IEEE 7003

IEEE Std 7003-2024, IEEE Standard for Algorithmic Bias Considerations, is a standard of the IEEE. 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". The same page gives access to the standard through the IEEE GET Program, and it gives this information: "The PDF of this standard is available at no cost". Use of the standard is fully voluntary (Important Notices and Disclaimers).

IEEE 7003 applies to algorithmic systems that can give different outcomes to different parties (Where IEEE 7003 applies). In 3.1, an algorithmic system gives an output decision from the automated processing of data. Most clauses apply to an autonomous intelligent system (AIS). For AIS, Clause 3.1 refers to the IEEE Standards Dictionary Online. The clauses are in a sequence around the bias profile, the record of bias consideration for an AIS (The bias profile).

Modulos holds IEEE 7003 as two Frameworks, MFF-25 for each AI System and OFF-25 for the Organization (How Modulos models the standard).

Where to start

Find your task below. Then open the IEEE 7003 page or section for that task.

• If you make, change, or decommission an AIS, start with Clause 4.
• If you are a procurer or a supplier, read Conformance.
• If you also use the EU AI Act, ISO/IEC 42001, or the NIST AI RMF, read How IEEE 7003 compares to other frameworks.
• If you set up IEEE 7003 in Modulos, read Operationalizing IEEE 7003 in Modulos.

Wanted and unwanted bias

In IEEE 7003, bias is a systematic difference in the treatment of some objects, persons, or groups, when compared with others (3.1).

IEEE 7003 does not make all bias a problem. Some bias is wanted, because it is necessary for the functional requirements of the AIS. It is also permitted to add wanted bias to the AIS to correct unwanted bias (5.1). Clause 4 starts bias consideration with two types of bias (4.1):

• wanted bias, which helps the AIS with its "functional objectives"
• unwanted bias, which can have a bad effect on the "functional objectives", or which can cause damage to stakeholders.

Clause 1.4 does not "deny" that operationally-justified bias has a "role" as a "fundamental" part of information classification and of decisions. The standard helps to identify how wanted bias and unwanted bias are different, and to tell other persons about this difference. Thus it helps to make clear the limits of "appropriate" use of the algorithmic system (1.4).

The Introduction of IEEE 7003 is not part of the standard. Its example of wanted bias is a "search engine": a good result has a bias that agrees with the "search-term" of the user.

The standard is for persons or organizations that make an AIS (1.2). It helps them to identify, measure, and tell users and regulatory authorities how the AIS uses bias. It also helps them to show that they used "best practices" in the design, the "development", the testing, and the evaluation of the AIS. The function of the "best practices" is to prevent unwanted differential impact on stakeholders (1.2).

For wanted and unwanted bias, the normative clauses require steps to examine the bias intent, and steps to mitigate unwanted bias (1.6). Clause 1.9 gives the limits of these steps (Conformance).

Metrics and the context of use

IEEE 7003 does not give the decision criteria or differential impacts that identify unwanted bias. This is because the attributes that make bias unwanted can change with the "social and societal context" in which the algorithmic system is used. Thus the standard gives methods that can be adjusted to each context of use (1.4).

But the standard also sets steps that include a "rationale". These are examples:

• IEEE 7003 requires that the risk and impact assessment starts when the business case is made. It also requires that the risk and impact metrics are identified with a "rationale" (8.4 a). See Clause 8.
• IEEE 7003 requires that those who do stakeholder identification record, with a "rationale", the procedure that they set and their decisions on each step (6.4 a). See Clause 6.
• Clause 9.2.4.3 b sets the selection of a metric with the testing scenarios and reference values. It also sets steps to "define and justify" the evaluation method, and to "define and justify" how to "interpret" its results. See Clause 9.2.
• Clause 9.3.4.1 sets a "rationale" for the recurrence of the ongoing evaluation, with steps to "define and document" it. See Clause 9.3.

Clause 5.2.5 has an item on the "thresholds" that can be set, but IEEE 7003 sets no "threshold". Annex D, Measuring Bias, is informative.

The bias profile: the through-life record

In IEEE 7003, the bias profile (Clause 5) is the repository of information that the tasks of bias consideration make and keep (1.6, 3.1). For each task, it holds the "initial", the "intermediate", and the "most recent versions" of the "documents" of that task (5.2). For more about the bias profile, read The bias profile.

The clauses of the standard

IEEE 7003 has nine clauses and six annexes. The normative Clauses 4 to 9 apply to points in the AIS life cycle (1.6).

The clauses are in the sequence in which they are usually used for the first time when an AIS is made. But IEEE 7003 requires that each clause is used again, iteratively, during all of the AIS life cycle, as applicable to how the AIS is made, how it operates, and how its conditions change (1.6).

Clause 5.2 gives five steps: requirements (Clause 4), stakeholder identification (Clause 6), data representation (Clause 7), system evaluation (Clause 9), and risk and impact assessment (Clause 8). The bias profile is not a step. It is the record that all steps use.

Clauses 4, 6, 7, 8, 9.2, and 9.3 each have five parts: the Purpose, the Inputs, the Outputs, the Actions (the tasks of the clause), and the Outcomes (for example 4.1 to 4.5). Each step after Clause 4 uses the bias profile as an input and gives an updated bias profile as an output.

Clause	Step	Outputs	Page
Clause 4	Requirements for Bias Consideration	The bias profile of the AIS, with the bias requirements and the boundaries of acceptability. A values statement for the AIS, with bias considerations (4.3).	Clause 4
Clause 6	Stakeholder Identification	An updated bias profile with five entries (6.3).	Clause 6
Clause 7	Data Representation	The data mapping, the metadata, and an updated bias profile (7.3).	Clause 7
Clause 8	Risk and Impact Assessment	An updated bias profile (8.3 a). A bias risk and impact assessment, initial or changed (8.3 b), that includes two risk inventories (8.4 d, e).	Clause 8
9.2	Evaluation of the AIS Design and Outputs	An updated bias profile and an evaluation document (9.2.3).	Clause 9.2
9.3	Ongoing evaluation	An updated bias profile and a changed evaluation document (9.3.3).	Clause 9.3

Some parts of IEEE 7003 are not normative:

• The Introduction is not part of the standard.
• NOTEs give only information and contain no requirements (footnote 10 to Clause 3).
• Annexes A to F are informative: A Conceptualizing Bias, B Stakeholder Identification, C Guidance for Risk and Impact Assessment, D Measuring Bias, E Cultural Aspects, and F Bibliography.
• Clause 2 gives no "normative references".

Thus only Clauses 1 to 9, but not their NOTEs, can set a requirement.

Where IEEE 7003 applies

IEEE 7003 applies to algorithmic systems, not to organizations. It applies to each algorithmic system that has a part in "selection, allocation, ranking, decision-making". It also applies to each algorithmic system that is part of a procedure in which some parties can get different outcomes from other parties (1.4). The computational approach of the system does not change this. The standard applies to rule-based, statistical, machine learning, and other computational approaches (1.4).

IEEE 7003 applies when one of these conditions occurs (1.4):

• The design of a new algorithmic system is made.
• A system is used with data that is different from its training data.
• A system is deployed in a new context of use.
• A system is changed.
• A system is decommissioned.

Clause 1.5 gives the limits of the scope. The scope includes methods and procedures that help designers and developers to make an AIS in which bias, when it is necessary, is wanted. The scope does not include systems or procedures that find and examine "problematic" algorithms while they are in use. Tracking and certification are not in the scope of the standard (1.5).

Two more items are not in the scope: the specification of mitigations (5.2, see The bias profile) and the computational representation of data (7.1, see Clause 7).

The "audience" of the standard includes developers, maintainers, operators, users, managers, and system auditors. It also includes system, software, and hardware suppliers, and procurers. These persons can be in supplier or procurer organizations, or in an organization that makes the AIS for use in that organization. The "audience" also includes data scientists, AI researchers, and their research organizations (1.7).

Conformance

Use of IEEE 7003 is fully voluntary (Important Notices and Disclaimers). IEEE 7003 is a standard, not a law.

Clause 1.3 tells how to read "shall", "should", "may", and "can":

• "Shall" shows a mandatory requirement for conformance, with no "deviation" permitted.
• "Should" shows that the standard recommends a step, but does not make it necessary.
• "May" shows that a step is permitted in the limits of the standard.
• "Can" shows that something is possible, or that a person or an object is able to do something.

Footnote 6 to Clause 1.3 does not let the standard use "must" for a mandatory requirement. Footnote 7 lets the standard use "will" only for facts.

Clause 1.8 gives three uses of the standard:

• A project or organization that claims conformance to IEEE 7003 can use the standard as its conformance document.
• A procurer and a supplier are permitted to include the standard in a contract or in "similar agreements", or to refer to it there. The condition is that the two parties agree that the supplier supplies systems and "services" "in accordance with" the standard.
• A project or organization is permitted to use the standard as an in-house standard. The condition is its decision to get information for users from a different part of the organization, "in accordance with" IEEE 7003.

To claim conformance, it is not necessary to use the nomenclature of IEEE 7003 for parts of the information for users, such as pages (1.8).

Clause 1.3 makes "shall" the mark of the requirements for conformance, and Clauses 4 to 9 set the steps (1.6). Thus conformance requires each "shall" of these clauses, not only one step. These are examples:

• a "rationale" for the stakeholder identification procedure and its decisions (6.4 a, see Clause 6)
• the data mapping again after retraining (7.4.2.3 b, see Clause 7)
• the risk and impact assessment, signed with accountability to the position that controls the AIS (8.4 f, see Clause 8)
• the formal acceptance of the assessment by the owner of the AIS (8.5, see Clause 8)
• a UI design that "reflects" the attributes of all identified stakeholders (9.2.4.4, see Clause 9.2).

IEEE 7003 does not include tracking or certification (1.5). A project or organization claims conformance. A record in Modulos, a Control, or a Requirement status does not make an AIS conform.

Clause 1.9 is the disclaimer of the standard. The standard sets "minimum criteria" that help to make sure that an AIS does not show unwanted bias. But the use of the "criteria" does not automatically make sure that the AIS conforms to its "system or mission objectives". It also does not automatically prevent "adverse consequences". Conformance does not release a party from its "social", "ethical", "moral", "financial", or legal obligations (1.9).

How Modulos models the standard

Modulos holds IEEE 7003 as two Frameworks with the name IEEE 7003: MFF-25 at the AI System level and OFF-25 at the Organization level. IEEE 7003 has no organization level, and its Clause 4 steps are for each AIS (1.6 a). OFF-25 holds the provisions that Modulos also reads at the Organization level. The page Operationalizing IEEE 7003 in Modulos gives the structure, the Requirement for each group of clauses, and the steps that Modulos recommends.

How IEEE 7003 compares to other frameworks

IEEE 7003 is a voluntary standard. Of the three instruments below, only the EU AI Act is law, and conformance to IEEE 7003 is not conformity with one of them.

EU AI Act. Regulation (EU) 2024/1689 is EU law. 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, especially where data outputs influence inputs for future operations" for the data sets of high-risk AI systems. Point (g) adds "appropriate measures to detect, prevent and mitigate possible biases identified according to point (f)". For high-risk AI systems that use "techniques involving the training of AI models", these rules apply to the "training, validation and testing data sets" (Article 10(1)). For other high-risk AI systems, they apply "only to the testing data sets" (Article 10(6)).

Article 15(4) has a rule for high-risk AI systems "that continue to learn after being placed on the market or put into service". These systems "shall be developed in such a way as to eliminate or reduce as far as possible the risk of possibly biased outputs influencing input for future operations (feedback loops), and as to ensure that any such feedback loops are duly addressed with appropriate mitigation measures".

Articles 10 and 15 are in Chapter III, Section 2. Article 16, point (a), tells "Providers of high-risk AI systems" to "ensure that their high-risk AI systems are compliant with the requirements set out in Section 2". Chapter III, Sections 2 and 3, apply from 2 December 2027 for Annex III systems and from 2 August 2028 for Annex I systems (Article 113, point (c)). For high-risk AI systems, Article 40(1) gives a presumption of conformity with the requirements of Section 2 where a system is in conformity with "harmonised standards or parts thereof the references of which have been published in the Official Journal of the European Union", "to the extent that those standards cover those requirements or obligations". Conformance to IEEE 7003 does not give this presumption of conformity. Regulation (EU) 2026/1744 (Digital Omnibus on AI) added Article 4a, "Processing of special categories of personal data for bias detection and correction", and removed Article 10(5).

ISO/IEC 42001:2023 is a voluntary standard for an AI management system in an organization. It requires the organization to "define a process" to examine the possible effects of AI systems on persons, groups of persons, and "societies" (6.1.4). The organization does this assessment at "planned intervals", and also when it is "proposed" that "significant changes" occur (8.4). Its Annex A gives "reference controls", for example A.5 (impact assessment of AI systems) and A.7 (data for AI systems), which the organization compares with its necessary controls (6.1.3 b).

NIST AI 100-1 (AI RMF 1.0, January 2023) is guidance: "The Framework is intended to be voluntary". Section 3.7, "Fair – with Harmful Bias Managed", gives "three major categories of AI bias to be considered and managed: systemic, computational and statistical, and human-cognitive", and the Core has four functions, "GOVERN", "MAP", "MEASURE", and "MANAGE" (Section 5).

The page Control overlap shows which Controls of the IEEE 7003 Frameworks are also in other Frameworks in Modulos.

Where to go next

Open one of these pages for the clauses of IEEE 7003, or for the use of IEEE 7003 in Modulos.

Bias requirements and the bias profile

Clauses 4 and 5: requirements for bias consideration and the bias profile

Stakeholders and data representation

Clauses 6 and 7: stakeholder identification and data representation

Risk, evaluation, and monitoring

Clauses 8 and 9: risk and impact assessment, and evaluation

Operationalizing in Modulos

The Frameworks, the Requirements, and a sequence for Modulos

FAQ about IEEE 7003

For short entries about IEEE 7003 that refer to the full rules, see the FAQ page.

See it in Modulos

In Modulos, use OFF-25 for your Organization and MFF-25 for each AI System to hold your IEEE 7003 work.

Request a demo

Sources of this page

• IEEE Std 7003-2024, IEEE Standard for Algorithmic Bias Considerations, approved 11 December 2024 and published 24 January 2025 (IEEE SA page). Provisions used:
  • the Important Notices and Disclaimers, the Introduction, and Annexes A to F
  • Clauses 1.2 to 1.9, Clause 2, and 3.1
  • 4.1, 4.3, 5.1, 5.2, and 5.2.5
  • 6.3, 6.4 a, 7.1, 7.3, and 7.4.2.3 b
  • 8.3, 8.4 a, d, e, and f, and 8.5
  • 9.2.3, 9.2.4.3 b, 9.2.4.4, 9.3.3, and 9.3.4.1.
• The IEEE SA page for IEEE 7003-2024, for the status, the dates, and the access through the IEEE GET Program: IEEE SA.
• EU AI Act, Regulation (EU) 2024/1689, as of 27 July 2026, with the changes of Regulation (EU) 2026/1744: EUR-Lex. Articles 4a, 10(1), 10(2), 10(6), 15(4), 16(a), 40(1), and 113 (point (c)).
• Regulation (EU) 2026/1744 (Digital Omnibus on AI), for the changes to Articles 4a and 10 of the EU AI Act: EUR-Lex.
• ISO/IEC 42001:2023, 6.1.3, 6.1.4, 8.4, and Annex A.5 and A.7: ISO.
• NIST AI 100-1, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023, Sections 3.7 and 5: NIST.

Disclaimer

This page gives general information and is not legal advice. Use of IEEE 7003 is voluntary. But a procurer and a supplier can make it a condition of their contract (1.8). Conformance does not release a party from its legal or other obligations (1.9).