Delegation Limits for Administrative Discretion to Machine Systems

1. Introduction

The increasing use of artificial intelligence, machine-learning models, automated scoring systems, predictive analytics and algorithmic decision-support tools creates a modern version of a traditional administrative-law problem: to what extent may a public authority transfer, surrender, or effectively outsource a statutory discretion to a machine?

Traditional administrative law assumes that a statutory power is entrusted to a legally identified decision-maker—such as a minister, regulator, tribunal or local authority. That decision-maker may often use officials, experts, software and administrative machinery, but the legal responsibility for exercising the discretion remains with the authority to which Parliament or the legislature entrusted it.

The problem becomes more difficult where the machine does more than calculate objective information and instead determines:

• eligibility;
• risk classifications;
• enforcement priorities;
• benefit entitlement;
• licensing outcomes;
• immigration or policing risk;
• tax investigations;
• regulatory sanctions;
• allocation of scarce public resources; or
• the factual or policy assumptions upon which the final decision depends.

Recent scholarship specifically identifies this as a challenge to traditional non-delegation doctrine because a nominally human "final decision" may in practice merely rubber-stamp an algorithmic output.

The central principle can therefore be stated as:

A public authority may ordinarily use a machine as an instrument for exercising statutory discretion, but it cannot use automation to evade the legal limits, procedural duties, accountability requirements or substantive judgment imposed upon the legally authorised decision-maker.

2. Meaning of Delegation to Machine Systems

There are at least four different situations.

A. Mechanical assistance

The machine performs calculations based on legally fixed criteria.

Example: calculating a statutory filing deadline or determining whether a fee has been paid.

This normally presents the least difficulty because the machine has little or no independent discretion.

B. Decision support

The machine produces a recommendation, but an authorised official genuinely considers the recommendation together with other relevant evidence.

Example:

Algorithm → identifies potentially fraudulent applications → officer investigates → officer independently decides.

This is generally easier to reconcile with traditional administrative law.

C. Algorithmic determination

The machine determines the outcome subject only to nominal human approval.

Example:

Algorithm → classifies applicant as "high risk" → official automatically accepts classification → licence refused.

Here the machine may have acquired de facto decision-making authority.

D. Fully automated decision-making

The statutory outcome is generated automatically without meaningful human consideration.

This raises the strongest delegation, legality, procedural fairness and accountability concerns.

3. The Core Legal Principle: Delegatus Non Potest Delegare

The traditional maxim means, broadly, that a person to whom a power has been delegated cannot further delegate it unless authorised to do so.

Indian Supreme Court authority has expressly described the principle as requiring a statutory power to be exercised by the body or officer to whom it was entrusted unless sub-delegation is authorised by express words or necessary implication. It also emphasises a strong presumption against sub-delegating legislative, judicial or disciplinary powers.

Applied to machines, the question becomes:

Has the legislature authorised the public authority to make the machine part of the decision-making architecture, or has the authority effectively transferred its own statutory judgment to an algorithm?

The distinction is important because software itself is not normally a legal office-holder.

4. Six Major Legal Tests

A machine-assisted administrative decision should therefore be examined through at least six questions.

1. Source of authority

Where does the power originate?

If Parliament gives a minister a discretionary power, the minister cannot simply transfer the substantive discretion to an AI vendor unless the statutory framework permits that arrangement.

2. Nature of the discretion

Is the function:

• ministerial;
• administrative;
• investigative;
• technical;
• quasi-judicial;
• disciplinary; or
• adjudicatory?

The more legally significant and judgment-dependent the function, the stronger the argument for retaining human responsibility.

3. Degree of machine autonomy

Does the system:

• calculate;
• recommend;
• rank;
• classify;
• decide; or
• automatically implement the decision?

The further the system moves from calculation towards substantive determination, the greater the delegation problem.

4. Human independence

A nominal human signature is insufficient if the official is effectively compelled by the machine's output.

The relevant question is:

Did the authorised decision-maker genuinely exercise judgment?

5. Statutory purpose

The machine must operate consistently with the purpose for which Parliament granted the discretion.

This follows from the principle in Padfield: statutory discretion must be exercised to promote the policy and objects of the legislation.

6. Reviewability

Courts must be able to determine:

• what the machine did;
• what data it used;
• what criteria it applied;
• what legal authority authorised its use;
• who ultimately made the decision;
• whether relevant matters were considered; and
• whether irrelevant or impermissible considerations influenced the result.

5. Case Law

Case 1 — Carltona Ltd v Commissioners of Works [1943] 2 All ER 560

Principle

This is the classic authority permitting substantial administrative delegation within government departments.

During wartime administration, functions formally vested in a minister were exercised by responsible departmental officials. The court recognised the practical impossibility of requiring ministers personally to exercise every administrative function. Public administration necessarily operates through officials acting under ministerial authority.

Relevance to machine systems

Carltona provides an important analogy for algorithmic administration.

A minister does not personally need to:

• calculate every figure;
• examine every document;
• process every application; or
• perform every routine administrative operation.

A computer system can similarly be treated as administrative machinery.

But Carltona should not be interpreted as allowing unrestricted transfer of substantive statutory discretion to technology.

The important distinction is:

Administrative assistance ≠ transfer of legal responsibility.

Case 2 — Barnard v National Dock Labour Board [1953] 2 QB 18

This is one of the strongest authorities for the limits of delegation.

The Local Dock Labour Board possessed statutory disciplinary powers over dock workers. The Board attempted to delegate those powers to its port manager.

The Court of Appeal held that the disciplinary power could not lawfully be delegated in that manner because it was essentially judicial or quasi-judicial in character. The fact that the Board had a large workload did not itself authorise delegation.

The court distinguished ordinary administrative functions from substantive adjudicative powers.

Application to AI

Suppose legislation gives a regulator the power to impose a regulatory penalty after considering:

• evidence;
• explanations;
• mitigating circumstances;
• proportionality; and
• culpability.

The regulator cannot necessarily say:

"The AI determined the penalty, therefore the statutory discretion has been exercised."

Barnard suggests that where the legislature entrusts a substantive judgment to a particular body, the body must actually exercise that judgment unless delegation is legally authorised.

Thus:

AI recommendation: potentially permissible.

AI adjudication replacing the statutory decision-maker: much more problematic.

Case 3 — H. Lavender & Son Ltd v Minister of Housing and Local Government [1970] 3 All ER 871

This case concerns fettering of discretion and acting under the dictation of another authority.

The Minister responsible for deciding a planning appeal effectively treated the position of another minister as determinative. The court regarded this as an improper surrender of the discretion which the legislation had entrusted to the relevant minister.

Application to algorithms

This principle is particularly important for machine systems.

Suppose legislation says:

"The licensing authority may grant or refuse a licence after considering the relevant circumstances."

The authority creates an AI model which produces:

"Risk score = 94; therefore refuse."

If the authority treats the score as determinative and does not genuinely consider the individual case, the problem is not merely technological accuracy.

It may be a failure to exercise the statutory discretion.

The machine has effectively become the decision-maker.

Case 4 — R v Minister of Agriculture, Fisheries and Food, ex parte Padfield [1968] UKHL 1

Padfield establishes a fundamental proposition concerning statutory discretion.

A statutory discretion cannot be treated as completely unlimited. It must be exercised consistently with the policy and objects of the statute.

The House of Lords rejected the proposition that describing a discretion as "unfettered" eliminates judicial control.

Application to machine systems

The same principle applies when discretion is embedded in software.

Suppose a statute aims to:

• protect vulnerable persons;
• promote fair access;
• prevent discrimination; or
• balance competing public interests.

An algorithm cannot lawfully convert the discretion into a rigid mathematical rule that frustrates that legislative purpose.

For example:

Statutory objective: individualised assessment
Algorithmic rule: automatic rejection whenever risk score > 80

If the statutory scheme requires consideration of individual circumstances, the algorithm may unlawfully transform discretion into an inflexible rule.

Case 5 — R (Bridges) v Chief Constable of South Wales Police [2020] EWCA Civ 1058

This is one of the leading modern cases concerning automated facial-recognition technology in public administration.

South Wales Police used Live Automated Facial Recognition technology to compare individuals captured by cameras with images on police watchlists.

The Court of Appeal found the deployment unlawful on several grounds, including that the legal framework did not sufficiently determine where AFR could be deployed and that the police had failed to comply with the Public Sector Equality Duty.

Importance

Bridges demonstrates that a public authority cannot avoid ordinary public-law obligations merely because its decision-making is technologically mediated.

The authority remained responsible for:

• identifying the legal boundaries of deployment;
• establishing adequate safeguards;
• considering equality implications;
• ensuring that discretion was constrained by law.

Machine-delegation lesson

Technology does not become legally autonomous simply because its operation is automated.

The public authority remains accountable for the architecture within which the machine operates.

Case 6 — State v Loomis, 881 N.W.2d 749 (Wis. 2016)

The Wisconsin Supreme Court considered the use of the COMPAS algorithmic risk-assessment system in criminal sentencing.

The court permitted consideration of the algorithmic assessment while imposing important limitations and warnings concerning its use. The official Wisconsin case record identifies the decision as State v Eric L. Loomis, 881 N.W.2d 749 (2016).

The case is important because the algorithm did not formally become the sentencing authority.

Principle for administrative law

The case illustrates a possible distinction between:

algorithmic assistance

and

algorithmic determination.

An algorithm may provide information relevant to a legally authorised human decision-maker without itself becoming the legal decision-maker.

But the safeguards become critical where:

• the methodology is opaque;
• the output has significant consequences;
• the official cannot independently evaluate it; or
• the output becomes practically determinative.

Case 7 — Federation of Dutch Trade Unions (FNV) v State of the Netherlands, ECLI:NL:RBDHA:2020:1878 — SyRI

The District Court of The Hague considered the Dutch government's System Risk Indication (SyRI), an algorithmic system used to identify possible fraud involving social-security and other public schemes.

The court held that the statutory framework permitting the system failed to provide sufficient safeguards and transparency and violated Article 8 ECHR requirements concerning private life. The case is widely recognised as a landmark judicial examination of algorithmic government.

Significance

SyRI demonstrates that the legality of an algorithm may depend not merely on the final decision but also upon:

• data inputs;
• risk indicators;
• system design;
• transparency;
• safeguards;
• proportionality; and
• the legal basis for processing information.

This is particularly important for delegation analysis because the substantive discretion may be hidden inside the design of the system itself.

Case 8 — OQ v Land Hessen / SCHUFA Holding (Scoring), C-634/21

The Court of Justice of the European Union considered automated credit scoring under Article 22 GDPR.

The Court held that automated establishment of a probability value by a credit-information agency constitutes automated individual decision-making where a third party strongly relies on that score to establish, implement or terminate a contractual relationship.

Importance for administrative discretion

The case illustrates an important legal principle:

A system cannot escape regulation merely because its output is formally described as a "recommendation" or "score."

If the downstream decision-maker strongly relies upon the automated output, the system may have substantive decision-making significance.

That reasoning is highly relevant to public administration.

An agency cannot necessarily avoid delegation controls simply by saying:

"The AI does not make the final decision; the officer does."

If the officer is institutionally or practically bound by the AI's classification, the distinction between recommendation and decision may become artificial.

6. Comparison of the Cases

7. The "Human-in-the-Loop" Problem

One of the most important issues is the difference between formal human control and substantive human control.

Consider:

AI generates score → officer receives score → officer approves result in 98% of cases.

Formally, the officer makes the decision.

Substantively, however, the algorithm may be exercising the decisive influence.

This creates what can be called rubber-stamp delegation.

Modern scholarship has specifically identified this problem: traditional delegation doctrine often focuses on who formally makes the final decision, even though an algorithm may have already determined the substantive parameters of that decision.

8. Algorithmic Fettering of Discretion

An algorithm can unlawfully fetter discretion in several ways.

Fixed thresholds

Score below 50 = approve
Score above 50 = reject

Automatic exclusions

Any applicant satisfying X is automatically rejected.

Mandatory algorithmic recommendations

Officer must follow the AI recommendation unless exceptional circumstances exist.

Institutional automation bias

Officials may be technically permitted to disagree but practically discouraged from doing so.

Resource-driven automation

An authority may say:

"There are too many cases for individual consideration."

But administrative convenience does not automatically justify surrendering statutory discretion.

This is strongly illustrated by Barnard, where workload considerations did not justify the unlawful transfer of the disciplinary function.

9. Delegation Versus Use of Expertise

Not every use of AI is delegation.

This distinction is crucial.

Lawful technical assistance

Statute → authority decides → AI calculates data → authority evaluates data → authority decides.

Potential unlawful delegation

Statute → authority asks AI → AI determines substantive outcome → authority automatically approves.

Even stronger delegation concern

Statute → private vendor designs model → model determines eligibility → public authority merely implements vendor output.

The third situation raises questions concerning:

• sub-delegation;
• contracting-out;
• statutory authority;
• procedural fairness;
• transparency;
• accountability;
• conflicts of interest;
• judicial review; and
• constitutional responsibility.

10. Private AI Vendors and Delegation

The problem becomes particularly acute when government purchases a proprietary system.

Suppose a statute gives a public regulator discretion to determine whether a company is fit to operate.

The regulator purchases a proprietary AI system.

The vendor determines:

• what variables matter;
• how variables are weighted;
• what threshold constitutes risk;
• what data are excluded;
• how false positives are treated;
• how recommendations are ranked.

The regulator then follows the result.

The apparent legal delegation may therefore be hidden inside software procurement and model design.

The relevant legal question becomes broader:

Who actually determined the substantive policy embedded in the administrative decision?

If the answer is a private developer rather than the legally authorised public authority, conventional delegation principles become highly relevant.

11. Delegation of Policy Is More Sensitive Than Delegation of Processing

A useful distinction is:

Processing decisions

• sorting documents;
• calculating amounts;
• checking completeness;
• identifying duplicate applications.

These generally present lower delegation concerns.

Policy decisions

• determining who constitutes a risk;
• determining acceptable levels of enforcement;
• deciding which factors deserve greater weight;
• determining what constitutes sufficient evidence;
• establishing eligibility thresholds.

These are substantially more sensitive.

The public authority should ordinarily retain control over such normative choices because they can determine the substance of the statutory scheme.

12. Duty to Give Reasons

Automated administration also creates a reason-giving problem.

A public authority normally needs to be capable of explaining the legal basis and material reasoning behind an adverse administrative decision.

An explanation such as:

"The algorithm gave you a score of 83"

may be inadequate if the affected person cannot understand:

• which facts were relied upon;
• which factors mattered;
• whether relevant circumstances were considered;
• whether erroneous data were used;
• whether a statutory exception was considered; or
• why the particular legal consequence followed.

The SyRI and Bridges litigation demonstrates the increasing importance of transparency and legal safeguards surrounding algorithmic government.

13. Judicial Review of Machine-Assisted Discretion

Courts can potentially review an automated administrative decision under familiar grounds.

A. Illegality

Was the machine used within the statutory authority?

B. Improper delegation

Did the legally authorised decision-maker actually exercise the discretion?

C. Fettering

Was the official effectively bound by an inflexible algorithmic rule?

D. Relevant and irrelevant considerations

Did the system use legally relevant information?

Did it incorporate irrelevant proxies?

E. Procedural fairness

Could the affected person challenge erroneous information or methodology?

F. Equality and discrimination

Did the system produce unlawful discriminatory effects?

G. Proportionality

Was the algorithmic intervention proportionate to the legitimate governmental objective?

H. Reason-giving

Can the authority adequately explain the decision?

I. Irrationality

Was the decision or underlying system so unreasonable that it falls outside lawful administrative discretion?

Recent scholarship argues that judicial review of algorithmic administration may need to examine upstream elements such as objectives, thresholds and data selection, rather than examining only the final automated output.

14. A Proposed "Delegation Boundary" Test

A useful analytical framework is:

Level 1 — Pure calculation

Generally low delegation concern

Example: calculating statutory interest.

Level 2 — Information classification

Moderate concern

Example: sorting applications into categories.

Level 3 — Recommendation

Higher concern

Example: recommending which licence applications should be investigated.

Level 4 — Presumptive decision

High concern

Example: system automatically marks applications for rejection subject only to formal review.

Level 5 — Fully autonomous substantive discretion

Highest concern

Example: AI independently determines entitlement to a statutory benefit where legislation entrusts the judgment to a named public authority.

The more the system moves towards Level 5, the stronger the need for:

• express statutory authority;
• defined parameters;
• meaningful human judgment;
• auditability;
• reasons;
• review mechanisms;
• correction procedures;
• accountability; and
• judicial supervision.

15. Indian Administrative-Law Relevance

The principles are particularly compatible with Indian administrative-law doctrine.

Indian courts have recognised the principle that statutory discretion is ordinarily intended to be exercised by the authority to which it has been entrusted, unless sub-delegation is authorised expressly or by necessary implication. The Supreme Court's reasoning in Sahni Silk Mills (P) Ltd v Employees' State Insurance Corporation expressly discusses delegatus non potest delegare and the special sensitivity of quasi-judicial powers.

Consequently, an Indian public authority using an AI system would potentially need to establish:

1. statutory authority for the underlying decision;
2. legal authority for delegation or technological implementation;
3. appropriate limits on machine-generated recommendations;
4. meaningful human consideration;
5. compliance with natural justice;
6. consideration of relevant circumstances;
7. absence of arbitrary classification;
8. equality and Article 14 compliance;
9. protection of applicable constitutional and statutory rights; and
10. availability of effective judicial review.

16. Key Distinction: Automation Is Not Necessarily Delegation

It would be incorrect to say that every automated administrative act is unlawful.

Modern administration necessarily relies upon machinery.

The better proposition is:

Automation becomes legally problematic when the machine is allowed to determine matters that the law requires the authorised public decision-maker to determine, or when the authority uses automation in a manner that unlawfully fetters, obscures or replaces its statutory discretion.

That distinction reconciles Carltona with Barnard.

Carltona recognises administrative machinery.

Barnard limits transfer of substantive statutory judgment.

Lavender prohibits surrender of discretion.

Padfield requires statutory purpose to govern discretion.

Bridges, SyRI, Loomis and SCHUFA demonstrate how those principles become increasingly important when automated systems exercise substantial practical influence.

17. Emerging Doctrine: De Facto Delegation

The most important development is likely to be a shift from asking:

"Who formally signed the decision?"

to asking:

"Who—or what—actually exercised the decisive judgment?"

This would allow courts to examine:

• model architecture;
• training objectives;
• thresholds;
• variables;
• data sources;
• ranking mechanisms;
• default settings;
• human override rates;
• reasons for accepting or rejecting recommendations;
• vendor involvement; and
• institutional practices surrounding the system.

This approach is particularly important because recent academic analysis argues that conventional non-delegation doctrine may inadequately address algorithmic decision-making where the formal human decision-maker retains nominal authority but the substantive choices have already been embedded in the system.

18. Conclusion

The law of administrative delegation does not require government to abandon computers, algorithms or AI. Rather, it requires a distinction between using technology to exercise public power and allowing technology to acquire the substance of public power without adequate legal authority.

The major principles emerging from the case law are:

1. Routine administrative processing may be delegated or automated.
2. Statutory powers must ordinarily remain with the legally designated decision-maker.
3. Quasi-judicial and disciplinary powers receive particularly strong protection against delegation.
4. An authority cannot surrender its discretion to an algorithm merely by calling the algorithm a recommendation.
5. A human approval step must involve genuine independent judgment rather than mechanical endorsement.
6. Algorithmic rules cannot frustrate the statutory purpose for which discretion was granted.
7. The legality of an automated system can extend beyond its final output to its design, data, thresholds and deployment.
8. Private procurement of AI does not transfer public-law responsibility to the private vendor.
9. Transparency, reasons, equality, proportionality and procedural fairness remain applicable notwithstanding automation.
10. The decisive question is increasingly whether the authorised authority retained meaningful legal control over the substantive exercise of discretion.

Accordingly, the strongest legal boundary is not "human versus machine" but "lawfully authorised human judgment versus unaccountable substitution of machine-generated judgment." The modern administrative state can use machines extensively, but the statutory responsibility for public power cannot simply disappear into the algorithm.