Regulatory Models Exceeding Interpretability Thresholds

1. Introduction

Regulatory models exceeding interpretability thresholds refers to situations in which a regulatory framework, algorithm, risk model, tariff model, artificial-intelligence system, or other decision-making mechanism becomes so complex that affected persons, regulators, courts, and even the institutions operating the system can no longer adequately understand how inputs are transformed into regulatory outcomes.

Interpretability is important because regulation is not merely about producing a technically correct result. A regulatory decision must ordinarily be capable of being explained, justified, reviewed, challenged, and held accountable.

The problem becomes particularly significant in energy governance. Electricity markets increasingly depend upon algorithmic dispatch, automated balancing, dynamic tariffs, forecasting systems, network-congestion models, renewable-generation forecasts, credit-risk models, and AI-assisted regulatory supervision. When these systems become opaque, the legal question is no longer simply “Is the model accurate?” but also:

“Can the exercise of regulatory power be sufficiently understood and justified to satisfy legality, procedural fairness, transparency, and judicial review?”

This creates what may be called an interpretability threshold: beyond a certain level of complexity, the regulatory value of a model may be undermined because its decisions cannot be meaningfully explained.

2. Meaning of the Interpretability Threshold

An interpretability threshold is the point at which the complexity of a regulatory model exceeds the ability of relevant actors to understand its:

1. inputs;
2. methodology;
3. assumptions;
4. decision rules;
5. weighting of competing factors;
6. outputs; and
7. relationship between the model and the legal authority under which it operates.

The threshold is not necessarily a mathematical number. It is primarily a legal and institutional concept.

For example, suppose an electricity regulator uses an AI model to determine whether a distribution company is entitled to recover ₹500 crore in network expenditure.

If the regulator can explain:

• what data were used;
• what statutory criteria were applied;
• what assumptions were adopted;
• why particular costs were accepted;
• why others were rejected; and
• how the final amount follows from the regulatory methodology,

the model remains legally manageable.

But if the regulator merely states:

“The AI model determined that ₹500 crore represents the efficient expenditure.”

without being able to explain the model's reasoning, the regulatory process risks becoming functionally unreviewable.

3. Why Interpretability Matters in Regulation

Regulatory power differs from ordinary private decision-making because it can directly affect:

• property;
• prices;
• licences;
• market access;
• electricity supply;
• investment;
• environmental obligations;
• consumer rights;
• competition;
• public expenditure.

Consequently, regulatory models must generally operate within principles such as:

A. Legality

The decision must have a lawful statutory foundation.

B. Procedural fairness

Affected parties should have a meaningful opportunity to understand and contest material considerations.

C. Reasoned decision-making

Regulators should ordinarily provide sufficient reasons for significant decisions.

D. Transparency

The methodology should be sufficiently disclosed to permit meaningful scrutiny.

E. Accountability

Decision-makers must remain responsible for decisions rather than transferring responsibility to an opaque model.

F. Judicial review

Courts must retain the practical ability to examine whether the regulator acted within its powers.

4. Regulatory Complexity and the Problem of the Black Box

The problem can be illustrated through a simplified regulatory chain:

Data → Algorithm → Model → Prediction → Regulatory Decision → Legal Consequence

If each stage is comprehensible, accountability is relatively straightforward.

But increasingly sophisticated systems may look like:

Massive datasets → machine learning → multiple interacting variables → adaptive optimisation → probabilistic output → automated recommendation → regulatory decision

The difficulty is that the relationship between input and output may no longer be intuitively understandable.

This creates the black-box regulatory problem.

A regulator may know that the model is statistically effective without knowing why a particular company, consumer group, project, or transaction received a particular regulatory classification.

5. Interpretability Is Different From Accuracy

One of the most important distinctions is between accuracy and interpretability.

A model may be highly accurate but poorly interpretable.

For example:

Regulatory law cannot necessarily accept the proposition:

“The model is accurate, therefore its decision is legally acceptable.”

Accuracy answers the question whether the model predicts well.

Interpretability addresses whether the exercise of regulatory power can be understood and justified.

These are different legal values.

6. Interpretability as a Component of Reasoned Decision-Making

A regulatory authority generally cannot avoid responsibility by saying that an external technical model produced the result.

The legal responsibility remains with the decision-maker.

This principle is particularly important where:

• tariffs are determined algorithmically;
• licences are granted or denied using risk scores;
• network investments are prioritised by algorithms;
• environmental compliance is assessed automatically;
• electricity consumers are classified according to predictive models;
• market manipulation is identified through AI;
• wholesale electricity prices are influenced by automated systems.

The regulator therefore needs to establish an explainability bridge between the model's output and the legal decision.

7. Key Case Law

7.1 R v Secretary of State for the Home Department, ex parte Doody [1994]

The House of Lords established an important principle concerning procedural fairness and reasons.

The case concerned prisoners and the exercise of administrative discretion. The House of Lords recognised the importance of providing sufficient information and reasons where fairness requires it.

Relevance

The principle is highly relevant to algorithmic regulation.

Where an automated regulatory system produces an adverse decision, procedural fairness may require sufficient explanation of the basis upon which the decision was reached.

The principle can therefore be translated into modern regulatory settings:

The greater the consequences of an automated decision, the stronger the justification for meaningful explanation.

8. R (Miller) v Secretary of State for Exiting the European Union [2017] UKSC 5

The UK Supreme Court emphasised constitutional principles concerning the exercise of executive power.

Although Miller was not an AI case, it demonstrates a broader proposition:

Public power must remain connected to lawful constitutional authority.

This is relevant where regulators increasingly rely upon sophisticated technical systems.

An algorithm cannot itself become the source of legal authority.

The legal authority must remain grounded in:

• legislation;
• delegated legislation;
• regulatory rules;
• licences; or
• other recognised sources of public law.

Thus:

Complexity cannot create jurisdiction.

9. R (Privacy International) v Investigatory Powers Tribunal [2019] UKSC 22

This case concerned judicial review and the limits of attempts to exclude judicial supervision.

Its broader constitutional importance lies in the protection of judicial review of executive action.

Relevance to algorithmic regulation

If a regulator says:

“The model is too technically complex for the court to review,”

that cannot ordinarily be treated as a complete answer.

Technical complexity may make judicial review more difficult, but it should not automatically make public power immune from legal scrutiny.

The interpretability threshold therefore interacts directly with the principle of reviewability.

10. R (Brent London Borough Council) v Secretary of State for Communities and Local Government [2010] EWCA Civ 562

The case is relevant to the broader administrative-law requirement that decision-makers properly understand and apply the relevant legal framework.

An algorithmic system cannot substitute for the legal judgment that legislation requires.

A model may assist a regulator, but it cannot silently determine the meaning of statutory criteria.

11. State of Orissa v Dr (Miss) Binapani Dei AIR 1967 SC 1269

The Supreme Court of India established an important principle that administrative decisions having civil consequences must comply with principles of natural justice.

The case is particularly significant because it demonstrates that administrative action cannot be treated as legally insignificant merely because it is classified as administrative.

Relevance to AI-based regulation

If an algorithmic model produces a decision having serious civil consequences, the use of technology does not eliminate natural justice.

For example:

• denial of an electricity licence;
• adverse tariff treatment;
• cancellation of regulatory approval;
• exclusion from an energy market;
• penalties for alleged market manipulation.

The affected party may still require procedural fairness.

12. Maneka Gandhi v Union of India (1978) 1 SCC 248

This is one of India's most important administrative-law decisions.

The Supreme Court significantly expanded the understanding of fairness, reasonableness and procedure under Article 21 of the Constitution.

The decision rejected the idea that merely following a formally prescribed procedure automatically satisfies constitutional requirements.

Relevance

An algorithmic decision-making system may formally follow a prescribed procedure while nevertheless being practically opaque.

Therefore:

Formal automation does not necessarily equal substantive fairness.

A regulatory model that cannot explain a severe adverse decision may raise questions concerning whether the procedure was genuinely fair.

13. Mohinder Singh Gill v Chief Election Commissioner (1978) 1 SCC 405

The Supreme Court stressed the importance of reasons underlying administrative decisions.

The case is famous for the proposition that an order must stand on the reasons contained in the decision itself and cannot ordinarily be supplemented later by an entirely different justification.

Application to regulatory algorithms

Suppose a regulator issues:

“The model identifies the applicant as high risk.”

If the regulator later attempts to justify the decision through entirely different reasons, the process may become problematic.

The regulatory institution should therefore document:

• the relevant model;
• the factors considered;
• the applicable legal criteria;
• the material reasoning;
• the final decision.

14. Kranti Associates Pvt. Ltd. v Masood Ahmed Khan (2010) 9 SCC 496

The Supreme Court strongly emphasised the importance of recording reasons in administrative and quasi-judicial decisions.

The judgment is especially relevant to the interpretability problem.

A reasoned decision:

1. demonstrates application of mind;
2. reduces arbitrariness;
3. facilitates judicial review;
4. informs the affected party;
5. strengthens public confidence.

AI relevance

If an algorithm supplies an unexplained output, the regulator should not simply reproduce that output.

The authority must translate the technical result into legally intelligible reasons.

This can be described as the human-readable justification requirement.

15. Tata Cellular v Union of India (1994) 6 SCC 651

The Supreme Court established important principles concerning judicial review of administrative decisions, particularly in government contracts.

The Court emphasised that judicial review concerns the decision-making process, rather than simply substituting the court's decision for that of the administrative authority.

Relevance

The case is highly relevant to complex regulatory models.

A court does not necessarily need to reconstruct every mathematical operation inside an AI system.

Instead, it may ask:

• Was the correct legal power used?
• Was the correct methodology adopted?
• Were relevant factors considered?
• Were irrelevant factors excluded?
• Was the process arbitrary?
• Was there procedural unfairness?
• Was the decision supported by reasons?

Thus, interpretability is necessary to make process review meaningful.

16. Reliance Infrastructure Ltd. v Maharashtra Electricity Regulatory Commission

Indian electricity regulation provides a particularly important setting for this problem.

Electricity regulators such as CERC and SERCs routinely make technical and economic determinations concerning:

• tariffs;
• power procurement;
• transmission;
• distribution;
• capital expenditure;
• fuel costs;
• power purchase agreements;
• regulatory assets.

Judicial review and appellate scrutiny of electricity-regulatory decisions demonstrate the importance of transparent regulatory methodology.

The Appellate Tribunal for Electricity (APTEL) has repeatedly scrutinised whether electricity regulators have properly applied statutory and regulatory principles when determining tariffs and other regulatory questions.

The lesson is significant:

Technical expertise does not eliminate the requirement of legally intelligible reasoning.

17. Energy Watchdog v CERC (2017) 14 SCC 80

This Supreme Court decision concerned regulatory treatment of power-purchase agreements and changes in circumstances affecting electricity generation.

The Court interpreted contractual and regulatory principles governing electricity generation.

Relevance

Energy regulation frequently involves highly technical economic calculations. Nevertheless, the legal outcome must remain connected to:

• statutory authority;
• contractual obligations;
• regulatory principles; and
• legally recognised doctrines.

An increasingly sophisticated model cannot displace these legal foundations.

18. Gujarat Urja Vikas Nigam Ltd. v Essar Power Ltd.

The Supreme Court has repeatedly recognised the specialised role of electricity regulatory commissions.

This jurisprudence demonstrates a delicate balance:

Judicial deference to regulatory expertise ≠ judicial surrender to regulatory opacity.

Courts generally respect specialised regulators, but regulatory expertise does not create an unlimited zone beyond legal scrutiny.

This distinction becomes crucial as regulatory agencies employ AI and quantitative models.

19. The EU GDPR and SCHUFA Jurisprudence

European data-protection law provides an important comparative dimension.

The GDPR contains protections concerning automated decision-making and requires transparency concerning processing.

The Court of Justice of the European Union's jurisprudence concerning automated decision-making, including the SCHUFA litigation, has strengthened attention to the legal significance of automated scoring.

The broader lesson is that algorithmic decisions cannot always be treated as legally neutral simply because they arise from statistical processing.

For regulators, this supports a broader principle:

Where an algorithm materially determines legal or similarly significant outcomes, meaningful information about the decision-making logic becomes legally important.

20. The SyRI Case — Netherlands

The Dutch court's decision concerning the SyRI welfare-fraud risk system is particularly important for algorithmic governance.

The court found the system incompatible with human-rights requirements because its operation lacked sufficient transparency and safeguards against abuse.

The case illustrates a fundamental problem:

A government cannot simply invoke technological complexity to justify opaque interference with individuals.

Although SyRI concerned social-security fraud detection rather than energy regulation, the principle is highly transferable to automated regulatory systems.

21. The Bridges Case — Facial Recognition

In R (Bridges) v Chief Constable of South Wales Police [2020] EWCA Civ 1058, the UK Court of Appeal considered automated facial-recognition technology.

The case demonstrates the importance of:

• legal authority;
• safeguards;
• proportionality;
• transparency; and
• adequate legal frameworks.

The case is especially useful for understanding algorithmic governance because the technology was not rejected merely because it was innovative.

Instead, the court examined whether the legal framework sufficiently controlled its use.

22. Regulatory Models in Electricity Markets

The interpretability problem becomes especially serious in electricity markets.

Consider an AI-based electricity pricing model.

It could process:

• demand;
• weather;
• generation availability;
• transmission congestion;
• fuel prices;
• renewable intermittency;
• storage levels;
• interconnector flows;
• consumer behaviour;
• market bids.

The system could then recommend a wholesale electricity price.

The model might be economically sophisticated but difficult to explain.

This creates several legal questions:

1. Who is responsible?

The regulator?

The market operator?

The software provider?

The electricity exchange?

2. What is the legal basis?

Which statute or regulation authorises the model?

3. Can participants challenge the result?

If market participants cannot understand the model, effective challenge becomes difficult.

4. Can a court review it?

If the model is incomprehensible even to the regulator, judicial review may become practically weakened.

23. Regulatory Models and Tariff Determination

Tariff regulation provides another example.

A regulator could use a machine-learning model to determine the efficient expenditure of a distribution utility.

Inputs could include:

• asset age;
• outage frequency;
• customer density;
• network topology;
• weather;
• historical expenditure;
• operational efficiency;
• regional benchmarks.

The model may calculate an allowable revenue requirement.

But if the utility cannot determine why its expenditure was rejected, the regulatory process may lack meaningful contestability.

Therefore, the regulator should disclose at least:

1. the legal methodology;
2. material variables;
3. major assumptions;
4. treatment of exceptional circumstances;
5. principal reasons for the result;
6. validation mechanisms;
7. human oversight.

24. Regulatory Models and Natural Justice

Interpretability directly affects audi alteram partem—the right to be heard.

A hearing becomes meaningless if the affected party does not know what it is responding to.

Suppose an energy company is denied approval because:

“The predictive model assigned a 92% probability of regulatory non-compliance.”

The company should reasonably be able to ask:

• Which factors generated the score?
• Was the data accurate?
• Were outdated data used?
• Was the company compared against inappropriate entities?
• Were exceptional circumstances ignored?
• Was the model validated?
• Was human judgment involved?

Without meaningful answers, the right to be heard may become largely symbolic.

25. The Problem of Model Drift

Interpretability becomes even more difficult when models evolve.

An adaptive regulatory system may change its internal parameters as new data arrive.

Consequently:

Model A at time T1 ≠ Model A at time T2

This creates a regulatory-record problem.

A person may challenge a decision made in 2026, but the model may have been retrained by 2027.

Which model should the regulator produce for judicial review?

This makes model preservation increasingly important.

Regulators should maintain:

• model versions;
• training datasets where legally appropriate;
• parameter histories;
• decision logs;
• audit trails;
• human overrides;
• validation reports.

26. Interpretability and Regulatory Accountability

A major danger is responsibility displacement.

This occurs when human officials say:

“The algorithm made the decision.”

Legally and institutionally, this is problematic.

An algorithm cannot ordinarily bear public-law responsibility.

Responsibility should remain with the institution legally empowered to make the decision.

Therefore:

Algorithmic assistance

is different from

Algorithmic delegation of legal authority.

The first can be legitimate.

The second requires much stronger legal justification.

27. The Interpretability Cascade

When one opaque model is incorporated into another regulatory model, opacity can multiply.

For example:

AI demand forecast → AI generation forecast → AI congestion model → AI pricing model → regulatory tariff decision

Each layer depends on the previous layer.

This creates an interpretability cascade.

An error or unexplained assumption at an early stage may propagate through the entire regulatory system.

The final decision may appear mathematically precise while its underlying assumptions remain uncertain.

28. Regulatory Complexity and Delegation

There is also a constitutional concern concerning sub-delegation.

A legislature may delegate regulatory authority to an expert agency.

The agency may then use a technical model.

But if the agency effectively allows an external software system to determine substantive regulatory outcomes without adequate legal control, the question arises:

Has regulatory discretion effectively been delegated again to a private or automated system?

This is particularly problematic when the model is supplied by:

• private technology companies;
• consultants;
• foreign vendors;
• proprietary software providers.

29. Proprietary Algorithms and Regulatory Secrecy

A particularly difficult issue arises where the model is protected as a trade secret.

A technology provider may argue:

“The regulator cannot disclose the model because it is proprietary.”

But the affected party may respond:

“Without disclosure, I cannot meaningfully challenge the decision.”

This creates tension between:

• intellectual-property protection;
• commercial confidentiality;
• regulatory transparency;
• procedural fairness;
• judicial review.

The appropriate legal solution may not always require disclosure of the entire source code.

Instead, regulators may require:

• independent auditing;
• explanation of material variables;
• documentation of methodology;
• access under confidentiality arrangements;
• regulator-controlled testing;
• reproducibility mechanisms.

30. A Proportionality-Based Interpretability Requirement

Not every regulatory decision requires the same level of explanation.

A useful approach is proportionality.

Low-impact decision

Minimal explanation may be sufficient.

Medium-impact decision

Greater disclosure of methodology may be necessary.

High-impact decision

Detailed explanation, human review, auditability and challenge mechanisms should generally become stronger.

Fundamental-rights decision

The strongest safeguards may be required.

Thus:

Interpretability requirements should increase with regulatory impact.

31. Proposed Legal Test

A useful legal framework for courts and regulators could ask six questions.

Test 1 — Authority

Does the regulator possess lawful authority to use the model?

Test 2 — Relevance

Are the variables used by the model legally relevant?

Test 3 — Transparency

Can the affected party understand the principal basis of the decision?

Test 4 — Explainability

Can the regulator provide meaningful reasons connecting the model's output to the legal decision?

Test 5 — Reviewability

Can an independent body meaningfully review the decision?

Test 6 — Accountability

Does a human institution remain legally responsible?

Failure across these dimensions may indicate that the model has exceeded an acceptable interpretability threshold.

32. Energy-Law Implications

For energy regulators, interpretability is particularly important because electricity systems involve essential public infrastructure.

An opaque regulatory model could affect:

• electricity tariffs;
• grid access;
• renewable-energy procurement;
• transmission planning;
• generation dispatch;
• capacity payments;
• consumer protection;
• reliability standards;
• market manipulation investigations;
• carbon regulation;
• energy-storage participation.

Errors can therefore produce consequences not merely for individual companies but for entire electricity systems.

33. Regulatory Design Principles

Regulators should adopt an interpretability-by-design approach.

This should include:

1. Model documentation

Every material regulatory model should have documentation explaining its purpose and methodology.

2. Version control

Every decision should be linked to the precise model version used.

3. Auditability

Independent experts should be able to test the system.

4. Human oversight

Important decisions should retain accountable human decision-makers.

5. Reasoned outputs

The system should generate explanations rather than only numerical results.

6. Challenge procedures

Affected parties should have mechanisms to challenge model outputs.

7. Bias testing

Models should be tested for systematic distortions.

8. Periodic review

Models should be reassessed as markets and technologies change.

34. A Conceptual Formula

The legal problem can be expressed conceptually as:

Regulatory Legitimacy = Legal Authority + Procedural Fairness + Explainability + Accountability + Reviewability

When model complexity increases:

Complexity ↑ → Explainability ↓ → Reviewability ↓ → Accountability risk ↑

This does not mean that complex models are inherently unlawful.

Rather:

The greater the complexity, the greater the institutional safeguards required to preserve legality and accountability.

35. Relationship With Judicial Deference

Courts often defer to expert regulators on technical matters.

This is sensible because judges may not possess specialised expertise in:

• electricity economics;
• engineering;
• network optimisation;
• financial modelling;
• AI.

But deference should not become blind acceptance.

A court can defer to technical expertise while still requiring:

• lawful authority;
• rational methodology;
• adequate reasons;
• procedural fairness;
• non-arbitrariness.

The distinction is therefore:

Technical deference ≠ legal deference without limits.

36. The Emerging Principle of Explainable Regulation

A new regulatory principle can be conceptualised as Explainable Regulation.

Under this principle, where an automated or highly complex model materially influences a regulatory decision, the regulator should be capable of explaining:

1. what the model was designed to determine;
2. what information it considered;
3. what principal factors influenced the outcome;
4. what limitations it possesses;
5. how human officials evaluated the result;
6. how the result connects to the governing law.

This does not necessarily require complete technical transparency.

The objective is legally meaningful transparency.

37. Critical Evaluation

The concept of regulatory models exceeding interpretability thresholds exposes a fundamental tension in modern governance.

On one side:

Complexity enables better regulation.

Advanced models can improve:

• forecasting;
• grid management;
• fraud detection;
• market surveillance;
• tariff modelling;
• resource allocation.

On the other side:

Complexity can weaken democratic accountability.

If no affected party can understand the decision, regulation may become technologically sophisticated but legally fragile.

The challenge is therefore not to eliminate complex models but to build institutions capable of governing them.

38. Conclusion

Regulatory models exceeding interpretability thresholds represent a major emerging problem in administrative, constitutional, and energy law.

The central legal principle is straightforward:

A regulator cannot avoid legal accountability merely because the mechanism used to make a decision is technologically complex.

Cases such as Maneka Gandhi, Mohinder Singh Gill, Kranti Associates, Tata Cellular, Doody, Privacy International, Bridges, and the SyRI litigation collectively support broader principles of legality, fairness, reasons, proportionality, transparency, and judicial review that are increasingly relevant to algorithmic regulation.

For energy governance, the issue is particularly significant because AI and quantitative models are becoming embedded in tariff determination, market surveillance, grid management, procurement, forecasting and system planning.

The appropriate response is not necessarily to prohibit sophisticated models. Instead, regulators should establish an interpretability threshold linked to regulatory impact: the more significant the legal consequences of a model-driven decision, the stronger the requirements for explanation, auditability, human oversight, procedural fairness and judicial review.

Ultimately, regulatory intelligence must not become regulatory opacity. A model may assist the regulator, but it should not make the exercise of public power unintelligible.