1. Introduction

Artificial intelligence is increasingly used for forecasting rather than merely processing historical information. AI systems can predict electricity demand, renewable-energy generation, equipment failures, wholesale prices, congestion, system imbalances, weather conditions, and consumer demand. In an electricity system, such forecasts may influence decisions about generation scheduling, reserve procurement, transmission operation, storage dispatch and demand response.

The central legal question is therefore not simply whether an AI forecast is accurate. It is:

When an AI forecast can materially affect safety, reliability, economic interests or fundamental rights, must a human decision-maker have the authority and practical ability to reject or override it?

The emerging answer is increasingly yes for high-risk applications, although the precise legal obligation depends on the jurisdiction, sector and consequences of the forecast.

The EU AI Act expressly requires human oversight for high-risk AI and gives designated persons the ability, where appropriate, to disregard, override or reverse AI output and to interrupt the system. (EUR-Lex) In the energy sector, Ofgem's current guidance specifically addresses AI used for predictions and forecasting and recommends explanation of the methodology and factors influencing predictions. (Ofgem)

2. Meaning of Human Override in AI Forecasting

A human override mechanism is a legal, organisational and technical arrangement under which an authorised person can reject, modify, suspend or reverse an AI-generated forecast or a decision based upon it.

For example:

AI forecast:

Expected electricity demand at 18:00 = 82 GW.

Human operator:

The forecast is inconsistent with an unexpected weather event and current system conditions. I reject the forecast and use an alternative operational estimate.

A genuine override system therefore requires more than merely placing a person somewhere in the workflow.

There should be:

Authority to reject the forecast.

Information sufficient to understand the forecast.

Time to intervene.

Technical functionality allowing intervention.

Training to recognise erroneous forecasts.

Documentation of the override.

Accountability for the resulting decision.

Post-event review of the AI's performance.

A human who can technically click "override" but is required by organisational policy to accept the AI output does not possess meaningful control.

3. Why AI Forecasts Require Human Override

A. Forecasts are inherently probabilistic

An AI forecast is not necessarily a statement of fact. It is normally a prediction based upon historical and real-time data.

For example:

P(Dt∣Xt)P(D_t | X_t)

may represent the predicted probability of electricity demand DtD_t given available information XtX_t.

Unexpected circumstances can make the historical relationship unreliable.

Examples include:

extreme weather;

sudden generation outages;

cyber incidents;

transmission failures;

unexpected industrial demand;

unusual consumer behaviour;

market manipulation;

extraordinary geopolitical events.

Human operators may possess contextual information that the model does not.

B. Automation bias

One of the most important risks is automation bias—the tendency of humans to accept computer-generated recommendations simply because they are generated by a sophisticated system.

The EU AI Act expressly requires human oversight arrangements to address the possibility of people over-relying on AI output. It also requires appropriately trained personnel to understand the system's capabilities and limitations. (EUR-Lex)

Thus, human override is intended to prevent:

AI recommendation → automatic human acceptance → consequential decision

from becoming a disguised form of automated decision-making.

4. Human Override as a Legal Principle

Human override can be understood through five connected legal principles.

4.1 Accountability

Someone must remain legally responsible for the consequential decision.

An electricity-system operator should not be able to say:

"The AI made the decision."

AI does not ordinarily displace the legal responsibilities of the licensed utility, system operator, regulator or public authority.

The UK's Data and AI Ethics Framework similarly states that people should remain responsible for decisions made with AI support. (GOV.UK)

4.2 Procedural fairness

Where an AI forecast substantially affects an individual or organisation, procedural safeguards may require meaningful human consideration.

A human must be capable of asking:

Is the forecast reliable?

What information produced it?

Are there unusual circumstances?

Is the model operating outside its validated conditions?

Does contradictory evidence exist?

Should the forecast be rejected?

4.3 Proportionality

Human intervention should increase as the potential consequences increase.

A simple weather forecast used for routine planning may require limited oversight.

An AI forecast used to determine whether an electricity system should disconnect a large industrial consumer, curtail renewable generation or initiate emergency measures requires considerably stronger safeguards.

This risk-based approach is consistent with Article 14 of the EU AI Act, which requires human oversight to be proportionate to the risks, level of autonomy and context of use. (EUR-Lex)

4.4 Transparency

An operator cannot effectively override an AI system if the operator cannot understand what the system is predicting and why.

Therefore, an AI forecasting system should ideally provide:

forecast value;

confidence interval;

relevant input variables;

model version;

data timestamp;

known limitations;

anomaly warnings;

comparison with alternative forecasts.

Ofgem's 2026 energy-sector guidance specifically adds consideration of explanations concerning the methodology and factors influencing AI predictions. (Ofgem)

4.5 Contestability

A forecast should be capable of being challenged.

The principle can be expressed as:

No consequential AI forecast should become legally unchallengeable merely because it was generated by a machine.

This is particularly important when the forecast influences consumer treatment, market access, system restrictions or allocation of scarce electricity resources.

5. EU AI Act and Article 14

The strongest statutory example is 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. The objective is to prevent or minimise risks to health, safety and fundamental rights. (EUR-Lex)

Article 14(4) is particularly significant.

The designated human should, where appropriate and proportionate, be capable of:

understanding the AI system's capabilities and limitations;

monitoring its operation;

detecting anomalies and malfunctions;

recognising automation bias;

correctly interpreting its output;

deciding not to use the AI;

disregarding, overriding or reversing its output; and

intervening in or stopping the system.

(EUR-Lex)

This establishes an important distinction:

Formal human involvement

A human sees the forecast.

Meaningful human oversight

A competent human understands the forecast, can question it and has actual authority and technical capability to reject it.

The second is much closer to the modern regulatory concept of human oversight.

6. Case Law

Case 1: State v. Loomis, 881 N.W.2d 749 (Wis. 2016)

This is one of the most important judicial decisions concerning algorithmic decision support.

The Wisconsin Supreme Court considered the use of the COMPAS risk-assessment system during criminal sentencing. The court permitted consideration of the algorithmic assessment, but imposed significant cautions concerning its use. (Justia Law)

The court recognised that the assessment should not simply determine the sentence. It emphasised the importance of other independent factors and judicial discretion.

Significantly, the judgment discussed the need for professional judgment and the possibility of overriding the computed risk assessment when circumstances warranted it. (Justia Law)

Principle for AI forecasting

The analogy to energy forecasting is strong:

AI output may inform a legally responsible decision-maker, but it should not automatically substitute for that decision-maker's judgment.

For example, an AI system might forecast a low probability of a transmission constraint. If the system operator possesses credible real-time evidence indicating that the forecast is unreliable, the operator should be able to depart from the forecast.

Loomis therefore supports the distinction between:

considering AI output
and
being controlled by AI output.

7. Case 2: OQ v Land Hessen (SCHUFA), Case C-634/21

The Court of Justice of the European Union decided SCHUFA on 7 December 2023.

The case concerned automated credit scoring under Article 22 GDPR. SCHUFA generated a probability score concerning an individual's ability to meet financial obligations, and that score was used by third parties in deciding whether to establish a contractual relationship. (Infocuria)

The Court held that automated establishment of such a probability value can itself constitute automated decision-making where the third party's decision is effectively determined by the score. (Infocuria)

Importance for AI forecasts

This case is highly relevant because it demonstrates that formal human involvement does not necessarily eliminate automation.

Suppose:

AI produces Forecast A → human routinely accepts Forecast A → consequential decision follows.

Calling the human a "reviewer" does not necessarily make the process genuinely human-controlled.

The important question is whether the human actually exercises independent judgment.

8. Case 3: SyRI — The Hague District Court, ECLI:NL:RBDHA:2020:1878

The Dutch SyRI litigation concerned a government system used to identify possible fraud involving social benefits, allowances and taxation.

The Hague District Court held that the legislation regulating SyRI violated Article 8 of the European Convention on Human Rights because the interference with private life did not satisfy the required legal safeguards. (Rechtspraak)

Although SyRI was not an electricity forecasting case, it illustrates an important principle:

A technically sophisticated algorithm does not remove the need for legally adequate safeguards when automated systems affect individuals.

For energy systems, this principle becomes relevant where AI forecasts are used for:

consumer profiling;

energy poverty interventions;

demand-response targeting;

disconnection risk;

dynamic pricing;

allocation of scarce electricity services.

9. Indian Constitutional Principles

India does not yet have a Supreme Court decision establishing a general statutory "human override right" for AI forecasts. Therefore, care should be taken not to present the following cases as direct AI-forecast precedents.

Nevertheless, Indian constitutional jurisprudence supplies principles that can inform the design of AI-based administrative systems.

K.S. Puttaswamy v Union of India

The Supreme Court's privacy jurisprudence establishes legality, legitimate state purpose and proportionality as important requirements when state action interferes with protected interests. (Sci API)

For AI forecasting, this can support a framework in which public authorities must ask:

Is the use of AI legally authorised?

What legitimate objective is being pursued?

Is the AI method rationally connected to that objective?

Are safeguards adequate?

Is a less intrusive or less risky method available?

The principle becomes particularly relevant where AI forecasting involves extensive personal or consumer data.

10. Indian Judicial Treatment of AI Itself

A particularly important recent development is the Supreme Court of India's 2026 decision in Pooja Ramesh Singh v Jammu and Kashmir Bank Ltd., 2026 INSC 668.

The Court addressed the use of AI-generated false or hallucinated legal authorities. According to the Supreme Court's official judgment summary, the Court set aside decisions that relied upon fabricated AI-generated citations and emphasised the need for human control and verification of AI outputs. (Scientific Ministries)

The Court distinguished legitimate assistance from allowing AI-generated material to substitute for human legal reasoning.

Although this case concerns AI-generated legal material rather than forecasting, its broader significance is directly relevant:

AI output must remain subject to human verification when it enters a legally consequential decision-making process.

This is an especially useful Indian authority for explaining why "human override" should mean genuine verification rather than merely formal human presence.

11. Application to Electricity Forecasting

The issue becomes particularly important in electricity regulation.

AI forecasts may concern:

The UK's current energy-policy work illustrates this direction. The government's 2026 vision for an AI-enabled clean-energy system recognises that future systems may use autonomous optimisation but identifies meaningful human oversight as a central risk, particularly where humans remain formally accountable while lacking the ability to scrutinise AI decisions. (GOV.UK)

12. Minimum Legal Requirements for AI Forecasting

A robust regulatory framework should establish at least the following requirements.

1. Human authority

A designated person must have legal and organisational authority to reject the forecast.

2. Technical override

The software must actually permit the operator to:

reject;

replace;

modify;

suspend; or

ignore

the AI output.

3. Competence

The person exercising oversight should understand:

model limitations;

uncertainty;

data quality;

known failure modes;

operational consequences.

4. Explainability

The operator should receive sufficient information to understand the forecast.

5. Confidence information

The system should distinguish between:

high-confidence prediction

and

uncertain prediction.

6. Anomaly detection

The system should automatically alert the operator when current conditions depart substantially from training or validation conditions.

7. Override logging

Every significant override should record:

original AI forecast;

time;

operator;

reason for override;

replacement forecast;

subsequent actual outcome.

8. Independent review

Repeated overrides should trigger model investigation.

9. No retaliation for legitimate override

Operators should not be discouraged from overriding AI because of commercial or managerial pressure.

10. Emergency stop

Critical systems should have a safe mechanism for suspending automated operations.

13. Human Override and Energy-System Safety

The principle can be expressed as a hierarchy:

Level 1 — AI advisory

AI provides a forecast; human decides.

Level 2 — AI recommendation

AI recommends an operational action; human approves or rejects.

Level 3 — AI-assisted automation

AI executes routine actions within predefined limits; human can intervene.

Level 4 — Conditional autonomy

AI operates autonomously within defined safety boundaries but escalates abnormal circumstances to humans.

Level 5 — Full autonomy

AI independently controls the system with little or no meaningful human intervention.

The legal risk generally increases as the system moves from Level 1 toward Level 5, particularly when failure can affect electricity reliability, safety or fundamental rights.

14. Human Override Is Not the Same as Human Presence

This is one of the most important legal distinctions.

Consider two systems.

System A

AI predicts demand of 90 GW.

The operator sees the prediction and must accept it unless a supervisor approves an exception.

System B

AI predicts demand of 90 GW.

The operator sees:

forecast;

confidence interval;

historical error;

weather anomalies;

alternative model prediction;

relevant system conditions.

The operator can reject the forecast immediately and the system records the reason.

System B provides substantially more meaningful human oversight.

The UK's Data and AI Ethics Framework similarly stresses identifying who is responsible, defining how much of the process is automated and ensuring that humans can intervene in risky or high-impact systems. (GOV.UK)

15. Human Override and Liability

An important legal question is:

Who is liable when an AI forecast is wrong?

Possible actors include:

AI developer;

software vendor;

electricity generator;

distribution company;

transmission operator;

system operator;

forecasting service provider;

regulator;

individual decision-maker.

A human-override framework should therefore allocate responsibility contractually and institutionally.

For example:

Developer:
Responsible for model design, testing and disclosed limitations.

Utility:
Responsible for appropriate deployment and monitoring.

System operator:
Responsible for operational decisions within its statutory authority.

Human supervisor:
Responsible for exercising assigned oversight appropriately.

Regulator:
Responsible for regulatory supervision and compliance requirements.

Human override should not become a mechanism for transferring all AI-related liability to an individual operator.

16. Human Override and Automation Bias

There is a paradox:

The more accurate an AI system becomes, the more likely humans may be tempted to accept it automatically.

Therefore, a successful override framework should not merely teach operators how to override. It should teach them when they should consider overriding.

Possible trigger conditions include:

confidence below a prescribed threshold;

unusual weather;

sensor disagreement;

sudden demand movement;

model drift;

data missingness;

cyber-security alerts;

disagreement between forecasting models;

unprecedented system conditions.

This creates a risk-triggered override architecture.

17. Ofgem and Energy-Specific Governance

Ofgem's updated 2026 guidance is particularly significant for energy law because it directly addresses ethical AI use in the energy sector.

It covers AI used in:

predictions and forecasting;

grid management;

consumer interactions;

data analytics.

The guidance has also considered explainability, transparency, black-box systems and AI assurance. (Ofgem)

Ofgem is also proceeding with a 12-month AI technical sandbox pilot, intended to allow controlled testing of AI systems in the energy sector and generate evidence concerning system behaviour and regulatory risks. (Ofgem)

This demonstrates a regulatory movement from simply asking:

"Does AI work?"

toward:

"Can AI be safely governed, audited and controlled?"

18. Legal Doctrine Emerging from the Case Law

The cases discussed above collectively suggest several principles.

Principle 1 — AI is evidence, not necessarily authority

Loomis illustrates how algorithmic assessments can assist decision-makers without becoming determinative. (Justia Law)

Principle 2 — Human involvement must be meaningful

SCHUFA demonstrates that formal human involvement does not necessarily remove the legal significance of automated decision-making. (Infocuria)

Principle 3 — Algorithmic systems require safeguards

SyRI demonstrates that automated governmental systems affecting individuals require legally adequate safeguards. (Rechtspraak)

Principle 4 — AI output requires verification

The Supreme Court of India's 2026 decision concerning AI-generated hallucinated authorities reinforces the importance of human verification when AI output enters a consequential legal process. (Scientific Ministries)

Principle 5 — Risk determines the intensity of oversight

The EU AI Act expressly adopts a risk- and context-sensitive approach to human oversight. (EUR-Lex)

19. Proposed Legal Test for AI Forecasts

A regulator or court could conceptually ask five questions:

Test 1 — Consequence

What happens if the forecast is wrong?

Test 2 — Autonomy

How much operational authority has been delegated to the AI?

Test 3 — Contestability

Can an authorised human reject the forecast?

Test 4 — Capability

Does that human have sufficient information, expertise and time to challenge it?

Test 5 — Accountability

Is responsibility clearly allocated for the resulting decision?

If the answer to the third or fourth question is effectively "no", the system may have formal human oversight without meaningful human control.

20. Conclusion

Human override requirements in AI forecasting represent a transition from "human-in-the-loop" as a slogan to human control as a legally enforceable governance mechanism.

The central principle is not that every AI forecast must be manually checked. That would often be impractical, especially in high-frequency electricity systems. Rather, the level of human oversight should correspond to the risk, autonomy and consequences of the AI application.

The EU AI Act provides the clearest statutory formulation by requiring appropriately designed human oversight and, where appropriate, the ability to disregard, override or reverse AI output. (EUR-Lex)

State v. Loomis demonstrates the importance of retaining independent human judgment when algorithmic assessments influence consequential decisions. (Justia Law) SCHUFA demonstrates why merely inserting a human into an automated process may not be enough. (Infocuria) SyRI demonstrates the importance of safeguards around algorithmic governmental systems. (Rechtspraak) Indian constitutional principles of legality and proportionality, together with the Supreme Court's recent treatment of AI-generated material, provide additional foundations for requiring verification and accountable human reasoning in consequential AI-assisted processes. (Sci API)

For electricity forecasting, the strongest legal model is therefore:

AI predicts → human understands → human verifies → human may override → action is recorded → outcome is audited.

That architecture preserves the efficiency of AI while ensuring that responsibility for critical energy decisions remains with accountable human institutions.