Artificial Intelligence (AI) as a highly transformative technology take on a special role as both an enabler and a threat to UN Sustainable Development Goals (SDGs). AI Ethics and emerging high-level policy eforts stand at the pivot point between these outcomes but is barred from efect due the abstraction gap between high-level values and responsible action. In this paper the Responsible Norms (RAIN) framework is presented, bridging this gap thereby enabling efective high-level control of AI impact. With efective and operationalized AI Ethics, AI technologies can be directed towards global sustainable development.
Several recent and comprehensive reviews make clear
that there is a strong connection between large-scale
change and developments in Artificial Intelligence (AI)
[
far as to term this technological progression a “vector of
hope” [
or negative, of AI on large-scale sustainability and the SDGs is taken by AI Ethics. It is widely recognized that efective soft and hard policies on AI technologies are needed to ensure positive outcomes. Many attempts at high-level soft policy already exists by intergovernmental organisations, e.g. the European Commission’s “Guidelines for Trustworthy AI” (GTAI), but also by professional bodies, e.g. IEEE. Such policy documents focus The IJCAI-ECAI-22 Workshop on Artificial Intelligence Safety (AISafety
LGOBE 0000-0003-3113-2631 (M. Brännström); 0000-0001-9499-1535 (A. Theodorou); 0000-0001-7409-5813 (V. Dignum)
© 2022 Copyright 2022 for this paper by its authors. Use permitted under Creative Commons
values [
An often mentioned problem of the high-level guidelines is that they are at times too abstract to be applied to any particular case and and at other times too specific by mentioning problems which might not exist in a particular application. There is no particular level of abstraction that solves this problem for high-level policy as guidelines either become too abstract or too extensive.
technical domain typically consist of not a single actor,
the AI developer, but an interplay between developers,
procurers, customers and users [
play and successful bridging this abstraction gap between high-level policy and particular application is of central importance in establishing socially-beneficial AI.
The abstraction gap is not only a problem from a regulatory perspective. For the individual developer, procurer, or any other actor dealing with emerging AI applications where the gap severs the link between design and organisational choices on one hand and outcomes, ethical or otherwise, on the other. As there is no clear link between the particularities of an AI application and high-level ethical goals, there is no clear path forward even for actors on all levels who desire to act responsibly.
ment and analysis. This contribute to increase the divides and inequalities already present in society, decrease the transparency of AI Ethics itself and undermine trust in
ing towards the SDGs. Efects like these are even more
prominent in areas where both expertise and efective
governance structures with a strong ethical focus is
lacking; leading to an AI ethical void in the most sensitive into contextualised lower-level norms[
The solution to bridging the gap is context awareness text Z’, and has been well described formally in depth in
and structure, in policies, tools, assessment procedures, [
In this paper, a solution for bridging the abstraction represented using counts-as and modal logic [
The paper is structured as follows: First, a brief theoret- is a socio-technical system. ical background is provided. Then, RAIN is described in detail using examples (given in italics). Finally, the paper concludes with a discussion how RAIN aids the transition 3. RAIN towards the SDGs on all levels as well as directions for future work.
pared to VSD in the hierarchical breakdown process of
values in a way that makes the resulting norms
hier2. Background archy with its context-sensitive requirements reusable.
The RAIN norms hierarchy is also just as apt for
Value Sensitive Design (VSD) is a methodology for cen- questionnaire-type assessment questions as automatic
tering design around abstract high-level values, by em- tests or continual monitoring, and, thus, extending the
bedding values into the socio-technical context where use case of the Glass-Box to the socio-technical sphere.
they are being used, value-conflicts and key concrete de- A reusable norms hierarchy go a long way to reduce
sign requirements can be identified [
The VSD process can be made more structured using archy, then exploring the connection to socio-technical the count-as operator to break down high-level norms context, scoring mechanisms and, finally how to derive assessment results and projections of assessments on Given this scafold, particular a policy can be seen particular policies. as sub-context providing additional detail primarily to the abstract concepts HLV and issue by specifying sub3.1. Overview of the RAIN Pipeline classes. While policies are not typically written in such a structured manner we can use this scafold to frame the The RAIN framework can be seen as consisting of four content and see them as sets of statements about HLV fundamental components. The heart of the framework and sets of statements about issues. The following parts is the RAIN Graph. The RAIN Graph contains a struc- of this framework will help to extract detail so framed. tured and contextualized norms hierarchy. The section Policies also frequently mention particular technical feabelow will detail it’s methods of construction from an tures or stakeholders, if so these too are seen as content existing policy. Building around the Graph is a three of the policy. part pipeline starting with the context layer, which cap- Ex. GTAI presents several issues around the HLV Privacy tures the context features of a particular AI application which is expressed as consisting of Right to Privacy, Right in order to determine which contexts of the RAIN Graph to Data Protection, and Data Governance. Among issues which are active. Having established this, the assessment are use of personal data in training and use of transmission layer can be used to determine compliance to identified and storage of personal data, all of which violate subsets of norms. Finally, the results layer concerns aggregation Privacy and thus AI Ethics. Some of these concerns are not and extraction of results from the Graph and Assessment. in the provided assessment-questions but in the descriptive text. 3.2. Building the RAIN Graph
The RAIN Graph captures AI policy in a structured manner. In this section it will be described how such a Graph With the scafold for the policies in place, we can follow can be derived from a high-level policy but also expanded this with a new context 0 ⪯ forming the basis of the to particular contexts not explicitly mentioned in such rain framework. We can describe 0 as: abepdoelifinceyd. aHsiagnhy-LpeovleiclyP, ogluicidye(lHinLeP,s)twanidllairnd tohristhseeclitkioen 0 ∶ ⊑ (4) which primarily bases itself upon High-Level ethical Val- 0 ∶ -1 ⊑ (5) ues (HLV) and presents challenges to these from the use 0 ∶ -2 ⊑ -1 of AI technology, solutions to such challenges, or re- 0 ∶ -3 ⊑ -2 quirements on action to alleviate such challenges. These challenges, regardless of which form they appear will be In the practical applications of the RAIN framework, a termed AI Issues. graded scoring model of maturity levels is used: 1 implies
The framework description will be aided by standard that the system violate the high-level requirements to
Description Logics extended with the context scope ( ∶ , a minor degree and each other level indicate lesser
dewhere applies in context ), counts-as (⇒ ) operators grees of compliance with more serious violations. Each
and context relation ⪯ as described by [
(3)
A RAIN Graph, can be described to contain the following concepts • value A parsimonious sub concept of HLV. Often values in policy are expressed using several sub values or norms, these are here separated. We will term the set of all values ∈ as • stakeholder Reflecting a perspective of concern for a stakeholder group. We will term the set of all stakeholder s ∈
as . • socio-technical feature A socio-technical use of technology. We will term the set of all sociotechnical features ∈
as . • RAI norms and contexts A norm (, ) ∈ represent a particular challenge caused by some socio-technical feature ∈ to a value ∈ with respect to a stakeholder concern ∈ . We will term the set of all such norms as . Every such norm ∈
will be embedded in a sub-context ⪯ 0 such that ∶ ⊑ ( ⊓ ⊓ ).
The sets , , , ,
are considered to be holding semantically distinct items. For the algorithms 1 and 2, we define the operation
merge to mean an addition that preserves semantic distinctness. In the case of RAI norms, multiple distinct issues with corresponding assessment lists can have the same semantics but will be distinct if assessment criteria are taken into account. If so they occupy the same norms context as they are activated by the same features.
These norms can be seen as representing the junction between a value, a subject, and a circumstance. Or value, subject, and action. Through these norms, it is possible to determine what features of a given context which are related to which values and for whom. These relationships are the main purpose of the RAIN Graph and how it helps to bridge the abstraction gap.
Since each of the RAIN nodes identifies a particular threat, it can be accompanied with a corresponding set of requirements alleviating that threat. In this manner, a context-sensitive assessment of how a given AI application complies with one or several policies can be expressed as the degree of which it fulfills the requirements selected by its features. Some types of the technical requirements can be verified in an automated manner; in other words, the RAIN Graph fulfills the interpretation stage of the Glass Box by identifying in which ways it is relevant to monitor an application with regards to ethical concerns. Other requirements, concerning organisational features, design choices, or documentation, require a wider socio- intervention by stakeholders. Such requirements instead lend themselves to manual assessment procedures. Formally we can represent these RAI norms and their accompanying assessment rules as their
Data: ( , , , , begin
), a RAIN Graph for hlv ⊑ HLV ∈ P do
merge component values of hlv to if Explicit stakeholders ∈ P then merge component stakeholder concerns of policy into if Explicit socio-technical features ∈ P then merge component stakeholder concerns of policy into for ⊑ issue ∈ P do ← Values ⊂ impacted by issue ← Stakeholder concerns impacted by issue ← Socio-technical features which must be present for issue to threaten with regards to merge concerns into own contexts ⪯ 0 where represents the active presence of a stakeholder and feature instance in the context of the application. The general structure of this context also including the foundation of the assessment layer can be expressed as follows: ∶ ⇒ ( ∧ ∧ ) ∶ ∶ ∶ -1 ≡ ∃ -2 ≡ ∃ -3 ≡ ∃ -1. -2. -3. (6) (7)
into the graph. A second algorithm described here, the
of concern and expands the policy with consideration of a potentially new area of socio-technical context.
Ex. Privacy. 3. Consider the characterization of each of these high level values in order to break down each of these high-level concepts into singular areas of
norms form the values of the Graph with respect concern. The key here is to be parsimonious. One concern or concept per item. ⊏ Privacy ∈ GTAI Ex: Right to Privacy, Data protection, Data GovernanceData: Data: ( , , , ) begin Algorithm 2: RAIN Expansion Algorithm
, a RAIN Graph a set of new socio-technical features to this policy. 5. If Stakeholders or particular socio-technical features are explicitly mentioned in the policy, repeat the previous step for them as well in the same manner. Ex. End Users and Developer are mentioned stakeholders in GTAI. 6. Go through each Issue raised by this policy and state it in connection to at least one Value and at least one Stakeholder. In addition determine what socio-technical feature which must be in place for the issue to exist. It might be a way of dealing with data, a particular technology, or a particular use case, for example. Ex: Handling of the personal data of End Users is required for issues of GDPR in GTAI.
as one or more RAI norms which state that this
Ex. personal data ∶ ⇒ (Personal data ∧ End User ∧ Data Governance).
treated in this manner one can consider the content of the Graph to contain the explicit parts of the policy. However since most policy have selected some level of abstraction and scope, it is likely that many intended issues related to AI
merge for ( , , ) ∈ × ×
into to do
do for Issues which threaten with respect ← Values ⊂ impacted by Issue ← Stakeholder concerns impacted by Issue ← Socio-technical features which must be present for Issue to threaten with regards to merge norm ( , , ) into and merge concerns Graph can be extended to cover new particular contexts. cial case of interest is when policies have values which
ered to the Graph. Ex. For the example in the
next section, features common to home automation,
voice control and human interaction such as e.g.
Remote Processing and Passive Recording.
Sociotechnical features such as Vulnerable End Users are
also relevant to the elderly care example below.
2. For each intersection between a value, a
stakeholder concern and a socio-technical feature,
consider the possible ways in which the value is
challenged with regards to the stakeholder concern.
are defined diferently. That ethical values lack a
universal definition is a common mentioned problem [
manner it is possible to define a RAIN Graphs coverage.
Two things must pertain for the RAIN Graph to have relates to in this domain can be seen as a guide to which coverage of some particular area of AI Ethics. parts of the context that are relevant. This helps to reduce the otherwise nebulous concept of a context into • A RAIN Graph have coverage of a particular policy a more narrow form. With regards to the RAIN Graph, if merging it to the RAIN Graph using the RAIN the context is whether the features are present in the Decomposition Algorithm (Algorithm 1) would socio-technical sphere of the application or not. result in no change to the Graph. The socio-technical use case of a project and who the • A RAIN Graph have coverage of a particular area stakeholders are are necessarily intertwined. Similar of socio-technical context with respect to the to how merging additional policy into a single RAIN policy it covers if merging its Features to the Graph, representing the relationships between use cases, RAIN Graph using the RAIN Expansion Algo- Stakeholders and Features will also naturally overlap rithm (Algorithm 2) would result in no change to and converge creating a reusable structure helping with the Graph. knowledge elicitation and transfer.
Ex. The graph in the example have coverage for GTAI, Features described readily lend themselves to ontology voice recognition, home automation and human interaction. representation and dynamic questionnaires and dialogue The process can be repeated to add coverage for national approaches can be used to extract the details of an applisafety guidelines and the AI policy of local jurisdiction. cation without placing unduly high demands of expertise Even the particular policy of a procuring organisation can on the people characterizing the system. be added by using Algorithm 1 and 2. Ex. The features Remote Processing, Personal Data, Anthropomorphic Human Interaction, Language Dependence, Vulnerable End Users, and Hazardous Robotics (stove) are 3.3. The RAIN pipeline present in the example home automation system. End Users, For the purpose of assessment, the RAIN Graph can be Developers, Procurers and Auditors are relevant stakeholdembedded into a pipeline with the following three steps: ers. These identified in the context layer of the pipeline. 3.3.2. Assessment layer • Context layer capturing the socio-technical context and identifying stakeholders, top level values and policies. The output of this layer is context features and activated values. • Assessment layer providing context-specific testable requirements, satisfying the identified norms on a five-step scale of compliance. • Result layer aggregating the result of the individual norms onto the high-level values as well as projections upon compatible policies of choice.
Given that a certain set of features and stakeholders have been asserted by the context layer, some of the contexts of G will be active, and their statements will apply. The purpose of the assessment layer is to see if the rules (7) with regards to these contextualized norms have been violated or not. Each of these assessment statements are connected to an appropriate type of test (e.g quiz, monitoring, supplied evidence).
In this manner, no assessment is required in the cases
The pipeline can be part of an assessment process, an where the context does not apply thereby preventing a iterative development process or automatic monitoring bloat of irrelevant assessment questions. Every assessof policy compliance. ment test that do apply can be constructed towards a
For the rest of this section, a voice-controlled home- particular feature and stakeholder rather than towards automation system will be used as a running example. A the high-level goals or attempted generalisations. Bepublic-sector procurer is evaluating the compliance of cause each negative assessment result violate a particular said system against the GTAI before its purchase and use norm, and if this norm counts as the high-level value, in elderly care. This particular example is just one low- then the violations of the assessment rules will also be level interaction, but it such small interactions aggregate violations of the high-level norms they connect to. into the large scale societal efects towards or against the Ex: RAIN assessment find that Remote Processing is used SDGs. The example is given in italics. without a use-case reason (it is used to collect marketing data). Security measures surround handling of the stove 3.3.1. Context layer and context features and support exists for multiple languages. Anthropomorphic language is a Transparency concern especially due to the Vulnerable End User feature..
and 2, defined as the most general semantics of a feature which must apply in order for a particular norm to be challenged.
Given that the RAIN Graph have coverage in a particular socio-technical domain, the set of features the Graph
When assessments have been performed, the result can be evaluated in several ways. A straightforward way is to enumerate the Values in set and determine what level of violation and thus maturity score which applies to each Value. This would be a RAIN Graph-specific result. Another straightforward way is to look to the context of a particular policy and similarly enumerate its particular HLVs together with the aggregated maturity level. A less straightforward but highly efective way is to provide a set of statements on the contents of , where each statement maps to a particular requirement of an external assessment covered by . For instance if a RAIN Graph covers GTAI, a set of statements on the graph can map the results to each of the assessment questions in the guidelines. This way a particular high-level policy can be assessed in a context-aware manner even if the policy itself is not constructed for the RAIN framework.
Given the structure embedded in the RAIN graph, results can also be aggregated on particular stakeholders or socio-technical features, giving a valuable and detailed description on how ethical compliance is distributed over the socio-technical landscape of the Application.
Ex. While the system get high maturity levels on national safety standards, the aggregated GTAI scores are strongly violated due to the Remote Processing, especially with regards to Privacy. The local Procurers internal guidelines are also found violated and the system is rejected. The developer of the system could adapt for on-site processing of recorded data to gain a higher Privacy maturity level. Such adaption is a concrete technical and business problem, not an abstract ethical concern. After a switch to local processing, a less anthropomorphic language-use might further raise maturity level. Here the combined interests of all actors contribute towards an application with features in line with applied policies, driving towards more ethically full-featured applications promoting sustainable and responsible development.
In this paper, we presented the RAIN framework; a structured methodology for translating high-level policy to concrete normative requirements and features. Using the count-as operator, we can formally represent the sociotechnical contexts where a policy is relevant for an application. Formal representation of value-context relations allows us to trace requirements and features to the values they represent in both a verifiable and transparent way. The framework allows a structured discussion and communication about AI systems in a low-overhead manner; enabling efective policy making, compliance checking and ethics in a concrete way. RAIN considers all levels of the emerging ecosystem of stakeholders: developers, procurers, users, regulators, and policymakers.
The RAIN Graph shifts the guidelines and assessment criteria from abstract values to contextualised features and requirements. In contrast to high-level AI Ethics, software development is already apt at working with such feature requirements. Expert knowledge embedded in the graph decreases the overhead of local practicalphilosophy and policy expertise and complicated organisational containment-strategies, thus increasing both the availability and impact of any policy.
As more AI products are marketed, complex software
containing multiple AI modules developed by multiple
developers procured by yet other public or commercial
organisations will become more common. Applying a
RAIN Graph based assessment from the module level up,
and from the top-organisational level down facilitates
full-chain modular policy compliance checking and
charting of responsibility. Cross-application of local
organisational policies and national and international guidelines
allows procurers to set their own terms and requirements
on their suppliers, enabling each layer of the chain to take
responsibility [
Continuing on alleviating the overhead on the AI ecosystem, our future work includes adding a functional model of AI systems to the graph representation which would extend the scope from high-level principles to a more direct multi-level treatment of explainability, contestability, and trust. Finally, our future work also includes field testing of tools and methodologies building on the presented framework.
tonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation. Brännström, Theodorou and Dignum thank the Knut and Alice Wallenberg Foundation for grant RAIN (2020:2012) that supported their eforts.