Principles

MAS FEAT gives 14 Principles for the use of AIDA in decisions about financial products and services. The Veritas Documents give an optional method for assessment of an AIDA system against these Principles.

AIDA and data subjects

In MAS FEAT § 3.1, the Principles "can guide all firms using AIDA to provide financial products and services". The name in MAS FEAT § 3.3 for these organizations is AIDA Firms. The Principles are "not intended to be prescriptive" and are "not intended to replace existing relevant internal governance frameworks", and AIDA Firms "should continue to comply with all applicable laws and requirements" (MAS FEAT §§ 2.1–2.2). For more about MAS FEAT, see About MAS FEAT.

MAS FEAT § 3.3 gives this information:

Name	MAS FEAT § 3.3
AIDA	“artificial intelligence or data analytics, which are defined as technologies that assist or replace human decision-making”
AIDA Firms	“firms that use AIDA to provide financial products and services”
AIDA-driven decisions	“decisions and decision outcomes made or supported by AIDA”
AIDA-driven decision-making	“the process of making decisions using AIDA”
Data subjects	“individuals and/or firms who are or may be affected by AIDA-driven decisions”
Models	“the specific AIDA algorithms and algorithmic processes used to arrive at AIDA-driven decisions”

MAS FEAT §§ 3.2–3.3 give this information about where the Principles apply:

• AIDA includes data analytics, not only artificial intelligence. In MAS FEAT § 3.2, “this definition is broad”, and the Principles "can apply to both complex and straightforward AIDA techniques".
• AIDA-driven decisions include decisions that AIDA helps to make, not only decisions that AIDA makes.
• Data subjects include “individuals and/or firms” who “are or may be affected by AIDA-driven decisions”.

Materiality

MAS FEAT § 2.3 lets AIDA Firms use materiality for all of the Principles:

Across all of the Principles set out in this document, firms can calibrate actions and requirements under their internal governance framework based on the materiality of the AIDA-driven decisions.

An AIDA Firm can use materiality to “calibrate” its internal governance “actions and requirements”. Materiality is not a condition for MAS FEAT to apply. In MAS FEAT § 3.2, the Principles "can apply to both complex and straightforward AIDA techniques", and AIDA Firms "are encouraged to calibrate their internal governance frameworks based on considerations of materiality".

MAS FEAT § 2.3 gives "non-exhaustive and indicative examples of considerations that may be used when determining materiality":

• "Extent to which AIDA is used in decision-making"
• "Complexity of AIDA model"
• "Extent of automation of process of AIDA-driven decision-making"
• "Severity and probability of impact on different stakeholders, including individuals"
• "Monetary and financial impact"
• "Regulatory impact"
• "Options for recourse available"

MAS FEAT does not give a method to calculate materiality from these examples. In MAS FEAT § 2.4: "As the Principles relate to and reinforce each other, readers are encouraged to view the Principles in totality."

The 14 FEAT principles

MAS FEAT § 4 gives AIDA Firms 14 Principles in four groups: Fairness (Principles 1–4), Ethics (Principles 5–6), Accountability (Principles 7–11), and Transparency (Principles 12–14). MAS FEAT gives the Principles "for firms to consider", and its examples are "purely illustrative" (§§ 2.1–2.2). The Veritas codes G, F, EA, and T are not Principle numbers (see Veritas assessment).

Fairness (4 principles)

MAS FEAT § 5.3 puts the four Fairness Principles in two groups. The first group, justifiability, contains Principles 1 and 2:

• Principle 1: “Individuals or groups of individuals are not systematically disadvantaged through AIDA-driven decisions unless these decisions can be justified.”
• Principle 2: “Use of personal attributes as input factors for AIDA-driven decisions is justified.”

The other group, accuracy and bias, contains Principles 3 and 4:

• Principle 3: “Data and models used for AIDA-driven decisions are regularly reviewed and validated for accuracy and relevance, and to minimize unintentional bias.”
• Principle 4: “AIDA-driven decisions are regularly reviewed so that models behave as designed and intended.”

MAS FEAT does not identify which decisions or personal attributes are “justified”. In MAS FEAT § 5.2, AIDA Firms "may have internal governance frameworks in place to assess and set standards of appropriate levels of justifications behind AIDA-driven decisions, including use of specific data attributes". “Illustration 2” (§ 5.5) is about a decision "whether to offer retirement-related financial services and products". In that example, "the use of age as an input factor for this AIDA-driven decision may be justified, given the relationship between age and retirement".

MAS FEAT does not give an interval for the regular reviews of Principles 3 and 4. Two examples show how an AIDA Firm can set the interval:

• “Illustration 3” (§ 5.6): "The frequency of exercises or tests depends on the materiality of the decision and the complexity of the model." In this example, the AIDA Firm "requires more frequent validation and testing" for "particularly material and complex deep learning models". For "less material and more straightforward models, quarterly, annual or even less frequent validations and tests suffice".
• “Illustration 4” (§ 5.7): "In addition to periodic reviews", the internal governance framework identifies "specific triggers for review, including unexpected changes that invalidate basic model assumptions".

These examples do not tell AIDA Firms to test continuously or to do a review after each change to data or models.

Ethics (2 principles)

MAS FEAT § 6.1 tells AIDA Firms: "Adherence to these ethical standards applies equally to the use of AIDA." In § 6.1, an AIDA Firm can show its ethical standards "through many ways, including company values, codes of conduct and mission statements", and these standards "may vary across firms and geographies".

• Principle 5: “Use of AIDA is aligned with the firm’s ethical standards, values and codes of conduct.”
• Principle 6: “AIDA-driven decisions are held to at least the same ethical standards as human-driven decisions.”

“Illustration 5” (§ 6.2) gives an example of Principle 6. In this example, the AIDA Firm has "recruitment rules that specify that gender should not be a consideration in making hiring decisions". To keep "the same ethical standards as human driven decisions", the model of the AIDA Firm "does not use gender as an input or derived attribute".

Accountability (5 principles)

MAS FEAT § 7.3 puts the five Accountability Principles in two groups. The first group, internal accountability, "is concerned with the AIDA Firm’s internal governance" and contains Principles 7 to 9:

• Principle 7: “Use of AIDA in AIDA-driven decision-making is approved by an appropriate internal authority.”
• Principle 8: “Firms using AIDA are accountable for both internally developed and externally sourced AIDA models.”
• Principle 9: “Firms using AIDA proactively raise management and Board awareness of their use of AIDA.”

The other group, external accountability, "is concerned with the AIDA Firm’s responsibility to data subjects" and contains Principles 10 and 11:

• Principle 10: “Data subjects are provided with channels to enquire about, submit appeals for and request reviews of AIDA-driven decisions that affect them.”
• Principle 11: “Verified and relevant supplementary data provided by data subjects are taken into account when performing a review of AIDA-driven decisions.”

In MAS FEAT § 7.1, "clear responsibility for and ownership of AIDA-driven decisions" in the AIDA Firm is important. This accountability "applies to all uses of AIDA, whether internally developed or externally sourced".

Principle 7 does not identify the internal authority. In “Illustration 7” (§ 7.4), the AIDA Firm uses its "internal approval structures in place for decision-making" when it sets approval authorities for AIDA-driven decision-making. In “Illustration 8” (§ 7.5), "When setting approving authorities for AIDA-driven decisions, the Firm’s internal governance framework takes into account materiality and complexity of the decision." For "highly-material decisions", that AIDA Firm sets approval at "Chief Executive Officer or Board levels", "with proper due diligence so that approving authorities have sufficient understanding of the data and model logic used for decision-making".

In Principle 9, the AIDA Firm tells "management and Board" about its use of AIDA, not only about the risks of AIDA. In “Illustration 9” (§ 7.6), the AIDA Firm "updates the Board on the use of AIDA within the Firm on a periodic basis".

MAS FEAT § 7.2 tells AIDA Firms that AIDA-driven decisions, "especially those that could significantly affect data subjects", "need to be based on as accurate an understanding of the data subject as possible". In § 7.2, "Data subjects could provide accurate information about themselves through channels that aim to facilitate communication and seek recourse." In “Illustration 10” (§ 7.7), the AIDA Firm "includes a contact number and email address" when it tells a data subject about an AIDA-driven decision. MAS FEAT does not give a time limit for the AIDA Firm to supply information or to do a review.

Principle 11 has two conditions on supplementary data. The data is "verified and relevant", and data subjects supply it.

Transparency (3 principles)

MAS FEAT § 8.1 identifies two possible effects of transparency:

• Increased transparency "could improve public understanding and confidence in AIDA".
• In the other direction, "excessive transparency could create confusion or unintended opportunities for individuals to exploit or manipulate AIDA models".

MAS FEAT § 8.1 tells AIDA Firms: "It is important to balance these considerations in determining the appropriate level of transparency in the use of AIDA." In § 8.1, "the materiality of the decision is also relevant" to the level of transparency.

• Principle 12: “To increase public confidence, use of AIDA is proactively disclosed to data subjects as part of general communication.”
• Principle 13: “Data subjects are provided, upon request, clear explanations on what data is used to make AIDA-driven decisions about the data subject and how the data affects the decision.”
• Principle 14: “Data subjects are provided, upon request, clear explanations on the consequences that AIDA-driven decisions may have on them.”

In Principle 12, the use of AIDA is "proactively disclosed" in general communication. Principles 13 and 14 apply "upon request" of a data subject, and the explanations are about the decisions about that data subject.

MAS FEAT § 8.2 tells AIDA Firms: "Clear explanations do not necessitate exposure of intellectual property or publishing of proprietary source codes." The explanations "could instead focus on facilitating the data subject’s understanding of the Firms’ use of AIDA".

“Illustration 13” (§ 8.4) shows the balance of § 8.1 in two examples:

• For "AIDA models for fraud detection", "Given materiality considerations and concerns of model manipulation or exploitation", the AIDA Firm "decides not to share about the AIDA model used, or provide explanations relating to this area".
• For "an AI phone operator", the same AIDA Firm "opts to begin the conversation by informing the data subject that they are interacting with an AI phone operator".

“Illustration 13” shows the decision of one AIDA Firm. MAS FEAT does not make this decision for all "AIDA models for fraud detection".

Veritas assessment

Financial services institutions (FSIs) can use and adapt the Veritas assessment method for optional assessment of AIDA systems against the MAS FEAT Principles.

In Veritas Document 3 “Appendix A”, each assessment item has a code. G is the generic checklist, F is the fairness checklist, EA is the ethics and accountability checklist, and T is the transparency checklist. These codes are not MAS FEAT Principle numbers.

Groups in a Veritas assessment

Veritas Document 3 § 1.3 identifies four groups in an AIDA system assessment:

Group	Function in Veritas Document 3 § 1.3
AIDA system developer	“the party that develops the AIDA system under study”, for example an FSI, “a vendor or a Fintech firm”.
AIDA system owner	Uses the AIDA system “to help make decisions for fulfilling their business goals”, for example “the business team in the FSI that is responsible for the AIDA system”.
AIDA system assessor	“assesses the FEAT assessment presented to them by FSI”, for example an “external audit firm”, or a model risk management group or internal audit group in the FSI.
FSI	“the organisation responsible for the AIDA system under study”. “The AIDA system owner sits within the FSI and is acting on its behalf.”

The AIDA system developer "may be part of the FSI if the system is developed in house or may be an organisation external to it if the system has been procured". The AIDA system assessor "may sit within the FSI (for an internal assessment) or be a separate body (for an external assessment)".

The optional assessment

Veritas Document 3 “Appendix A” tells FSIs that the “checklist” "is optional and is only relevant if an FSI chooses to take steps to adhere to FEAT". “Appendix A” also tells FSIs: "If FSIs decide to follow FEAT principles, the evidence to this checklist should be available on demand if required."

Veritas Document 3 § 3.2.3 recommends that FSIs adapt the “assessment process” and its “level of detail” to the AIDA system’s risk level. For "higher risk use cases", the examples include "more extensive analysis", "more frequent and in depth monitoring processes", and "more stringent review and approval processes". In these examples, "higher risk use cases may benefit from independent validation through third party audits". For "lower risk use cases", internal audits can be sufficient "where in house expertise is available".

MAS FEAT does not give a limit for a fairness metric, and it does not identify an audit of bias as a necessary step.

The sequence of the assessment

Veritas Document 3 recommends this sequence in its “Executive Summary”:

It is recommended to start with Ethics and Accountability assessment which helps to define the guiding principles at an organizational level, then perform the Fairness and Transparency assessments concurrently within the various steps of the AIDA system development lifecycle, as they are inter-related and are use case specific.

Veritas Document 3 “Appendix A” puts the assessment items in five steps. Step 0, "Principles to Practice", is the "Foundational Setup" for all use cases. In Step 0, the FSI completes the EA assessment items first. The FSI completes Step 0 at the start, and Steps 1 to 4 for each use case. Veritas Document 3C § 3.1 recommends review of the T1–T5 transparency standards at intervals after Step 0.

Step	Assessment items	Assessment
Step 0, Principles to Practice	G1–G4	“Is AIDA defined in the FSI?” (G1), and a framework for “roles and responsibilities”. A materiality framework with clear materiality levels. An “up-to-date inventory” of all material AIDA use cases.
Step 0, Principles to Practice	F0	A "standard, robust process" to identify "at risk groups", "personal attributes and potential proxies", and "explicit fairness objectives and associated measures and thresholds". Also "algorithmic methods to identify and reduce unfairness" (Document 3, “Appendix A” § 1.2.1).
Step 0, Principles to Practice	EA1–EA4	Organizational values, principles “for the ethical use of AI”, “context-relevant core concepts”, and “Is the workforce trained on values-based decision making or is there a pilot planned?” (Veritas Document 3B “Figure 1.1”).
Step 1, Define System Context and Design	G5–G8	The commercial objectives and their “quantitative measures”, and the “scope boundaries” of the system. G7: “Is AIDA use justified for this particular objective/use case (based on limitations of alternatives, proof of AIDA being used successfully elsewhere, availability of adequate data to train the AIDA system, valid legal basis, etc.)?” The materiality of the use case with the Step 0 framework.
Step 1, Define System Context and Design	EA5–EA6	For each use case, “a statement of relevant values, core concepts, principles, commitments, and specifications”, with the commitment owners and “the stakeholders with interest in holding the commitment owners accountable”. Also “relative priorities, which may or may not be tied to materiality”, for each commitment.
Step 2, Prepare Input Data	G9–G10	The data of the system, for example its data provenance and informed consent, and for personal data, “has a DPIA been conducted?”. DPIA is a “Data Protection Impact Assessment” (Document 3 § 3.1). Also the data pre-processing and feature engineering of the system.
Step 3, Build and Validate	G11	“Is the composition of the AIDA system defined?”, and the performance estimates of the system and their uncertainties.
Step 4, Deploy and Monitor	G12–G13	A monitoring and review regime that can find unusual operation, for example “significant model and/or data drift”. Also “fallback and/or mitigation plans” for “triggers from the system’s monitoring and review regime”.

The fairness checklist, the transparency checklist, and the EA assessment items of Step 4 are in the sections below. For the sequence of work that Modulos recommends, see A sequence that works.

Fairness checklist items

Veritas Document 3 “Appendix A” uses different numbers for its fairness checklist, F0 to F12. Its F6–F8 are about performance estimates, trade-offs, and "justified" "fairness outcomes" (§ 1.2.2). For F1–F12 below, see Veritas Document 3A “Appendix 8.13”:

Assessment items	Assessment
F1–F3	The “individuals and groups” “at risk of being systematically disadvantaged by the system” (F1). The possible harms and benefits “that are relevant to the risk of systematically disadvantaging the individuals and groups in F1” (F2). The fairness objectives and fairness metrics of the system, for the “individuals and groups in F1” and the harms and benefits in F2 (F3).
F4–F5	“Key errors, biases or properties present in the data” that can have an effect on fairness, and “how are these impacts being mitigated”.
F6–F8	The personal attributes used for the operation or fairness assessment of the system (F6). “Does the process of identifying personal attributes take into account ethical objectives of the system, and the people identified as being at risk of disadvantage?” (F7). F8: “Have you assessed and documented for every personal attribute and potential proxy for a personal attribute, why is its inclusion justified given the system objectives, the data, and the quantified performance and fairness measures?”
F9–F11	The “quantitative estimates” of the performance of the system against its fairness objectives, and the uncertainties of these estimates. The “achievable trade-offs” between the fairness objectives and the commercial objectives. “Have you justified (vs alternative tradeoffs) and documented the final fairness outcome for the AIDA System?” (F11).
F12	If the monitoring and review regime keeps the effects of the system aligned with its commercial objectives and fairness objectives (G5 and F3).

The FSI selects the fairness objectives and fairness metrics for each system (F3). Veritas Document 3A § 4.1.2 recommends that FSIs adapt the fairness “assessment process” and its “level of detail” to the AIDA system’s risk level. For "higher risk systems", its example is a "detailed justification for the choice of fairness metrics, methods of measurement, and included attributes".

Recourse and commitments

Veritas Document 3B “Figure 1.1” gives two ethics and accountability checklist items for Step 4, Deploy and Monitor:

• EA7: "Are recourse mechanisms available to data subjects being used? Is there an explanation if the use of recourse is too high or too low?"
• EA8: "Are controls in place to revisit and recalibrate the commitments and their specifications to ensure they’re properly incentivizing the right decisions and accurately measuring progress toward the commitments?"

EA7 connects to the "channels" for data subjects in MAS FEAT Principle 10 (see Accountability).

Transparency checklist items

Veritas Document 3C §§ 3.1–3.4 gives the transparency checklist items T1 to T17:

Assessment items	Step	Assessment
T1–T5	Step 0 (§ 3.1)	The FSI identifies “the factors it will use to determine whether external transparency is essential for a particular AIDA use case”. “Where an FSI has chosen to provide external transparency”, the “proactive or reactive communication” for each part of “the FSI’s customer lifecycle”. The FSI identifies “the factors it will use to determine the extent of, and audience for, internal transparency”. An “acceptable set of AIDA ML explanation method(s)”, and “minimum accuracy standards for such explanation methods”.
T6–T10	Step 1 (§ 3.2)	The use case team uses T1 to check if external transparency is necessary (T6). If it is necessary, the use case team uses T2 to identify “the proactive and reactive communication needed at each stage of the customer lifecycle” and “the form of such customer facing communication” (T7). The use case team uses T3 to identify “the level of internal transparency needed, and the audiences for the same” (T8). The use case team selects an explanation method “from the approved list in T4” (T9). The use case team checks that the method “meets the minimum accuracy requirement for this specific use case” with T5 (T10).
T11–T13	Step 3 (§ 3.4), internal transparency	“Internal transparency dashboards/reports”. Review and approval of these outputs by “relevant first and second line control teams”, “including model validation where relevant”. Mitigation actions “where the explanations have not met first/second line expectations”.
T14–T16	Step 3 (§ 3.4), external transparency	System functions for the agreed external transparency. Changes to customer service and complaint handling. Also “Have relevant staff been provided with appropriate training to address customer queries?”. Changes to the terms and conditions “to explain how data is used, how benefits and risks to individuals are associated with the processing, and how individuals may participate and object where appropriate”.
T17	Step 3 (§ 3.4), “Common to Both”	If the “AIDA system implementation” continues to “support the agreed internal and external explanations” after go-live, “throughout the lifetime of the AIDA system”.

The explanation methods and their accuracy standards can change with the type of use case and with materiality (Veritas Document 3C § 3.1). Veritas Document 3C § 3.2 recommends that the use case team records each decision and its “justification”.

Related information

For the MAS FEAT Frameworks in Modulos, their Requirements, and the steps that Modulos recommends, see Use MAS FEAT in Modulos. For the FAQ, see MAS FEAT FAQ. For the MAS Guidelines on Artificial Intelligence Risk Management for Financial Institutions, see MAS AI Risk Management Guidelines.

Sources

• MAS FEAT — Monetary Authority of Singapore, “Principles to Promote Fairness, Ethics, Accountability and Transparency (FEAT) in the Use of Artificial Intelligence and Data Analytics in Singapore’s Financial Sector”. Sections 2 to 8. MAS.
• Veritas Document 3 — Veritas Consortium, FEAT Principles Assessment Methodology (“Veritas Phase 2”, 2022-02). “Executive Summary”, sections 1.3, 3.1 and 3.2.3, and “Appendix A”. MAS Veritas Initiative.
• Veritas Document 3A — Veritas Consortium, “FEAT Fairness Principles Assessment Methodology” (“Veritas Phase 2”, 2022-02). Section 4.1.2 and “Appendix 8.13”. MAS Veritas Initiative.
• Veritas Document 3B — Veritas Consortium, “FEAT Ethics and Accountability Principles Assessment Methodology” (“Veritas Phase 2”, 2022-02). Section 1.1 and “Figure 1.1”. MAS Veritas Initiative.
• Veritas Document 3C — Veritas Consortium, “FEAT Transparency Principles Assessment Methodology” (“Veritas Phase 2”, 2022-02). Sections 3.1 to 3.4. MAS Veritas Initiative.

Disclaimer

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