Frequently asked questions about IEEE 7003

Each entry is short and refers to the section that gives the full rule.

Disclaimer

This page gives general information. It is not legal advice.

What is IEEE 7003?

IEEE Std 7003-2024 is the IEEE Standard for Algorithmic Bias Considerations. It is a voluntary standard for bias consideration in an autonomous intelligent system (AIS) during all of its life cycle. The IEEE SA Standards Board approved it on 11 December 2024, and IEEE published it on 24 January 2025. In October 2026, the IEEE SA page gives the status of the standard as “Active Standard”.

Its normative Clauses 4 to 9 are in a sequence around the bias profile. The Introduction, the NOTEs, and Annexes A to F are not normative. See The clauses of the standard.

Is IEEE 7003 a fairness-metrics standard?

IEEE 7003 does not give the decision criteria or differential impacts that identify unwanted bias (1.4), and it sets no “threshold”. The attributes that make bias unwanted can change with the “social and societal context” of use (1.4).

But Clause 9.2.4.3 b sets a step to select a metric for the evaluation of bias, with the testing scenarios and reference values. It also sets steps to “define and justify” the evaluation method. IEEE 7003 requires a “rationale” for the risk and impact metrics (8.4 a). Annex D, Measuring Bias, is informative. See Metrics and the context of use.

What is the difference between wanted and unwanted bias in IEEE 7003?

In IEEE 7003, wanted bias is bias that is necessary for the functional requirements of an AIS. It is also permitted to add wanted bias to an AIS, to correct unwanted bias (5.1). Unwanted bias can have a bad effect on the “functional objectives” of the AIS, or it can cause damage to stakeholders (4.1).

The attributes that make bias unwanted can change with the context of use (1.4). The normative clauses set steps to “consider” the bias intent, and steps to mitigate unwanted bias (1.6). But the specification of mitigations is not in the scope of the standard (5.2). See Wanted and unwanted bias.

What is the bias profile in IEEE 7003?

In IEEE 7003, the bias profile is a repository of information that the tasks of bias consideration make and keep (3.1). It is a record of the bias consideration for an AIS during all of its life cycle (5.1). IEEE 7003 puts its clauses in a sequence around the bias profile (1.6). Clause 4 sets it up (4.4 e). Each of Clauses 6 to 9 uses it as an input and gives an updated bias profile as an output.

The bias profile holds the “initial”, the “intermediate”, and the “most recent versions” of the “documents” of each task (5.2). See The bias profile (Clause 5).

Who does IEEE 7003 apply to?

IEEE 7003 applies to algorithmic systems, not to organizations, for all computational approaches. It applies to each algorithmic system that has a part in “selection, allocation, ranking, decision-making”, or in other procedures in which some parties can get different outcomes from other parties (1.4). IEEE 7003 is for persons such as suppliers, procurers, developers, operators, maintainers, system auditors, users, and managers (1.7). These persons can be in the organizations of the supplier and the procurer, or in an organization that makes and uses the system (1.7). IEEE 7003 is also for data scientists, AI researchers, and their research organizations (1.7).

It applies when a new system is in design, or when a system is changed or decommissioned. It also applies when a system uses data that is different from its training data. It also applies when a system is put in use in a new context of use (1.4). Its scope does not include systems or procedures that find and examine “problematic algorithms” in use, or tracking and certification (1.5). See Where IEEE 7003 applies.

How is conformance to IEEE 7003 claimed?

A project or organization claims conformance, and it can use IEEE 7003 as its conformance document (1.8). Conformance requires each “shall” of the normative clauses, not only one step (1.3, 1.6). Clause 1.8 also gives two more uses that are permitted. A procurer and a supplier are permitted to name IEEE 7003 in a contract, if they agree that the supplier supplies “services and systems” in “accordance” with it. A project or organization is permitted to use it as an in-house standard, if it makes the decision to get information for users from a different part of the organization in “accordance” with it. To claim conformance, it is not necessary to use the nomenclature of IEEE 7003 for the parts of the information for users, such as pages (1.8).

IEEE 7003 does not include tracking or certification (1.5). Conformance does not release a party from its legal or other obligations (1.9). A completed Review of a Requirement in Modulos is not conformance. It does not sign the risk and impact assessment (8.4 f) or give its formal acceptance (8.5). See Conformance and Work that the Frameworks do not hold.

Is IEEE 7003 mandatory?

No. Use of IEEE 7003 is “wholly voluntary” (the first pages of the standard). But a procurer and a supplier are permitted to include IEEE 7003 in a contract or in “similar agreements”, if they agree that the supplier supplies “services and systems” in “accordance” with it. An obligation then comes from the contract or the “similar agreements” (1.8). Conformance does not release a party from its legal or other obligations (1.9).

Laws can set different requirements. For example, Article 10(2), points (f) and (g), of the EU AI Act applies to the “training, validation and testing data sets” of high-risk AI systems that use “techniques involving the training of AI models” (Article 10(1)). For other high-risk AI systems, it applies “only to the testing data sets” (Article 10(6)).

Article 16, point (a), gives these obligations to “providers of high-risk AI systems”. These rules apply from 2 December 2027 for Annex III systems and from 2 August 2028 for Annex I systems (Article 113, point (c)). Conformance to IEEE 7003 does not give the presumption of conformity of Article 40(1). See Conformance and How IEEE 7003 compares to other frameworks.

How does Modulos model IEEE 7003?

Modulos holds IEEE 7003 as two Frameworks with the name IEEE 7003. Add MFF-25 to the AI System of each AIS, for the work from Clause 4 to Clause 9. Add OFF-25 to the organization AI System, for the provisions that Modulos also reads at the Organization level. If your organization has no organization AI System, contact Modulos before you add OFF-25.

IEEE 7003 has no Organization level, and its Clause 4 steps are for each AIS (1.6 a). Thus OFF-25 does not replace them. A Requirement status in Modulos does not make an AIS conform. See Recommended AI System structure and The Requirements.

How does IEEE 7003 compare to the EU AI Act, ISO/IEC 42001, and the NIST AI RMF?

IEEE 7003, the EU AI Act, ISO/IEC 42001, and NIST AI 100-1 are not equivalent. IEEE 7003 is a voluntary standard for bias consideration in an AIS.

The EU AI Act is EU law. Its Article 10(2), point (f), sets an “examination in view of possible biases that are likely to affect the health and safety of persons, have a negative impact on fundamental rights or lead to discrimination prohibited under Union law” for the data sets of high-risk AI systems. Point (g) sets measures against the biases that point (f) identifies. These rules apply to the “training, validation and testing data sets” of high-risk AI systems that use “techniques involving the training of AI models” (Article 10(1)). For other high-risk AI systems, they apply “only to the testing data sets” (Article 10(6)).

Article 16, point (a), gives these obligations to “providers of high-risk AI systems”, from 2 December 2027 (Annex III) or 2 August 2028 (Annex I) (Article 113, point (c)). Conformance to IEEE 7003 does not give the presumption of conformity of Article 40(1). Regulation (EU) 2026/1744 added Article 4a, “Processing of special categories of personal data for bias detection and correction”, and removed Article 10(5).

ISO/IEC 42001 is a voluntary standard for an AI management system, with an AI system impact assessment (6.1.4, 8.4). NIST AI 100-1 is voluntary guidance that gives “three major categories of AI bias” (Section 3.7). See How IEEE 7003 compares to other frameworks.

What does the IEEE 7003 bias-consideration setup require?

Clause 4 of IEEE 7003 (Requirements for Bias Consideration) gives two outputs for an AIS (4.3). The first output is the bias profile, with the bias requirements and the boundaries of acceptability. The other output is a values statement for the AIS, with bias considerations.

Clause 4.4 gives eight steps, for example to select an accountability structure, to set up the bias profile, and to “propose” boundaries of acceptability. Step 4.4 b adds external advocates if the persons in the bias consideration process do not “appropriately” represent the sensitive attributes. Step 4.4 a names the governance processes as “essential”, but the items of 4.2 c are inputs only “where available”. IEEE 7003 requires that Clause 4 is done again during all of the AIS life cycle, as “appropriate” (1.6). See Requirements for bias consideration (Clause 4).

What are boundaries of acceptability in IEEE 7003?

In IEEE 7003, Step 4.4 g is to “propose” boundaries of acceptability for an AIS in its context of use. It gives diversity, inclusion, disability, “culture”, and accessibility as examples. The bias profile of the AIS includes boundaries of acceptability, together with the bias requirements (4.3 a). IEEE 7003 does not give boundaries of acceptability for all systems. The attributes that make bias unwanted can change with the context of use (1.4).

IEEE 7003 requires that Clause 4 is done again, in iterations, during all of the AIS life cycle, as “appropriate” (1.6). See Requirements for bias consideration (Clause 4).

How does IEEE 7003 identify stakeholders?

In IEEE 7003 Clause 6, the persons who do stakeholder identification identify the impacted stakeholders and the influencing stakeholders of the AIS. IEEE 7003 requires that the attributes of all stakeholders are identified and used to make the groups against which bias is examined and mitigated (6.1). It requires that, as a minimum, stakeholder identification includes all the steps of 6.4 a to j.

IEEE 7003 also requires that the business case gives a first record of the stakeholders and, “explicitly or implicitly”, their initial “ranking” (6.2 a). When protected attributes are identified, it requires “consideration” of the biases, diversity, cultural context, and “competency” of the persons who do stakeholder identification (6.5). IEEE 7003 also requires that the stakeholder attributes are examined again, and changed if necessary, at each subsequent “stage” (6.1). See Stakeholder identification (Clause 6).

What data metadata does IEEE 7003 Clause 7 require?

IEEE 7003 Clause 7 requires that the data type or types of each feature are recorded, with their “reasons” (7.4.2). It also requires that the output of the metadata steps is recorded as metadata (7.4.2). Clause 7.4.2.1 sets a step to “consider, document and verify” the conditions in which each dataset is collected or made, and gives these conditions only as examples. It also sets four more steps, for example to identify possible proxies.

IEEE 7003 requires a record of the “reasons” for each decision to include, remove, or not include data and attributes (7.4.1). It also requires that these “reasons” are added to the bias profile. This applies while the data is collected, when tests are done on the data, while the data is used, and when the data is decommissioned. See Data sources, types, and metadata (Clause 7).

What is data-to-stakeholder mapping in IEEE 7003?

In IEEE 7003, data mapping maps the data against the attributes of the identified stakeholders (7.4.2.3). IEEE 7003 requires that the mapping makes it possible to compare and examine the identified stakeholders. It also requires that each imbalance between the attributes of stakeholders in the data is recorded. If the AIS is retrained, IEEE 7003 requires that the mapping is done again (7.4.2.3 b).

Before the AIS is “implemented”, IEEE 7003 also requires data exploration for possible sources of bias (7.4.2.2). IEEE 7003 also requires that the data is monitored continuously, as in Clause 9, and that the bias profile is changed with the results (7.4.2.3). See Data exploration and data mapping (Clause 7).

What does the IEEE 7003 risk and impact assessment require?

IEEE 7003 Clause 8 requires an initial bias risk and impact assessment (RIA) at system inception (8.1). It requires that the assessment starts when the business case is made, with risk and impact metrics and a “rationale” for them (8.4 a). The RIA is examined again and changed at each subsequent “stage”, after the evaluation of bias (8.1). IEEE 7003 also requires this when the bias profile, the AIS, the context of use, or the environment of the AIS changes (8.4 b).

IEEE 7003 also requires two risk inventories (8.4 d, e). It requires that the assessment and each subsequent change to it are signed, with accountability to the position that controls the AIS during its life cycle (8.4 f). When the RIA finds that risks are not removed or mitigated, IEEE 7003 also requires that the “relevant” accountable entity accepts accountability and responsibility for possible legal, regulatory, or other “harmful consequences” (8.4 f). The owner of the AIS gives formal acceptance of the assessment. After a change of ownership or environment, IEEE 7003 requires a full reassessment and its formal acceptance (8.5). See Risk and impact assessment (Clause 8).

What is the IEEE 7003 design and output bias evaluation?

IEEE 7003 Clause 9.2 gives a procedure to examine the design and outputs of an AIS for bias. IEEE 7003 requires that this procedure is part of the “design and development” steps, where more than one evaluation is permitted (9.2.1). The procedure uses risk-based testing, a metric with reference values, and a bias evaluation method with a “rationale” (9.2.3 b, 9.2.4.3). It has a step to recommend mitigations (9.2.4.3 d), but IEEE 7003 does not give the mitigations (5.2).

IEEE 7003 also requires that the pre-processing methods are examined for bias, and that “similar” procedures are followed when a model is retrained and changed (9.2.4.2). It requires a UI design that “reflects” the attributes of all the identified stakeholders (9.2.4.4). The evaluation document has nine items, two of them only where applicable (9.2.3 b). See Evaluation of the AIS Design and Outputs (Clause 9.2).

How does IEEE 7003 catch bias drift over time?

With ongoing evaluation: IEEE 7003 Clause 9.3 requires that continuous monitoring of an AIS helps to find drift during its operation, and also during its decommissioning (9.3.1). It is possible that the monitoring does not find all drift (9.3.1, 1.9). The procedure has a list of items to monitor, which 9.3.4.1 gives only as examples, and a recorded “rationale” for its recurrence (9.3.4.1).

IEEE 7003 requires that each iteration checks and records each type of drift that can cause bias, with the “plans” for each, “if any” (9.3.4.2). After mitigation steps, it requires one more evaluation cycle, as “appropriate” (9.3.4.1). Clause 9.3.4.2 c adds a “plan” for other drift, and monitoring if there is “evidence” of the “reinforcement” of external real-world bias. See Ongoing evaluation and drift monitoring (Clause 9.3).

In what order should an organization roll IEEE 7003 out?

Modulos recommends this sequence for the first iteration of IEEE 7003. First, add OFF-25 to the organization AI System and MFF-25 to the AI System of each AIS. Then collect the governance processes and policies that are available (4.2 c). For each AIS, do the Clause 4 work, and start the risk and impact assessment (RIA) when you make the business case (8.4 a). Then do stakeholder identification, data representation, and evaluation. The evaluation of the design and outputs is part of the “design and development” steps (9.2.1).

IEEE 7003 requires that each clause is applied again during all of the AIS life cycle, as “appropriate” (1.6). It also requires that the RIA is examined again and changed when the bias profile, the AIS, the context of use, or the environment of the AIS changes (8.4 b). See Recommended sequence of work.