we add instrumentation, test it, establish limits, build redundancy, monitor performance, and continuously compare what we expected with what actually happened. The greater the consequences of failure, the stronger the safeguards should be.

The transformation of power grids will probably require exactly that. More distributed systems, millions of connected resources, variable renewable generation, batteries, electric vehicles, microgrids, and consumers responding to prices and grid conditions create a coordination problem that is unlikely to be solved simply by putting more operators in front of more screens. The European Union has already launched AI.grids to develop AI models for grid planning and management ^[17].

Part of the intelligence required will remain human. Part will be conventional automation. A growing share will be artificial intelligence. The task for engineers and regulators will be to determine which decisions can be delegated, which limits must remain external to the algorithm, how models will be validated, what evidence will be required, when direct human oversight is necessary, and who remains accountable when something goes wrong.

We may indeed be entering a room that we do not yet fully know. But engineering has better tools than simply keeping the door closed.

Humans can leave the millisecond control loop without giving up command of the architecture.

References

1. SANTANA, Edvaldo. O quarto escuro da IA [The Dark Room of AI]. Valor Econômico, Oct. 6, 2026. Available at: https://valor.globo.com/opiniao/coluna/o-quarto-escuro-da-ia.ghtml. Accessed on: Oct. 7, 2026.
2. NATIONAL INSTITUTE OF STANDARDS AND TECHNOLOGY. Smart Grid Communications. NIST. Available at: https://www.nist.gov/programs-projects/smart-grid-communications-0. Accessed on: Oct. 7, 2026.
3. E-REDES. GridWise: A ferramenta que veio revolucionar a gestão da rede de baixa tensão [GridWise: The tool that revolutionized low-voltage network management]. Mar. 4, 2024. Available at: https://www.e-redes.pt/pt-pt/noticias/2024/03/04/gridwise-ferramenta-que-veio-revolucionar-gestao-da-rede-de-baixa-tensao. Accessed on: Oct. 7, 2026.
4. NATIONAL ENERGY SYSTEM OPERATOR. Probabilistic Machine Learning Solution for Dynamic Reserve Setting. Available at: https://www.neso.energy/about/innovation/our-innovation-projects/probabilistic-machine-learning-solution-dynamic-reserve-setting. Accessed on: Oct. 7, 2026.
5. FINGRID. Fingrid introduces Dynamic Line Rating in the transmission grid. Apr. 27, 2026. Available at: https://www.fingrid.fi/en/news/news/2026/fingrid-introduces-dynamic-line-rating-in-the-transmission-grid/. Fingrid half-year report, GlobeNewswire, July 23, 2026: https://www.globenewswire.com/news-release/2026/07/23/3331934/0/en/fingrid-group-s-half-year-report-1-1-30-6-2026.html. Accessed on: Oct. 7, 2026.
6. ANEEL. Supervisão, Controle e Proteção de Sistemas de Energia Elétrica [Supervision, Control, and Protection of Electric Power Systems]. Subtopic SC05. Available at: https://www.gov.br/aneel/pt-br/assuntos/programa-de-pesquisa-desenvolvimento-e-inovacao/temas-para-investimentos/sc. Accessed on: Oct. 7, 2026.
7. ANEEL. TE4: Inovações para Transmissão e Distribuição e Novas Tecnologias de Suporte, Inteligência Artificial, Realidade Virtual e Aumentada e Blockchain [TE4: Innovations for Transmission and Distribution and New Supporting Technologies, Artificial Intelligence, Virtual and Augmented Reality, and Blockchain]. PEQuI 2024-2028. Available at: https://www.gov.br/aneel/pt-br/assuntos/programa-de-pesquisa-desenvolvimento-e-inovacao/temas-estrategicos-do-pequi/te3-inovacoes-para-transmissao-e-distribuicao-e-novas-tecnologias-de-suporte-2013-inteligencia-artificial-realidade-virtual-e-aumentada-e-blockchain. Accessed on: Oct. 7, 2026.
8. TAESA. Projeto 0048: Metodologia e Ferramenta para Análise Automática de Ocorrências Utilizando Algoritmos de Aprendizado de Máquina [Project 0048: Methodology and Tool for Automatic Event Analysis Using Machine Learning Algorithms]. ANEEL code PD-07130-0048/2019. Available at: https://institucional.taesa.com.br/pesquisa/projeto-0048metodologia-e-ferramenta-para-analise-automatica-de-ocorrencias-utilizando-algoritmos-de-aprendizado-de-maquina/. Accessed on: Oct. 7, 2026.
9. TAESA. Projeto 0062: Análise Preditiva de Falhas por Inteligência Artificial [Project 0062: Predictive Failure Analysis Using Artificial Intelligence]. ANEEL code PD-07130-0062/2020. Available at: https://institucional.taesa.com.br/pesquisa/projeto-0062-analise-preditiva-de-falhas-por-inteligencia-artificial-sistema-para-identificacao-e-predicao-de-falhas-em-equipamentos-de-potencia-baseado-em-redes-de-sensores-e-ia/. Accessed on: Oct. 7, 2026.
10. LACTEC. Robô de Inspeção de Subestações [Substation Inspection Robot], R&D 08106-0002-2017. Available at: https://lactec.com.br/2024/08/26/o-que-sao-os-5ds-da-transicao-energetica/. Accessed on: Oct. 7, 2026.
11. CTG BRASIL. Plataforma IA para Gestão de BESS + DG em Utility Scale [AI Platform for Utility-Scale BESS + Distributed Generation Management]. ANEEL code PD-10381-0024. Available at: https://www.ctgbr.com.br/inovacao/nossos-projetos-aprovacao/. Accessed on: Oct. 7, 2026.
12. EUROPEAN UNION. Regulation (EU) 2024/1689, Artificial Intelligence Act, especially Annex III and Articles 14 and 15. Official text on EUR-Lex: https://eur-lex.europa.eu/eli/reg/2024/1689/oj. Annex III on the European Commission’s AI Act Service Desk: https://ai-act-service-desk.ec.europa.eu/en/ai-act/annex-3. Article 14 on the AI Act Service Desk: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-14. Article 15 on the AI Act Service Desk: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-15. AI Act risk levels, European Commission page: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai. Accessed on: Oct. 7, 2026.
13. OFGEM. Ethical AI use in the energy sector, version 2. 2026. Available at: https://www.ofgem.gov.uk/guidance/ethical-ai-use-energy-sector. Accessed on: Oct. 7, 2026.
14. OFGEM. Artificial intelligence. AI Reg Lab, AI technical sandbox, and AI assurance in the energy sector. Available at: https://www.ofgem.gov.uk/energy-regulation/technology-and-innovation/artificial-intelligence. Accessed on: Oct. 7, 2026.
15. INTERNATIONAL CONFEDERATION OF ENERGY REGULATORS; GNERC. IX World Forum on Energy Regulation, Concurrent Session 2C: The impact of AI on energy systems. Tbilisi, Sept. 22, 2026. Speakers: Jonathan Thurlwell, Ofgem; Hiromichi Tanoue, EGC; Victoria Mollard, AEMC; and Lindsay See, FERC. Based on the official program and the author’s personal notes from the session. Available at: https://www.wfertbilisi2026.com/en/programa/day-2. Accessed on: Oct. 7, 2026.
16. ALCÂNTARA, Márcio Venício Pilar. Energy Conservation 2.0, Plenary 3, Technology as a Driver of Energy Transformation, IX WFER. Tbilisi, Sept. 24, 2026. Official Plenary 3 program: https://www.wfertbilisi2026.com/en/programa/day-4. RELOP technical article: https://relop.org/wp-content/uploads/2025/05/Alcantara-Marcio-Rumo-a-Sustentabilidade-Energetica-Conservacao-de-energia-integrada-ao-consumidor-via-redes-inteligentes-e-reforma-regulatoria.pdf. Accessed on: Oct. 7, 2026.
17. EUROPEAN COMMISSION. Flagship projects on AI for grids and data centre sustainability. June 4, 2026. Available at: https://energy.ec.europa.eu/news/flagship-projects-ai-grids-and-data-centre-sustainability-2026-06-04_en. Accessed on: Oct. 7, 2026.

The opinions and analyses expressed in this article are personal and do not represent the positions, decisions, or institutional views of the Brazilian Electricity Regulatory Agency (ANEEL).