Re-training Obligations for Models

“Re-training obligations for models” generally refers to the legal and contractual responsibilities of an organisation to re-train, fine-tune, validate, or otherwise update an AI/ML model when its performance, accuracy, fairness, security, or compliance becomes inadequate. In an employment-law context, the issue can arise where employers use AI models for recruitment, performance evaluation, employee monitoring, promotion, dismissal, scheduling, or other HR decisions.

There is no single general legal rule requiring every AI model to be re-trained at fixed intervals. Instead, the obligation may arise from data-protection duties, anti-discrimination law, sector-specific regulation, contractual commitments, internal governance policies, or the need to correct demonstrably defective automated decision-making.

1. When re-training may become necessary

A model may require re-training or modification when:

• Performance deteriorates: The model's predictions become materially less accurate because the underlying data or environment has changed.
• Data drift occurs: Employee demographics, job roles, workplace practices, or other relevant characteristics change over time.
• Historical bias is identified: Training data contains discriminatory patterns that the model reproduces.
• Legal requirements change: A new law or regulatory requirement makes the existing model or its use non-compliant.
• Personal data is improperly used: The organisation needs to modify training practices or remove problematic data.
• Security risks emerge: Adversarial manipulation or other security problems make the model unreliable.
• The model affects employment rights: Higher scrutiny may be appropriate when automated systems influence hiring, promotion, compensation, discipline, or termination.

Re-training is not necessarily the only remedy. Depending on the problem, an employer may need to change the model, remove particular data, alter decision thresholds, introduce human review, suspend the model, or replace it altogether.

2. Employment discrimination and re-training

AI systems used in employment can reproduce discrimination contained in historical employment data. For example, if an employer historically hired fewer women for a particular position, a model trained on those records could learn a relationship between gender-associated characteristics and successful hiring outcomes.

If testing demonstrates discriminatory effects, simply continuing to use the model may expose the employer to legal risk. Possible corrective measures include:

1. auditing the training data;
2. identifying discriminatory variables or proxies;
3. re-training using appropriately reviewed data;
4. testing the revised model;
5. monitoring outcomes after deployment; and
6. maintaining meaningful human oversight.

The legal issue is therefore broader than merely asking whether an employer “re-trained” the model.

3. Data protection considerations

Where employee or applicant personal data is used for model training, organisations must consider applicable data-protection requirements concerning:

• lawful processing;
• purpose limitation;
• data minimisation;
• accuracy;
• retention;
• transparency;
• security; and
• rights relating to automated decision-making, where applicable.

Re-training with newly collected employee data does not automatically make the processing lawful. The employer must still have an appropriate legal basis and comply with applicable data-protection requirements.

4. Accuracy and ongoing monitoring

AI models are not necessarily static systems. A model that was accurate when deployed may become less reliable later.

For example, an employee-performance model trained using historical working patterns might become unreliable after:

• a transition to hybrid work;
• organisational restructuring;
• changes in job responsibilities;
• changes in workforce composition; or
• changes in the performance criteria used by the employer.

An organisation should therefore establish model monitoring and review procedures. A policy might specify performance thresholds that trigger investigation, re-validation, re-training, or suspension.

5. Human oversight

Re-training does not eliminate the importance of human decision-making. Where an AI model makes or substantially influences an employment decision, employers should consider whether a qualified person should review the result.

For example, an automated system might flag an employee as a “high-risk” disciplinary candidate. Before disciplinary action is taken, the employer should consider:

• the data relied upon;
• whether the information is accurate;
• whether the model produced an anomalous result;
• whether the employee has an opportunity to explain the circumstances; and
• whether the final decision is independently reviewed.

6. Documentation of re-training

Employers using AI systems should maintain records concerning:

• the original model and version;
• training datasets and their provenance;
• reasons for re-training;
• identified performance or bias problems;
• changes made to the training data;
• validation and testing results;
• approval of the new model;
• deployment date;
• responsible personnel; and
• post-deployment monitoring.

Such documentation can become important in litigation, regulatory investigations, audits, or employee disputes.

7. Contractual obligations

A re-training obligation can also arise from a contract. An AI vendor may promise that its system will:

• maintain specified accuracy levels;
• comply with applicable laws;
• undergo periodic testing;
• correct identified defects; or
• update its model when necessary.

If the vendor fails to perform those contractual obligations, the employer may have contractual remedies depending on the agreement.

8. Regulatory approach

Modern AI regulation increasingly focuses on risk management, monitoring, data quality, accuracy, and human oversight, rather than imposing a universal rule that every model must be re-trained at a particular interval.

Consequently, an organisation should be able to demonstrate that it has a rational process for determining when re-training is necessary and when another corrective measure is more appropriate.

Important Case Laws

1. State of Andhra Pradesh v. V. Sadanandam, (1989) 4 SCC 181

The Supreme Court recognised that decisions affecting public employment must comply with constitutional standards and cannot be based on arbitrary considerations.

Relevance: Where an automated employment system produces arbitrary or irrational outcomes, the employer may need to review the system and its underlying methodology rather than blindly rely on automated outputs.

2. E.P. Royappa v. State of Tamil Nadu, (1974) 4 SCC 3

The Supreme Court established that arbitrariness is antithetical to equality under Article 14.

Relevance: An AI-based employment decision that produces arbitrary outcomes may raise concerns where the employer is subject to constitutional standards. Model validation and corrective action become particularly important when automated processes produce unexplained or inconsistent results.

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

The Supreme Court developed the principle that state action affecting rights must satisfy standards of fairness and non-arbitrariness.

Relevance: Automated systems used by public authorities cannot be treated as beyond scrutiny merely because a computer or AI system generated the result.

4. Justice K.S. Puttaswamy (Retd.) v. Union of India, (2017) 10 SCC 1

The Supreme Court recognised privacy as a fundamental right and discussed principles concerning informational privacy and personal autonomy.

Relevance: Where employee or applicant personal data is used to train or re-train AI models, organisations must consider privacy and lawful-processing implications. Re-training cannot be treated as automatically permissible merely because the organisation already possesses the underlying information.

5. Justice K.S. Puttaswamy (Retd.) v. Union of India, (2019) 1 SCC 1

The Supreme Court's Aadhaar judgment examined proportionality, purpose limitation, data protection and safeguards surrounding the use of personal information.

Relevance: Large-scale use of personal information for algorithmic systems highlights the importance of limiting data use to legitimate purposes and implementing appropriate safeguards. These principles are relevant when employee data is incorporated into model training or re-training.

6. Anuradha Bhasin v. Union of India, (2020) 3 SCC 637

The Supreme Court emphasised the need for legality, proportionality and procedural safeguards when state measures affect fundamental rights.

Relevance: Although the case did not concern AI model re-training, its proportionality reasoning can be relevant where automated systems used by public employers significantly affect employees' rights or interests.

7. Swiss Ribbons Pvt. Ltd. v. Union of India, (2019) 4 SCC 17

The Supreme Court discussed the importance of rational classification and non-arbitrary decision-making.

Relevance: AI systems used in employment should operate according to rational and legally defensible criteria rather than opaque or arbitrary classifications.

8. Navtej Singh Johar v. Union of India, (2018) 10 SCC 1

The Supreme Court emphasised constitutional equality, dignity and protection against discriminatory treatment.

Relevance: Where an employment model systematically disadvantages individuals because of characteristics protected by applicable equality law, organisations should investigate the underlying data and model design and take corrective measures, which may include re-training.

Key Compliance Measures

An organisation using AI models for employment purposes should consider establishing a model lifecycle policy containing:

1. Initial validation before deployment.
2. Periodic performance testing.
3. Bias and discrimination testing.
4. Data-quality assessments.
5. Defined thresholds for triggering re-training.
6. Human review for significant employment decisions.
7. Version control for every model.
8. Documentation of training and re-training data.
9. Privacy and data-protection assessments.
10. Post-retraining validation and monitoring.
11. A process for employee complaints or challenges.
12. A mechanism to suspend or withdraw a model that creates unacceptable legal or operational risks.

Conclusion

A general legal obligation to re-train every AI model periodically does not exist. Instead, the need for re-training may arise from an organisation's obligations concerning accuracy, discrimination, privacy, fairness, security, contractual commitments, and responsible automated decision-making. In employment settings, organisations should not rely on an AI system merely because it was previously validated. Continuous monitoring, documented review, human oversight and corrective action—including re-training where appropriate—are important components of responsible AI governance.