Algorithmic bias is the AI risk that regulators, courts and customers ask about most. A hiring model that favours one group, a credit score that penalises a postcode, or a medical tool that performs worse for some…
Algorithmic bias is the AI risk that regulators, courts and customers ask about most. A hiring model that favours one group, a credit score that penalises a postcode, or a medical tool that performs worse for some patients can each trigger legal liability, reputational damage and real harm. Two international standards set out how to deal with it, and they take very different approaches.
ISO/IEC TR 24027:2021, Bias in AI systems and AI aided decision making, is a Technical Report from the international AI committee ISO/IEC JTC 1/SC 42. It explains what bias is, where it comes from, how to measure it and how to treat it. IEEE Std 7003-2024, the Standard for Algorithmic Bias Considerations, published on 24 January 2025, is a normative process standard. It requires organisations to make and document deliberate decisions about bias across an AI system's life cycle, recorded in a "bias profile".
This guide compares the two, explains how they fit with the EU AI Act and ISO/IEC 42001, and shows how to use both in one bias management programme.
- ISO/IEC TR 24027:2021
- Information technology — Artificial intelligence — Bias in AI systems and AI aided decision making; published November 2021
- Type
- Technical Report: informative guidance, no requirements
- Developed by
- ISO/IEC JTC 1/SC 42 (artificial intelligence); adopted in Europe as CEN/CLC ISO/IEC/TR 24027:2023
- IEEE Std 7003-2024
- IEEE Standard for Algorithmic Bias Considerations; approved 11 December 2024, published 24 January 2025
- Type
- Normative process standard with requirements
- Core artefact
- IEEE 7003: the bias profile · ISO/IEC TR 24027: no required artefact
- Related standards
- ISO/IEC TS 12791:2024 (treating unwanted bias); prEN 18283 (EU harmonised bias standard in development)
The short answer
ISO/IEC TR 24027 is a knowledge resource. It gives you shared vocabulary, a catalogue of bias sources, a set of fairness metrics and treatment techniques across the AI life cycle. It tells you what bias is and what you could do about it, but it doesn't require anything.
IEEE 7003 is a governance process. It requires you to decide which differences in treatment are acceptable for your system, identify who is affected, check whether your data represents them, evaluate the system, assess risks and impacts, and record all of it in a bias profile that follows the system through its life. It tells you how to manage bias decisions and prove you made them.
ISO/IEC TR 24027 answers
"What kinds of bias exist in AI systems, where do they come from, how can we measure them, and which techniques can treat unwanted bias?"
IEEE 7003 answers
"Have we deliberately defined, assessed, mitigated and documented bias for this system, with the reasoning on record, across its whole life cycle?"
What ISO/IEC TR 24027 covers
ISO/IEC TR 24027 addresses bias in AI systems, especially where AI supports human decisions. Its core insight is that not all bias is bad. Some bias is necessary to meet a system's objectives (desired bias), while other bias isn't intended and creates unfair outcomes (unwanted bias). The report also distinguishes bias from fairness: a system can be technically unbiased yet still unfair, for example a policy that rejects every candidate.
| Clause | Content |
|---|---|
| Clause 5: overview | Concepts of bias and fairness, and their relationship to AI-aided decision-making |
| Clause 6: sources of unwanted bias | Human cognitive bias, data bias (for example sampling, labelling and non-representative data) and bias from engineering decisions such as feature selection and model design |
| Clause 7: assessing bias and fairness | Metrics, many based on the confusion matrix, such as equalised odds, equality of opportunity, demographic parity and predictive equality |
| Clause 8: treating unwanted bias | Techniques across the life cycle: data collection and preparation, training, adversarial methods, rule-based systems, evaluation and use |
Because it is a Technical Report, it uses descriptive language rather than requirements. You can't be certified against it or claim conformance. Its value lies in giving engineers, data scientists, auditors and lawyers a common language for bias, and it is widely cited by later standards, including ISO/IEC TS 12791 and ISO/IEC 42001.
ISO/IEC TS 12791:2024, Treatment of unwanted bias in classification and regression machine learning tasks, builds directly on TR 24027. It turns the report's treatment ideas into a more structured, life-cycle-based method for identifying, mitigating and monitoring unwanted bias in common machine learning tasks. Most organisations that use TR 24027 for concepts use TS 12791 for method.
What IEEE 7003 requires
IEEE 7003 is part of the IEEE 7000 series on ethically aligned design. It defines bias broadly as any systematic difference in treatment, and treats bias as a design choice: something an organisation must state as wanted or unwanted, against a definition of what is acceptable in that context of use.
It is deliberately a governance and process standard. It defines what to consider and why, not which fairness metric or threshold to use. Its work is organised into five activities, all feeding one artefact:
Requirements for bias consideration
Define which differences in treatment are wanted or unwanted for the system, and to what extent, in its intended context of use.
Stakeholder identification
Identify who influences the system and who is affected by it, including groups at risk of unwanted bias.
Data representation
Check whether the data adequately represents all affected stakeholders, and document decisions on inclusion, exclusion and data governance.
System evaluation
Evaluate the system's behaviour, including validation data selection and stated application boundaries beyond which it shouldn't be used.
Risk and impact
Assess the risks and impacts of remaining bias on stakeholders, decide on mitigations, and manage user expectations.
Everything is recorded in the bias profile, a living record of every bias-related decision from data selection to model output. It is updated as the system, its data or its context changes, and it creates the traceability auditors and regulators ask for.
If you are building a bias management programme, I post practical breakdowns of AI standards, the EU AI Act and ISO 42001 regularly on LinkedIn. You're welcome to connect with me there and tell me which bias question your team is wrestling with.
IEEE 7003 vs ISO/IEC TR 24027: side by side
| Dimension | ISO/IEC TR 24027:2021 | IEEE Std 7003-2024 |
|---|---|---|
| Publisher | ISO and IEC (JTC 1/SC 42) | IEEE Standards Association |
| Document type | Technical Report (informative) | Standard (normative) |
| Uses "shall" requirements? | No | Yes |
| Can you claim conformance? | No | Yes, against its process requirements |
| Core focus | Concepts, sources, metrics and treatment techniques | Life-cycle process and documented decisions |
| Definition of bias | Desired versus unwanted bias, distinct from fairness | Any systematic difference in treatment, classified as wanted or unwanted in context |
| Fairness metrics | Describes specific metrics | Doesn't prescribe metrics or thresholds |
| Stakeholders | Discussed as context | A required activity |
| Documentation | No required artefact | Bias profile, maintained across the life cycle |
| Scope | AI systems and AI-aided decision-making | Algorithmic systems, including AI and autonomous systems |
| Best used for | Training, shared vocabulary, choosing metrics and techniques | Governance, accountability, audit evidence |
| Works best with | ISO/IEC TS 12791, ISO/IEC 42001, ISO/IEC 23894 | The IEEE 7000 series and any AI management system |
How the two fit together
These standards aren't competitors. They answer different questions, and together they cover the full job:
Use IEEE 7003 for "how"
Run the process: define acceptable differences, identify stakeholders, check data, evaluate, assess impact and keep the bias profile.
Use TR 24027 for "what"
Use its catalogue of bias sources to know what to look for, and its metrics to measure what IEEE 7003 asks you to evaluate.
Use TS 12791 to "fix"
Apply structured mitigation techniques for classification and regression models, then record the results in the bias profile.
TR 24027 lists several fairness metrics, and they can contradict each other. A model can satisfy demographic parity while failing equalised odds. IEEE 7003's insistence on documenting which differences are acceptable, and why, is what makes a metric choice defensible.
Where the law comes in
Neither standard is legally required. But both help organisations meet legal duties on bias that are now arriving.
| Legal requirement | What it asks | How the standards help |
|---|---|---|
| EU AI Act Article 10(2)(f)–(g) | Examine training, validation and test data for possible biases, and take measures to detect, prevent and mitigate them, for high-risk AI | TR 24027 identifies data bias sources and metrics; IEEE 7003's data representation activity and bias profile provide evidence |
| EU AI Act bias-data processing | Special-category data may be processed for bias detection and correction under strict safeguards | The IEEE 7003 bias profile documents why such data is necessary and how it is protected |
| EU AI Act Article 27 FRIA | Certain deployers assess fundamental-rights impacts, including discrimination | IEEE 7003's stakeholder and risk-and-impact activities feed directly into the assessment |
| Anti-discrimination law | Avoid unlawful discrimination in employment, credit, housing and services | Documented bias decisions support a defence and show due diligence |
| US state and local rules | For example, bias audits for automated hiring tools in New York City | TR 24027 metrics support audits; the bias profile records results |
CEN-CENELEC JTC 21 is developing prEN 18283, Concepts, measures and requirements for managing bias in AI systems, as a candidate harmonised standard under the AI Act. Once it is finished and cited in the Official Journal, it could give providers a presumption of conformity for bias-related requirements. Neither IEEE 7003 nor TR 24027 will carry that status, but both are useful preparation.
How the timelines fit together
- 2017
IEEE launches the P7003 project on algorithmic bias considerations.
- Nov 2021
ISO/IEC TR 24027 published.
- Dec 2023
ISO/IEC 42001 published, referencing bias treatment within its controls.
- 2023
TR 24027 adopted in Europe as CEN/CLC ISO/IEC/TR 24027:2023.
- 1 Aug 2024
EU AI Act enters into force, including Article 10 bias duties.
- Aug 2024
Work on prEN 18283 begins in CEN-CENELEC JTC 21.
- 2024
ISO/IEC TS 12791 published.
- 11 Dec 2024
IEEE 7003 approved.
- 24 Jan 2025
IEEE 7003-2024 published.
- 2 Dec 2027
EU AI Act high-risk duties, including Article 10, apply to Annex III systems.
A 6-step plan to use both standards
Train teams on TR 24027 vocabulary
Give product, data and legal teams a shared understanding of desired versus unwanted bias, bias sources and the difference between bias and fairness.
Start an IEEE 7003 bias profile for each priority system
Begin with systems that affect people's access to jobs, credit, housing, healthcare or public services, and those classed as high-risk under the AI Act.
Define acceptable differences with stakeholders
Decide, with input from affected groups, which differences in treatment are justified and which aren't, and record the reasoning.
Choose and justify metrics
Select fairness metrics from TR 24027 that fit the context, explain why in the bias profile, and test for intersectional groups, not just single attributes.
Treat bias and monitor it
Apply mitigation techniques from TR 24027 and TS 12791, set application boundaries, and monitor for drift after deployment.
Connect to your management system and the law
Link bias profiles to your ISO/IEC 42001 AI management system, AI Act Article 10 evidence and impact assessments, and track prEN 18283 as it develops.
Common mistakes to avoid
- Claiming "ISO 24027 certification", which doesn't exist for a Technical Report
- Treating IEEE 7003 as a fairness-metric standard, when it is a process standard
- Picking one fairness metric without explaining why it fits the context
- Testing bias once before launch and never again
- Removing protected attributes and assuming that removes bias, when proxies can remain
A prompt to build a bias profile for your AI system
Paste in your system details to get a first-draft bias profile. Your data science, legal and governance leads should review it together.
You are an AI fairness specialist with deep knowledge of IEEE Std 7003-2024, ISO/IEC TR 24027:2021, ISO/IEC TS 12791:2024 and EU AI Act Article 10. Our AI system: [purpose and decisions it informs; context of use; people affected; model type; training, validation and test data sources; attributes available (including any protected or proxy attributes); current testing; deployment setting; whether it is high-risk under the EU AI Act] Produce: 1. A draft IEEE 7003 bias profile with sections for requirements, stakeholders, data representation, system evaluation, and risk and impact. 2. A list of likely sources of unwanted bias, using the ISO/IEC TR 24027 categories (human cognitive, data, engineering). 3. Two or three fairness metrics suited to this context, with the reasoning for each and the trade-offs between them. 4. Mitigation options from TR 24027 and TS 12791 for each main risk. 5. Application boundaries: where the system should not be used. 6. How this evidence supports EU AI Act Article 10 and any fundamental rights impact assessment. Use only what I provide and mark uncertain points UNKNOWN.
ISO/IEC TR 24027 teaches you to see bias. IEEE 7003 makes you decide what to do about it and write down why. You need both the eyes and the record. — Kumail Mehdi
Frequently asked questions
What is the difference between IEEE 7003 and ISO/IEC TR 24027?
ISO/IEC TR 24027 is an informative Technical Report that explains bias concepts, sources, metrics and treatment techniques. IEEE 7003 is a normative process standard that requires organisations to make and document bias decisions across the AI life cycle in a bias profile.
Can I get certified to ISO/IEC TR 24027?
No. Technical Reports contain no requirements, so there is nothing to certify against. Organisations use TR 24027 as guidance, often alongside ISO/IEC 42001, which is certifiable as a management system.
What is an IEEE 7003 bias profile?
A living record of all bias-related considerations and decisions for an algorithmic system, from requirements and stakeholders to data, evaluation, risks and mitigations. It is maintained throughout the system's life cycle and provides traceability for audits.
Does IEEE 7003 tell you which fairness metric to use?
No. It defines what to consider and why, not which metric or threshold to apply. That choice depends on context, and must be justified and recorded. ISO/IEC TR 24027 describes metrics you can choose from.
Do these standards help with the EU AI Act?
Yes, but neither gives a presumption of conformity. They support Article 10 duties to examine and mitigate data bias for high-risk AI, and feed fundamental rights impact assessments. The EU's own harmonised bias standard, prEN 18283, is in development.
What is ISO/IEC TS 12791?
A 2024 Technical Specification on treating unwanted bias in machine learning classification and regression tasks. It builds on ISO/IEC TR 24027 and gives a more structured, life-cycle-based method for identifying, mitigating and monitoring bias.
Master AI governance in 6 months, 15 minutes a week
AI Governance – Decoded is my weekly newsletter that explains AI standards, the EU AI Act, ISO 42001 and the frameworks around them in plain language. One short, practical read every week.
Join the newsletterBuilding a bias management programme? Let's connect on LinkedIn. I'm always glad to compare notes on fair and accountable AI.
Sources
- IEEE SA — IEEE 7003 implementation overview (2026)
- DLA Piper — Landmark AI framework sets new standard for tackling algorithmic bias
- Modulos — IEEE 7003 algorithmic bias considerations
- ISO — ISO/IEC TR 24027:2021 Bias in AI systems and AI aided decision making
- iTeh Standards — ISO/IEC TR 24027:2021 overview
- iTeh Standards — ISO/IEC TS 12791:2024
- CEN-CENELEC — European standards supporting the AI Act (prEN 18283)
- DIN — prEN 18283 project page
- EUR-Lex — Regulation (EU) 2024/1689 (AI Act)
Kumail Mehdi
ISO 42001 Lead Auditor and AI strategist. 11 years in corporate leadership, 14 years running a digital agency. I help professionals, consultants and AI startups turn expertise into governed, AI-powered systems.
