TEC 57050:2023 - Standard for Fairness Assessment and Rating of Artificial Intelligence Systems

The Telecommunication Engineering Centre is the technical arm of the Department of Telecommunications. Its functions include framing standards, formulating essential requirements for mandatory testing and certification of telecom equipment, designating conformity assessment bodies, and testing and certification of telecom products, through four regional centres in New Delhi, Bangalore, Mumbai and Kolkata.¹

TEC's work on AI fairness began with a letter of 22 February 2022 inviting stakeholder inputs on a framework for fairness assessment of AI/ML systems. As reported by MediaNama, the letter asked about classifying systems by learning type, about biases such as selection and measurement bias, about fairness metrics such as demographic parity, equal opportunity and average odds, about what a fairness rating procedure should include, and about handling code and training data safely during third-party assessments. It described the exercise as a possible benchmark that could lead to certification by a neutral government agency.²³ TEC's consultations register lists a draft "Standard for assessing and rating Artificial Intelligence Systems for fairness" issued on 29 December 2022, with comments due by 13 February 2023, followed by the final standard.³ TEC's website records the final standard as released on 7 July 2023.⁴

The standard grew out of India's national policy framework. The National Digital Communications Policy 2018 calls for synergistic deployment and adoption of AI, and TEC's document states that its framework is built on NITI Aayog's Principles of Responsible AI, specifically equality, inclusivity and non-discrimination.¹ The Telecom Regulatory Authority of India's July 2023 recommendations had separately suggested that TEC could accredit laboratories for testing and certification of AI products and that a future AI authority should coordinate with TEC on that function.⁹

The standard's stated objective is to promote bias assessment by providing a standard procedure for assessing, measuring, presenting and transparently disclosing bias, and thereby to enhance trust in AI systems. It also aims to give a reference scale for comparing systems. TEC's introduction observes that fairness is subjective, so that "metrics or thresholds cannot be applied uniformly across use cases, domains, or industrial environments".¹

The document identifies several categories of user:¹

On legal status, the standard says the framework can primarily help a developer self-certify, and that "the intention is not to make the framework a mandatory regulatory assessment tool".¹

Sources of bias. The standard classifies bias as individual, group or multi-group, and as pre-existing, technical or emergent. It maps process factors and technical factors that can introduce bias at each component of the AI lifecycle, beginning with data (for example historical bias, under-representation, imputation of missing values, outlier treatment and labelling).¹

Step 1: bias risk classification. The auditee completes a questionnaire covering seven areas: the AI system, data, model, pipeline, interface and integration, human-in-the-loop, and deployment. Systems are classed as high, medium or low risk according to how far results affect a person's life, their rights and freedoms, and the principles of equality, inclusivity and non-discrimination. High-risk examples include disease-diagnosis algorithms and autonomous vehicles, as well as AI that decides eligibility for benefits. Medium risk covers AI affecting convenience or financial choices, and low risk covers AI with limited or no bearing on individuals.¹

Step 2: metrics, thresholds and benchmarks. The auditee selects protected attributes, privileged and unprivileged groups and the favourable outcome, then chooses fairness metrics and thresholds. The standard presents group-fairness metrics for classification, including demographic parity (with the four-fifths rule that the ratio between groups should not fall below 0.8), minimum allocation guarantees and counterfactual fairness, and balance metrics for clustering.¹

Step 3: bias testing. Testing combines process-based checks, metric-based checks and scenario testing. Scenario testing uses counterfactual fairness, edge cases, long-tail risk and adversarial tests to surface previously unidentified bias. Procedures differ according to whether the evaluator has open-box, grey-box or closed-box access to the model.¹

Scoring. The standard standardises results on a linear scale. A Bias Index is computed for each protected attribute by combining the individual metrics, with an ideal value of zero, and a single Fairness Score is computed for the whole model, with an ideal value of one.¹

The evaluation outcome report has three parts: a summary (including protected attributes, favourable outcome, whether the audit was a self-test or independent, level of data access and limitations), a tabulation of metrics with thresholds and measured values, and a detailed report setting out assumptions, justifications and tests performed.¹

Certificates are time-limited. A self-certified report is valid for three months or less, an audited report for six months or the period shown on the certificate, and a mandated certification for a maximum of one year. Systems that keep learning during deployment must be renewed at least once a quarter, and the preparer of the report is accountable for keeping it up to date.¹

The standard is candid about its limits. It covers pre-existing, technical and emergent bias and group and individual bias, but only for tabular data where each row is independent. Text, image and speech data and models such as reinforcement learning, GANs and autoencoders are left to future versions, as are the fairness of interfaces, pipelines, infrastructure and deployment, and risk evaluation for those components.¹ It also acknowledges that no framework can flex to cover every type of AI development process.¹

The sources reviewed for this entry did not identify published independent critiques of the standard, nor any regulator that has made certification under TEC 57050:2023 mandatory. Practical questions that the standard itself leaves open include who would accredit third-party auditors and how thresholds would be chosen in each sector.

TEC has since widened its AI standards programme. On 18 October 2023, shortly after publishing the fairness standard, it wrote to stakeholders proposing a voluntary standard for assessing and rating the robustness of AI systems in telecom networks and digital infrastructure, with inputs requested by 15 November 2023. The letter described robustness as the ability to perform reliably under diverse and unpredictable conditions, including adversarial attacks, scalability, resilience and safety.⁶

A draft of that standard, TEC 57070:2025 on robustness assessment and rating of AI systems, was put out for consultation on 6 June 2025 with comments due by 5 August 2025. Its abstract describes an assessment methodology that covers resilience to data shifts, integrity, reliability, explainability, transparency, privacy and security, and a rating method to benchmark robustness.³⁷

TEC also circulated a draft of TEC 57090:2025, a standard for the schema and taxonomy of an AI incident database in telecommunications and critical digital infrastructure, in May 2025, and its standards listing records the standard as released on 26 November 2025.³⁴⁸

As of 15 October 2026 TEC 57050:2023 is published on TEC's website and is listed among its standards. It remains voluntary, and the sources reviewed did not show it being referenced in any binding regulation.¹⁴ The robustness standard TEC 57070 was a draft in the sources reviewed; the final status of that standard was not confirmed.

The fairness standard sits alongside other Indian AI-governance instruments, including NITI Aayog's responsible AI principles on which it is built and the standards activity of the Bureau of Indian Standards, which is covered in a separate entry.