1. Introduction

Artificial intelligence is increasingly being incorporated into electricity-sector operations, including load forecasting, generation dispatch, demand response, fault detection, grid balancing, congestion management, outage prediction, voltage control, renewable-energy integration and automated protection systems. These applications can improve speed and accuracy, but electricity systems are safety-critical: an incorrect automated decision can cause cascading outages, equipment damage, instability, financial losses, or risks to public safety.

Human oversight of AI in electricity operations therefore means maintaining meaningful human responsibility and intervention capability over AI-assisted or AI-controlled operational decisions. It is not necessarily a requirement that a human approve every machine action. Rather, the level of human involvement should correspond to the risk, autonomy, speed and consequences of the AI system.

The modern regulatory approach is moving toward a combination of:

human supervision;

human override;

operational accountability;

explainability and accessible outputs;

logging and auditability;

testing and validation;

cybersecurity;

fail-safe mechanisms; and

appropriately trained system operators.

The EU AI Act expressly requires effective human oversight for high-risk AI systems, while electricity-sector reliability rules in the United States already regulate highly automated grid systems without necessarily requiring human initiation of every protective action. (EUR-Lex)

2. Meaning of Human Oversight

Human oversight can be understood as the institutional and technical capacity of qualified personnel to understand, supervise, challenge, intervene in, override, or discontinue an AI system when necessary.

There are several levels:

A. Human-in-the-loop

The AI produces a recommendation, but a human operator must approve the action.

Example:
AI recommends redispatching 500 MW from Generator A to Generator B. The system operator reviews the recommendation and authorises the dispatch.

B. Human-on-the-loop

The AI can execute actions automatically, but qualified personnel continuously monitor its operation and retain an intervention or override capability.

Example:
An automated voltage-control system adjusts reactive power while the control-room operator monitors the system and can disable the automation.

C. Human-in-command

The human establishes the objectives, operating limits and authority boundaries within which the AI may act.

This is particularly important for autonomous grid-control systems.

D. Human-out-of-the-loop

The AI operates without meaningful human intervention or realistic intervention capability.

This is potentially problematic where an AI system performs safety-critical electricity functions because a nominal "human supervisor" may not realistically be able to understand or stop an action occurring within seconds.

3. Why Human Oversight Is Particularly Important in Electricity Systems

Electricity networks are different from many ordinary digital applications because their physical state changes continuously.

An AI system may have to make decisions concerning:

frequency;

voltage;

power flows;

generation dispatch;

transmission constraints;

load shedding;

battery dispatch;

renewable-energy curtailment;

restoration following an outage;

protection systems;

demand response; and

system stability.

A defective AI decision can therefore have consequences beyond the individual transaction.

The EU Commission's AI Act guidance specifically identifies AI used for detecting anomalies in electricity-grid operation for purposes connected with critical functions such as load distribution, grid stability and shutdown procedures as potentially falling within the high-risk critical-infrastructure framework when the AI functions as a safety component. (AI Act Service Desk)

At the same time, the guidance distinguishes between AI that directly performs a safety function and AI that merely provides forecasts or recommendations to human operators. For example, a forecasting tool whose output is reviewed by grid operators may not automatically constitute a high-risk safety component. (AI Act Service Desk)

This distinction is legally significant because not every AI used by an electricity company requires the same degree of human control.

4. EU AI Act and Human Oversight

The strongest express statutory framework comes from the EU Artificial Intelligence Act, Regulation (EU) 2024/1689.

Article 14 requires high-risk AI systems to be designed so that they can be effectively overseen by natural persons during the period in which they are used. The objective is to prevent or minimise risks to health, safety and fundamental rights. (EUR-Lex)

The level of oversight must be proportionate to:

the risks associated with the AI;

the level of autonomy; and

the context in which it is used.

This is highly relevant to electricity operations.

Example

Suppose an AI system merely predicts tomorrow's electricity demand.

Human oversight may consist primarily of:

reviewing prediction accuracy;

identifying anomalous forecasts; and

deciding whether the forecast should be incorporated into planning.

But an AI system capable of automatically disconnecting major industrial loads or changing network configuration would require substantially stronger safeguards.

5. Practical Powers of the Human Operator

Effective oversight should give the operator genuine capabilities rather than merely placing a human name on an automated process.

The oversight framework should enable the operator, where appropriate, to:

1. Understand the AI output

The operator should receive information sufficient to interpret the recommendation or action.

2. Detect anomalies

Operators should be able to identify circumstances where the AI is operating outside expected conditions.

3. Override the AI

An operator should have an effective mechanism to prevent or reverse an AI action where technically feasible.

4. Interrupt the system

Where continuing automation creates unacceptable risk, the operator should be capable of stopping or isolating the AI system.

5. Reject recommendations

AI recommendations should not automatically become operational instructions.

6. Escalate unusual events

AI systems should have procedures for transferring difficult or uncertain situations to appropriately qualified personnel.

These principles correspond closely to Article 14 of the EU AI Act, which requires human oversight to be designed into the system rather than treated merely as an organisational afterthought. (EUR-Lex)

6. Human Oversight and NERC Electricity Reliability Standards

The North American electricity system provides an important comparative example.

The North American Electric Reliability Corporation (NERC) maintains mandatory reliability standards covering areas including:

transmission operations;

emergency preparedness;

protection and control;

critical infrastructure protection;

personnel performance and training;

balancing;

modelling; and

system operations. (NERC)

An important point is that electricity reliability law does not always require human initiation of every automated action.

NERC's CIP-002 framework, for example, specifically addresses automated load-shedding systems capable of shedding 300 MW or more without human operator initiation. Such systems can nevertheless fall within cybersecurity categorisation because their failure or compromise can have significant consequences for Bulk Electric System reliability. (NERC)

This demonstrates an important legal principle:

Human oversight does not necessarily mean human approval of every real-time grid action.

In some circumstances, an automatic protection system must act faster than a human operator could.

The proper legal question is therefore whether the overall system has adequate human governance, testing, monitoring, intervention and accountability mechanisms.

7. Human Oversight Versus Automatic Protection

This distinction is crucial.

Consider an under-frequency load-shedding system.

If frequency suddenly collapses, an automatic protection mechanism may disconnect predetermined loads within milliseconds or seconds.

Requiring a human to approve that action could actually increase risk.

Consequently, a sensible legal framework distinguishes between:

Automatic protection

and

Autonomous decision-making.

Automatic protection generally operates according to predetermined engineering rules.

AI autonomy may involve:

dynamic learning;

changing strategies;

probabilistic recommendations;

adaptive optimisation;

novel responses to previously unseen situations.

The latter creates a stronger need for human supervision because the system may behave differently from traditional deterministic protection equipment.

8. Human Oversight and Operator Competence

Oversight is ineffective if the human operator does not understand the technology sufficiently to challenge it.

NERC's PER-006-1 requires personnel to receive specific training on matters essential to reliable real-time operation of the Bulk Electric System. (NERC)

For AI-enabled electricity systems, training should increasingly cover:

AI limitations;

false positives;

false negatives;

model uncertainty;

data-quality problems;

model drift;

adversarial manipulation;

automation bias;

appropriate override procedures;

cybersecurity;

failure modes; and

circumstances requiring escalation.

Thus, human oversight is partly a competence obligation.

A control-room operator cannot exercise meaningful oversight if the AI's output is effectively treated as unquestionable.

9. Automation Bias

One of the major risks is automation bias.

Automation bias occurs when humans place excessive trust in automated recommendations.

For example:

AI predicts that a transmission corridor can safely carry additional power. The operator accepts the recommendation without independently checking an unusual weather event or equipment condition.

Even though a human technically made the final decision, the human may have exercised little meaningful judgment.

Therefore, the legal concept of oversight should not be reduced to:

"A human clicked the approval button."

Instead, meaningful oversight requires an opportunity for independent assessment and intervention.

10. Auditability and Record-Keeping

Human oversight also requires evidence showing:

what the AI recommended;

what data it relied upon;

what action it took;

whether an operator intervened;

whether the operator accepted or rejected the recommendation;

what warnings were displayed;

whether the AI was operating within its authorised parameters; and

what happened after an incident.

This is important for regulatory investigations and liability proceedings.

NERC's cybersecurity framework similarly emphasises protection of systems whose compromise could cause misoperation or instability of the Bulk Electric System. (NERC)

For AI systems, logging therefore becomes part of the accountability architecture.

11. Cybersecurity as Part of Human Oversight

An AI system cannot be meaningfully supervised if an attacker can manipulate:

its training data;

operational inputs;

communications;

control commands;

model parameters; or

software configuration.

NERC's CIP standards are designed to protect critical cyber systems against compromise capable of causing misoperation or instability. (NERC)

Consequently, AI oversight should include:

AI safety + cybersecurity + operational governance.

A human operator supervising corrupted AI outputs is not exercising effective oversight.

12. Indian Electricity Law Context

India does not yet have a single comprehensive statutory framework specifically regulating AI oversight in electricity-grid operations comparable to Article 14 of the EU AI Act.

Nevertheless, human responsibility can be derived from the broader electricity regulatory structure.

The Electricity Act, 2003, grid codes, CERC regulations and system-operation arrangements establish institutional responsibilities for secure and coordinated electricity-system operation.

Indian electricity jurisprudence has repeatedly recognised that grid operation involves technical coordination and maintenance of system frequency and stability.

For example, in Power Grid Corporation of India Ltd. v. Chhattisgarh State Electricity Regulatory Commission, the Appellate Tribunal for Electricity discussed the UI mechanism and its objective of maintaining grid frequency and ensuring smooth and integrated operation of the regional and national grid. (Indian Kanoon)

The case is not an AI case, but it is relevant to the legal principle that grid operation is a regulated technical responsibility rather than merely a private commercial activity.

Similarly, Indian courts have examined the physical and operational structure of interconnected electricity grids. In Hindalco Industries Ltd. v. Gujarat Electricity Transmission Corporation Ltd., the Gujarat High Court considered the operation of interconnected transmission systems and the technical requirements associated with parallel operation. (Indian Kanoon)

These cases can therefore be used as foundational authorities when analysing how responsibility for AI-assisted grid decisions could fit into existing electricity law.

13. Case Law: National Grid Electricity Transmission Plc v ABB Ltd

A useful comparative authority is National Grid Electricity Transmission Plc v ABB Ltd & Others, concerning transmission equipment and competition-related issues.

The litigation illustrates an important broader principle: electricity-network operators and technology suppliers remain subject to legal responsibilities concerning the performance and consequences of infrastructure and equipment.

The case is particularly useful in an AI context because AI will increasingly become embedded within electricity infrastructure rather than existing as an isolated software product. (vLex)

Its relevance is therefore conceptual:

Deploying sophisticated technology does not automatically eliminate the legal responsibility associated with operating critical electricity infrastructure.

For AI systems, this supports the argument that utilities should establish clear allocation of responsibility between:

utility operators;

AI developers;

equipment manufacturers;

system integrators; and

control-room personnel.

14. Case Law: National Grid Plc v Gas and Electricity Markets Authority

In National Grid Plc v Gas and Electricity Markets Authority, the UK courts considered regulatory responsibility and the circumstances in which National Grid could be held responsible for conduct under the applicable electricity-market framework.

The Court of Appeal noted the significance of the statutory requirement concerning intentional or negligent infringement and considered the regulatory circumstances surrounding National Grid's conduct. (Competition Appeal Tribunal)

The case is relevant to AI governance because it demonstrates that regulated electricity operators cannot necessarily avoid regulatory responsibility simply because operational decisions involve complex systems or third-party arrangements.

In an AI environment, an electricity operator could therefore face regulatory scrutiny where an AI-assisted process causes a violation of an operational or market rule.

15. Case Law: Power Grid Corporation of India Ltd v Chhattisgarh State Electricity Regulatory Commission

This Indian authority is particularly relevant to the principle of system reliability.

The Appellate Tribunal emphasised that the UI mechanism was designed to maintain grid frequency within the prescribed band and support smooth integrated operation of the regional and national grid. (Indian Kanoon)

Applied to AI systems, the principle suggests that an AI deployment should not be assessed merely according to whether it improves commercial efficiency.

It must also be assessed according to whether it supports:

frequency stability;

secure grid operation;

coordinated dispatch;

compliance with grid requirements; and

system reliability.

16. Case Law: India Energy Exchange Ltd v CERC

A particularly interesting recent Indian development concerns algorithmic electricity-market processes.

In India Energy Exchange Ltd v Central Electricity Regulatory Commission, decided in 2026, the litigation concerned CERC's directions relating to implementation of market coupling and a shadow pilot involving automated market-coupling processes and software developed by Grid-India.

The judgment discusses issues concerning the development and validation of the coupling engine, historical-data testing, software implementation and transparency surrounding the algorithmic basis of the process. (Indian Kanoon)

Although this is not a case specifically about AI human oversight, it is highly relevant to algorithmic governance in electricity markets.

It illustrates why regulators may need to consider:

transparency of algorithms;

validation;

testing;

stakeholder consultation;

software development;

operational accountability; and

regulatory supervision.

This provides a useful bridge between conventional electricity regulation and future AI-controlled electricity markets.

17. Liability for AI Errors

Human oversight also affects the question of liability.

Suppose an AI dispatch system recommends an unsafe generation schedule and the operator accepts it.

Potentially relevant parties could include:

AI provider — defective design, inadequate documentation or foreseeable model failure;

utility/deployer — inadequate governance, training or monitoring;

system operator — negligent failure to follow operational procedures;

integrator — improper integration with operational technology;

equipment manufacturer — hardware or protection-system defects; and

regulatory authority — potentially relevant only in circumstances recognised by applicable public-law principles.

The existence of AI therefore creates a potential distributed responsibility problem.

A good regulatory framework should prevent the "AI did it" defence from becoming a mechanism for avoiding accountability.

18. The Reasonable Human Operator Standard

A useful legal model would ask:

Would a reasonably competent electricity-system operator, having access to the available information and appropriate AI oversight tools, have recognised the danger and taken reasonable corrective action?

This resembles familiar negligence concepts while adapting them to AI-assisted operations.

Relevant factors could include:

seriousness of potential harm;

speed required for intervention;

reliability of the AI;

operator training;

quality of available data;

warnings generated by the AI;

availability of an override;

documented operating procedures;

previous system failures; and

industry standards.

The standard should not require humans to outperform AI mathematically. Rather, it should require reasonable operational supervision consistent with the risks of the system.

19. Human Override

Human override should be carefully designed.

An override mechanism should be:

accessible;

technically reliable;

independently tested;

protected against unauthorised use;

available within the relevant operational timeframe;

documented; and

subject to periodic testing.

However, an override should not necessarily be required for every automated protection function.

For example:

Automatic protective relay:
Immediate autonomous action may be necessary.

AI-based strategic dispatch:
Human review may be appropriate.

AI forecasting tool:
Human verification may be sufficient.

AI capable of autonomous network reconfiguration:
More extensive monitoring, authority boundaries and intervention mechanisms may be necessary.

20. Risk-Based Model of Human Oversight

A useful legal framework can be represented as follows:

The precise legal requirement will depend on the jurisdiction, technology and operational context.

21. Institutional Governance

Electricity companies should establish an AI Operations Governance Framework containing:

A. AI inventory

Identify every AI system used in operational processes.

B. Risk classification

Classify systems according to their potential consequences.

C. Authority matrix

Specify who can:

approve;

modify;

override;

deactivate; and

investigate the system.

D. Validation

AI models should undergo testing before deployment.

E. Continuous monitoring

Performance should be monitored for drift and abnormal behaviour.

F. Incident response

There should be procedures for AI malfunction and cyber compromise.

G. Human training

Operators should understand both the capabilities and limitations of the system.

H. Audit trails

Operational decisions should remain reconstructable.

I. Independent review

High-risk systems should periodically undergo technical and regulatory review.

22. Emerging EU Energy-AI Governance

The European Commission's 2026 energy-policy work explicitly recognises the importance of transparency, explainability and human oversight for AI incorporated into critical energy infrastructure. It also proposes an AI energy-safety transformation group and regulatory sandboxes for testing and validation of energy AI applications. (EUR-Lex)

This indicates a movement from merely regulating AI software toward treating AI safety as part of critical-energy-system governance.

23. Key Legal Principles

The doctrine of human oversight in electricity operations can therefore be reduced to several principles:

Principle 1 — Accountability cannot disappear through automation

The introduction of AI does not automatically eliminate the legal responsibility of the electricity operator.

Principle 2 — Oversight must be meaningful

A human must have sufficient information, competence and authority to intervene where necessary.

Principle 3 — Oversight must be risk-proportionate

A forecasting model and an autonomous grid-control system cannot reasonably be regulated in exactly the same way.

Principle 4 — Automation can sometimes be necessary

Emergency protection systems may legitimately act faster than human operators.

Principle 5 — Override capability matters

Where human intervention is technically feasible and safety-relevant, effective intervention mechanisms should be available.

Principle 6 — Cybersecurity is part of AI governance

An AI system cannot be considered reliable if its operational inputs or commands can be manipulated.

Principle 7 — Records support accountability

AI recommendations, decisions, overrides and incidents should be auditable.

24. Conclusion

Human oversight of AI in electricity operations is becoming an essential component of energy-law governance. The objective is not to prevent automation but to ensure that automation remains subject to appropriate legal, technical and institutional controls.

The EU AI Act provides an explicit framework requiring effective human oversight for relevant high-risk AI systems and requiring that oversight to correspond to the system's risk, autonomy and context. (EUR-Lex) NERC's reliability framework demonstrates a complementary model in which highly automated electricity functions may operate without direct human initiation while remaining subject to cybersecurity, reliability and personnel requirements. (NERC)

Indian electricity jurisprudence, including Power Grid Corporation of India Ltd v Chhattisgarh State Electricity Regulatory Commission, provides an important foundation concerning the legal importance of secure and integrated grid operation. (Indian Kanoon) More recent proceedings concerning algorithmic market coupling also demonstrate the growing importance of software validation, transparency and regulatory supervision in India's electricity sector. (Indian Kanoon)

Ultimately, the appropriate legal model is not "human versus AI." It is AI-assisted electricity governance under clearly defined human responsibility. AI may forecast, optimise, detect and, in carefully controlled circumstances, automatically act. But the surrounding legal architecture should establish who has authority, who has responsibility, when intervention is required, how intervention occurs, and how the decision can subsequently be audited.

Key authorities

Regulation (EU) 2024/1689 (EU AI Act), Article 14 — Human Oversight. (EUR-Lex)

EU AI Act — Critical Infrastructure / Electricity guidance. (AI Act Service Desk)

NERC CIP-002 — BES Cyber System Categorization and automated load-shedding systems. (NERC)

NERC PER-006-1 — Training for real-time Bulk Electric System operations. (NERC)

Power Grid Corporation of India Ltd v Chhattisgarh State Electricity Regulatory Commission — grid frequency and integrated grid operation. (Indian Kanoon)

Hindalco Industries Ltd v Gujarat Electricity Transmission Corporation Ltd — interconnected grid and technical operation. (Indian Kanoon)

India Energy Exchange Ltd v CERC (2026) — algorithmic market coupling, software development, validation and regulatory transparency. (Indian Kanoon)

National Grid Plc v Gas and Electricity Markets Authority — regulatory responsibility in electricity markets. (Competition Appeal Tribunal)