Docs

Appearance

Stakeholders and data representation ​

IEEE Std 7003-2024 (IEEE 7003) gives two steps of bias consideration for an autonomous intelligent system (AIS). These steps are stakeholder identification (Clause 6) and data representation (Clause 7). Stakeholder identification finds the impacted stakeholders and the influencing stakeholders of the AIS, and their attributes (6.1). Data representation examines how fully the data of the AIS shows the attributes of the impacted stakeholders (7.1).

IEEE 7003 requires that each of its clauses is used again during the AIS life cycle, as applicable to how the AIS is made, how it operates, and how its conditions change (1.6). See The clauses of the standard. For the Modulos Requirement of each of these clauses, see The Requirements on the page Operationalizing IEEE 7003 in Modulos.

Stakeholder identification (Clause 6) ​

IEEE 7003 Clause 6 gives the tasks of the persons who do stakeholder identification for an AIS. Clause 6 helps to make sure that all related stakeholders are identified (6.1). These are the stakeholders on which bias has an effect, in the AIS and in its context of use. They are also the stakeholders that have an effect on this bias. IEEE 7003 requires that the attributes of all stakeholders are identified. It also requires that these attributes are used to make the groups for which bias is examined and mitigated (6.1).

This step gives the initial set of stakeholders as its output (6.1). At each subsequent step of bias consideration, IEEE 7003 requires that the stakeholder attributes are examined again and possibly changed (6.1). These subsequent steps are data representation, evaluation, and risk and impact assessment. Clause 1.6 c gives this information: stakeholder identification is for identification and not for evaluation.

The Definitions of Clause 3.1 ​

IEEE 7003 Clause 3.1 gives these entries:

  • Stakeholder — a person or other item on which the AIS has, or can have, an important effect. A person or other item that has, or can have, an important effect on the AIS is also a stakeholder.
  • Impacted stakeholder — a stakeholder on which the outcomes of the AIS have an effect.
  • Influencing stakeholder — a stakeholder that has an effect on the AIS in one or more of its parts or properties.
  • Group — a subset of objects in a “domain” that the same attributes connect together.
  • Protected attribute — an attribute that has protection by law. The jurisdiction of the context of use gives this protection (NOTE, which gives information only).
  • Sensitive attribute — an attribute that is a protected attribute, or that is not a protected attribute but has “societal or cultural sensitivity”. The context of use gives the result for each attribute: the attribute is sensitive, or it is not sensitive (NOTE).

Inputs ​

IEEE 7003 gives five inputs to stakeholder identification (6.2):

  • a) The business case for the AIS. IEEE 7003 requires that the business case contains a first record of the stakeholders, which is the input to stakeholder identification. It also requires that the business case gives the initial sequence of the stakeholders. The sequence shows how important each stakeholder is to the organization, or which stakeholders come first for it. The business case can give this sequence directly, or it can contain information that shows the sequence.
  • b) The technical requirements of the AIS.
  • c) The context of use that applies to the AIS.
  • d) Other information that is necessary for the persons who do stakeholder identification, for the tasks of Clause 6.
  • e) The bias profile.

Tasks (6.4) ​

IEEE 7003 requires that stakeholder identification includes all of these tasks, and it can include more tasks (6.4):

  • a) Record the procedure that the persons who do stakeholder identification make, and give the “rationale” for it. Record the decisions for each task of 6.4, and give the “rationale” for each decision.
  • b) Apply “diverse perspectives”. Get input from the influencing stakeholders and impacted stakeholders that were identified before, to identify more stakeholders and their attributes.
  • c) Identify the impacted stakeholders.
  • d) Identify the influencing stakeholders.
  • e) Identify the attributes of the identified stakeholders. Include the “inherent” attributes and the attributes that come from the context of use.
  • f) Identify the protected attributes. Record that this protection is necessary. Also record each other step that is done to help make sure that these attributes have correct protection.
  • g) Examine the identified attributes to find if they show bias.
  • h) Assess and change the bias considerations that the requirements for bias consideration first identified (Clause 4).
  • i) Measure the effect of the influencing stakeholders, and record it.
  • j) Use these measured effects in the risk and impact assessment.

Some of these tasks refer to informative Annex B for a method to do them.

Outputs and outcomes ​

IEEE 7003 requires that the information from the 6.4 tasks goes into the related section of the bias profile. The updated bias profile then has these five entries (6.3 a):

  1. The stakeholder reference set.
  2. The stakeholders in a sequence by how important they are.
  3. The type of each stakeholder: influencing stakeholder, impacted stakeholder, or the two.
  4. The attributes that each stakeholder has.
  5. The attributes that get protection, with the “rationale” for this protection.

IEEE 7003 Clause 6.5 gives these outcomes:

  • Clause 6 gives the tasks that make the stakeholder reference set, which contains all stakeholders and their attributes. IEEE 7003 requires that this set is used during all of the AIS life cycle. It gives information to the data representation, evaluation, and risk and impact assessment steps. The stakeholder reference set is part of the bias profile, and it is changed after the 6.4 tasks are complete.
  • IEEE 7003 requires that each stakeholder is identified as an impacted stakeholder, an influencing stakeholder, or the two.
  • IEEE 7003 requires that the attributes include the “inherent” attributes of a stakeholder. It also requires that they include the attributes that are related because of the context of use.
  • IEEE 7003 requires that the procedure finds which attributes are protected attributes. It is possible that no attribute is a protected attribute. Protected attributes can make special procedures necessary in subsequent steps of bias consideration.
  • When protected attributes are identified, IEEE 7003 requires that the biases, the “competency”, the diversity, and the cultural context of the persons who do stakeholder identification are thought about.

The persons who do stakeholder identification ​

IEEE 7003 Clause 5.2.2 gives two “goals” for stakeholder identification. The first is to identify the groups that can have an effect on the AIS. The other is to identify the groups on which the AIS can have an effect. Clause 5.2.2 also gives this information: to help get to these “goals”, it is necessary to think about these three items:

  • The “resourcing” of the persons who have the task of stakeholder identification.
  • Their representation in relation to the stakeholders that the business requirements identify.
  • The relation of the AIS to systemic biases and cultural biases.

Clause 1.4 also gives this information: a team with diversity can help to get a full bias consideration. Such a team can also help to decrease unwanted bias in the AIS during its life cycle. This is not a requirement.

Clause 4.4 b (external advocates) is a different task with a different condition. See Requirements for bias consideration (Clause 4).

Data sources, types, and metadata (Clause 7) ​

IEEE 7003 Clause 7 is the data representation step of bias consideration for an AIS. Clause 7 gives a procedure to record the types and the sources of the data (7.1). The procedure also includes the context of use, as the business case gives it. It also includes the representation of the identified stakeholders in the data.

IEEE 7003 requires that the data for use in the AIS is examined (7.1). This is to know the possible effect of the use of this data on the identified stakeholders. In Clause 7, data representation refers to how fully the data shows the attributes of the impacted stakeholders. Clause 7 does not include the computational representation of the data (7.1).

IEEE 7003 Clause 5.2.3 gives the task of data representation: to map the data of the system to the identified stakeholders. At this step, the primary task is to make sure that data for the system is available, or can be collected or made. IEEE 7003 requires that the sources of data are recorded (5.2.3). This record includes the data governance and how the data was collected. It also includes if the data has a use for the AIS, and which biases the data can cause in the AIS.

Inputs and outputs ​

IEEE 7003 gives three inputs to data representation: the business case, the data and metadata, and the stakeholder identification section of the bias profile (7.2). The outputs are the data mapping, the metadata, and the updated bias profile (7.3).

Decisions about data ​

IEEE 7003 requires that decisions about data are made to give data that is sufficiently representative of the identified stakeholders (7.4.1). IEEE 7003 requires that the “reasons” for each decision to include, remove, or not include data and attributes in the datasets are recorded (7.4.1). It 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. IEEE 7003 also requires that each “concern” and each “limitation” from the tasks of Clause 7 is recorded (7.4.1).

Metadata ​

IEEE 7003 requires that the output of the metadata tasks is recorded as metadata (7.4.2). It also requires that the data type or types of each feature in the dataset are recorded, with the “reasons” for the selection. IEEE 7003 requires that these “reasons” are “based on” the “anticipated” outcome of the AIS, the context of use, and the available data (7.4.2).

IEEE 7003 Clause 7.4.2.1 gives this task: think about, record, and check the conditions in which each dataset is collected or made. Clause 7.4.2.1 gives these conditions as examples:

  • a) The source of the dataset, and how the dataset came from that source.
  • b) If the dataset came from a different jurisdiction.
  • c) The “reason” for the selection of the source of the dataset.
  • d) The use that the dataset was first collected for.
  • e) How the dataset is collected, for example crowd-sourced, synthesized, scraped, in a set that does not change, or real-time.
  • f) If the supply of the dataset was voluntary.
  • g) If an opt-out was available.
  • h) If the data was collected in conditions where it can show human biases or “patterns of behavior” that can result from discrimination or from “assumptions” about groups.
  • i) If the dataset is “free” or “purchased”.
  • j) When the dataset was collected.
  • k) If the dataset is anonymized or pseudonymized.
  • l) If the features in the dataset are related to features in other datasets of the AIS. For example, the features can be “complementary”: each adds data that the other does not have. They can also be “corroborative” of each other. They can show the same set of persons or sets that do not overlap. The sets can have a different representation of sensitive attributes and non-sensitive attributes.

Clause 7.4.2.1 also gives four tasks:

  • m) Identify all possible proxies in the dataset or datasets.
  • n) For a synthetic dataset, examine the generating algorithm for bias that it can cause.
  • o) Find if the data has “fitness for purpose” for the business case.
  • p) Do a data quality assessment.

Data exploration and data mapping (Clause 7) ​

IEEE 7003 Clause 7 also includes data exploration and data mapping for an AIS. IEEE 7003 requires that the data for possible use in the AIS is examined for possible sources of bias, before the AIS is “implemented” (7.4.2.2). This data exploration includes all of these investigations, and it can include more:

  • a) Examine sensitive attributes and non-sensitive attributes, to find if they can be sources of bias.
  • b) Examine the correlation, the causation, or the two, between non-sensitive attributes and sensitive attributes.

The identification of possible proxies is a different task (7.4.2.1 m).

In IEEE 7003, the mapping in the data bias profile connects the data attributes with the impacted stakeholders (7.4.2.3). IEEE 7003 requires that data mapping makes it possible to compare and examine the identified stakeholders. For example, the attributes of the stakeholders can have different representation, or the sample size can be small. Clause 7.4.2.3 gives these three tasks:

  • a) Use the stakeholder identification to map the data against the stakeholder attributes.
  • b) If the AIS is retrained, IEEE 7003 requires that the mapping is done again.
  • c) IEEE 7003 requires that each imbalance between the attributes of stakeholders in the data is recorded.

IEEE 7003 also requires that the data is monitored continuously, as in Clause 9 (7.4.2.3). It also requires that the bias profile is changed to show the results (7.4.2.3). See Ongoing evaluation and drift monitoring (Clause 9.3).

IEEE 7003 Clause 7.5 gives the outcome of data representation: the users of IEEE 7003 know how the data represents the stakeholders. This information is necessary for the risk and impact assessment, for the evaluation of bias, and for help to mitigate unwanted bias.

Other frameworks ​

To compare IEEE 7003 with the EU AI Act, ISO/IEC 42001, and the NIST AI RMF, see How IEEE 7003 compares to other frameworks.

Where to go next ​

Bias requirements and the bias profile

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

Risk, evaluation, and monitoring

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

IEEE 7003 overview

The standard, where it applies, conformance, and other frameworks

Operationalizing in Modulos

The Frameworks, the Requirements, and a sequence for Modulos

Sources of this page ​

  • IEEE Std 7003-2024, IEEE Standard for Algorithmic Bias Considerations, approved 11 December 2024, published 24 January 2025 (IEEE SA page). This page uses these parts of IEEE 7003:
    • The front pages, and Clauses 1.4, 1.6, 1.8, 1.9, and 3.1.
    • Clauses 4.4 b, 5.2.2, and 5.2.3.
    • Clause 6 (6.1 to 6.5) and Clause 7 (7.1 to 7.5).

Disclaimer

This page gives general information only. It does not replace the help of a person who knows the law. The use of IEEE 7003 is voluntary. A procurer and a supplier can agree to include IEEE 7003 in their contract (1.8). Conformance to IEEE 7003 does not release a party from its obligations in law or its other obligations (1.9). Compare this page with IEEE Std 7003-2024.