Let it RAIN for Social Good Mattias Brännström1,∗ , Andreas Theodorou1 and Virginia Dignum1 1 Umeå University, Universitetstorget 4, 901 87 Umeå, Sweden Abstract 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 efforts stand at the pivot point between these outcomes but is barred from effect 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 effective high-level control of AI impact. With effective and operationalized AI Ethics, AI technologies can be directed towards global sustainable development. Keywords AI assessment, value-sensitive design, AI ethics, accountability 1. Introduction on advocating high-level ethical principles such as fair- ness, transparency, accountability, and respect for human Several recent and comprehensive reviews make clear values [5]. that there is a strong connection between large-scale An often mentioned problem of the high-level guide- change and developments in Artificial Intelligence (AI) lines is that they are at times too abstract to be applied [1, 2]. All of the 17 Sustainable Development Goals to any particular case and and at other times too spe- (SDGs) for Sustainable Development are believed to be cific by mentioning problems which might not exist in moderately or strongly affected by AI technology. Stud- a particular application. There is no particular level of ies show that 59 of the sustainable development targets abstraction that solves this problem for high-level policy might actually be inhibited by AI and there is reason to as guidelines either become too abstract or too extensive. believe this is a low estimate [1]. The large scale pre- A gap thus appears between high-level policy and any dicted effects of AI take up a complicated role as some practical application [5, 6]. progress towards sustainability might be dependent on Further exacerbating the problem is that the socio- AI for the required changes. Some studies even go as technical domain typically consist of not a single actor, far as to term this technological progression a “vector of the AI developer, but an interplay between developers, hope” [2]. Research gaps exist regarding the large scale procurers, customers and users [7, 8]. It is within this effects in the interplay between AI technologies and so- socio-technical multi-actor sphere where the effects of ciety where AI related change could instead exacerbate AI on society develop [7, 8, 6]. Understanding this inter- negative narratives and global inequalities [3, 2, 4]. play and successful bridging this abstraction gap between A central role in determining the outcome, positive high-level policy and particular application is of central or negative, of AI on large-scale sustainability and the importance in establishing socially-beneficial AI. SDGs is taken by AI Ethics. It is widely recognized that The abstraction gap is not only a problem from a regu- effective soft and hard policies on AI technologies are latory perspective. For the individual developer, procurer, needed to ensure positive outcomes. Many attempts or any other actor dealing with emerging AI applications at high-level soft policy already exists by intergovern- where the gap severs the link between design and organ- mental organisations, e.g. the European Commission’s isational choices on one hand and outcomes, ethical or “Guidelines for Trustworthy AI” (GTAI), but also by pro- otherwise, on the other. As there is no clear link between fessional bodies, e.g. IEEE. Such policy documents focus the particularities of an AI application and high-level eth- ical goals, there is no clear path forward even for actors The IJCAI-ECAI-22 Workshop on Artificial Intelligence Safety (AISafety on all levels who desire to act responsibly. 2022), July 24–25, 2022, Vienna, Austria Currently, bridging this gap require expert involve- ∗ Corresponding author. ment and analysis. This contribute to increase the divides Envelope-Open mattias.brannstrom@umu.se (M. Brännström); and inequalities already present in society, decrease the andreas.theodorou@umu.se (A. Theodorou); virginia.dignum@umu.se (V. Dignum) transparency of AI Ethics itself and undermine trust in GLOBE http://www.recklesscoding.com (A. Theodorou) AI technologies. In other words, negatively contribut- Orcid 0000-0003-3113-2631 (M. Brännström); 0000-0001-9499-1535 ing towards the SDGs. Effects like these are even more (A. Theodorou); 0000-0001-7409-5813 (V. Dignum) prominent in areas where both expertise and effective © 2022 Copyright 2022 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0) governance structures with a strong ethical focus is lack- CEUR Workshop Proceedings http://ceur-ws.org ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org) ing; leading to an AI ethical void in the most sensitive into contextualised lower-level norms[11]. Linguistically areas for increasing global inequality [4]. counts-as represents the construct ‘X counts as Y in con- 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, [12]. Counts-as enables the expression of values in spe- and in communicating that context across actors. AI cific contexts as sub-norms finally connecting to concrete Ethics without the specific context lacks solutions and design choices. low-level technical approaches loses sight of the goals Building upon VSD is the Glass-Box framework [13]. and the larger effects [6, 5, 9]. A continual chain encapsu- The Glass-Box approach demonstrate how the the same lating all levels of the socio-technical landscape is needed procedure can be used to to retrieve testable requirements. to ensure relevance to both actual applications and the The Glass Box consists of two phases which inform each larger society [8, 7]. When such a chain is explicit, it other: interpretation and observation. The interpretation can drive the transformative effects of these emerging stage translate values into specific design requirements technologies towards sustainable change in line with the by using the VSD approach where the relationship be- SDGs. tween values, norms, and requirements can be formally In this paper, a solution for bridging the abstraction represented using counts-as and modal logic [13, 12]. gap is presented: the Responsible AI Norms (RAIN) frame- Once found, the low-level requirements inform the work. RAIN breaks down abstract high-level policies observation stage of the approach. The requirements, into actionable norms by connecting them with socio- now linked to high-level values, can be automatically or technical contexts in a structured way. The formal struc- even continually assessed to determine to which degree a ture of RAIN provides clarity in the connections be- solution fulfills its stated values. However, two key limi- tween policies and actual AI applications on all levels tations remains: The interpretation step must, as VSD, be and thereby enables effective policy-making and policy done separately for each AI application, something that compliance with low overhead. The formal specifica- requires significant expertise and may produce diverse tions also enable reproducibility and auditability of the results. It also focus on measurable requirements, which RAIN-produced requirements. limits the approach to the technical compliance while AI 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 The RAIN framework provides further structure com- future work. pared to VSD in the hierarchical breakdown process of values in a way that makes the resulting norms hier- 2. 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 [10, 11]. VSD starts the overhead of Ethical AI. It shifts the focus towards by placing the focus on a socio-ethical value relevant for application features which is more readily dealt with by the use case at hand. From this perspective, any associ- current technical B2B-landscape. It also enables a struc- ated values, stakeholders, and technologies are found by tured way to work with and communicate around policy iterative exploration. Harms and benefits of each identi- and AI Ethics between societal actors. In addition the fied group of stakeholders are determined and connected RAIN framework also serve as a kind of knowledge elici- to relevant values and values are prioritized. After this tation from experts. This embedded knowledge can be of mapping has taken place, conflicts between values, tech- aid in settings where such expertise might not be avail- nological solutions and project goals can be brought to able. Such clarity around tangible impacts, responsibility forefront in the design process. Key here is the explo- and concrete ethical choices is key for transparent and ration of how a value impacts design by exploring the accountable use of AI technology which drives towards intersections between value, stakeholders and techno- the SDGs rather than inequality and exploitation. logical use case. The goal of the process is to facilitate In this section the RAIN framework will be described discussion and understanding. in detail, starting with the creation of the norms hier- 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 scaffold, particular a policy can be seen particular policies. as sub-context providing additional detail primarily to the abstract concepts HLV and issue by specifying sub- 3.1. Overview of the RAIN Pipeline classes. While policies are not typically written in such a structured manner we can use this scaffold 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 fea- below 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 3.2.2. The RAIN scaffold and scoring model The RAIN Graph captures AI policy in a structured man- ner. In this section it will be described how such a Graph With the scaffold for the policies in place, we can follow can be derived from a high-level policy but also expanded this with a new context 𝑅0 ⪯ 𝑇𝑐 forming the basis of the to particular contexts not explicitly mentioned in such rain framework. We can describe 𝑅0 as: a policy. High-Level Policy (HLP) will in this section 𝑅0 ∶𝑣𝑎𝑙𝑢𝑒 ⊑ 𝐻 𝐿𝑉 (4) be defined as any policy, guideline, standard or the like which primarily bases itself upon High-Level ethical Val- 𝑅0 ∶𝑣𝑖𝑜𝑙𝑎𝑡𝑒-1 ⊑ 𝑣𝑖𝑜𝑙𝑎𝑡𝑒 (5) ues (HLV) and presents challenges to these from the use 𝑅0 ∶𝑣𝑖𝑜𝑙𝑎𝑡𝑒-2 ⊑ 𝑣𝑖𝑜𝑙𝑎𝑡𝑒-1 of AI technology, solutions to such challenges, or re- 𝑅0 ∶𝑣𝑖𝑜𝑙𝑎𝑡𝑒-3 ⊑ 𝑣𝑖𝑜𝑙𝑎𝑡𝑒-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 (𝑥 ∶ 𝑦, a minor degree and each other level indicate lesser de- where 𝑦 applies in context 𝑥), counts-as (⇒𝑐 ) operators grees of compliance with more serious violations. Each and context relation ⪯ as described by [12]. The for- level is given a concrete definition as to what type of re- malism make relationships exact and explicit, something quirements it contains. Possible attached meaning to the which is required for the framework to work in repro- scoring model is not the focus of this paper. Hence, we ducible inter-operability and communication of concerns use a 3-tiered score model that will be used as an example between actors. The formalisation also lends itself readily how scoring mechanisms are tied to the framework (4,5). to implementation. Different numbers of levels and different definitions of each level work in the same way. 3.2.1. A scaffold for High-level AI Policy The scoring model here described results in a threshold model of aggregation. Within each category the aggre- Before breaking down and structuring any AI policy, we gated score will be the worst score within that category. start by defining a simple scaffold in which to understand This approach counteracts ‘ethics washing’ the approach them. We specify that: of doing something less-relevant well to make up for major failures in more relevant areas. It also helps to 𝑇𝑐 ∶𝐻 𝐿𝑉 ⊑ 𝐴𝐼 𝐸𝑡ℎ𝑖𝑐𝑠 (1) highlight the ethical issues with an application where 𝑇𝑐 ∶𝑒𝑡ℎ𝑖𝑐𝑎𝑙𝐴𝐼 ≡ 𝐴𝐼 ⊓ ¬∃𝑣𝑖𝑜𝑙𝑎𝑡𝑒.𝐴𝐼 𝐸𝑡ℎ𝑖𝑐𝑠 (2) the most difference can be made. 𝑇𝑐 ∶𝑖𝑠𝑠𝑢𝑒 ≡ ∃𝑣𝑖𝑜𝑙𝑎𝑡𝑒.𝐴𝐼 𝐸𝑡ℎ𝑖𝑐𝑠 (3) 3.2.3. The RAIN Graph That is, in the top context HLV are sub-concepts of AI Ethics (1). Ethical AI is AI which does not violate AI A RAIN Graph, 𝐺 can be described to contain the follow- Ethics (2) and an issue is something that do (3). ing concepts • value A parsimonious sub concept of HLV. Often Algorithm 1: RAIN Decomposition Algorithm values in policy are expressed using several sub Data: P, a policy values or norms, these are here separated. We Data: 𝐺(𝑉 , 𝑆, 𝐹 , 𝑁 , 𝑁𝑐 ), a RAIN Graph will term the set of all values 𝑣 ∈ 𝐺 as 𝑉 begin • stakeholder Reflecting a perspective of concern for hlv ⊑ HLV ∈ P do for a stakeholder group. We will term the set of merge component values 𝑣 of hlv to 𝑉 all stakeholders 𝑠 ∈ 𝐺 as 𝑆. if Explicit stakeholders ∈ P then • socio-technical feature A socio-technical use merge component stakeholder concerns 𝑠 of technology. We will term the set of all socio- of policy into 𝑆 technical features 𝑓 ∈ 𝐺 as 𝐹. • RAI norms and contexts A norm 𝑛(𝑠, 𝑓 ) ∈ 𝐺 if Explicit socio-technical features ∈ P then merge component stakeholder concerns 𝑠 represent a particular challenge caused by some of policy into 𝑆 socio-technical feature 𝑓 ∈ 𝐹 to a value 𝑣 ∈ 𝑉 with respect to a stakeholder concern 𝑠 ∈ 𝑆. We will for 𝑖 ⊑ issue ∈ P do term the set of all such norms 𝑛 as 𝑁. Every such 𝑉𝑖 ← Values 𝑣 ⊂ 𝑉 impacted by issue 𝑖 norm 𝑛 ∈ 𝐺 will be embedded in a sub-context 𝑆𝑖 ← Stakeholder concerns impacted by 𝑛𝑐 ⪯ 𝑅0 such that 𝑛𝑐 ∶ 𝑛 ⊑ (𝑓 ⊓ 𝑣 ⊓ 𝑠). issue 𝑖 𝐹𝑖 ← Socio-technical features which must The sets 𝑉 , 𝑆, 𝐹 , 𝑁 , 𝑁𝑐 are considered to be holding se- be present for issue 𝑖 to threaten 𝑉𝑖 with mantically distinct items. For the algorithms 1 and 2, we regards to 𝑆𝑖 define the operation merge to mean an addition that pre- merge concerns 𝑆𝑖 into 𝑆 serves semantic distinctness. In the case of RAI norms, 𝑁 merge features 𝐹𝑖 into 𝐹 multiple distinct issues with corresponding assessment merge norm 𝑛(𝑉𝑖 , 𝑆𝑖 , 𝐹𝑖 ) into 𝑁 and 𝑁𝑐 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. own contexts 𝑛𝑐 ⪯ 𝑅0 where 𝑛𝑐 represents the active pres- ence of a stakeholder and feature instance in the context 3.2.4. RAI Norms and contexts of the application. The general structure of this context The RAI norms are the central content of the RAIN Graph. also including the foundation of the assessment layer can These norms can be seen as representing the junction be expressed as follows: between a value, a subject, and a circumstance. Or value, 𝑛𝑐 ∶𝑁𝑎 ⇒𝑐 (𝑣 ∧ 𝑠 ∧ 𝑓 ) (6) subject, and action. Through these norms, it is possible 𝑛𝑐 ∶𝐴𝑠𝑠𝑒𝑠𝑠𝑚𝑒𝑛𝑡-1 ≡ ∃𝑣𝑖𝑜𝑙𝑎𝑡𝑒-1.𝑁𝑎 (7) to determine what features of a given context which are related to which values and for whom. These relation- 𝑛𝑐 ∶𝐴𝑠𝑠𝑒𝑠𝑠𝑚𝑒𝑛𝑡-2 ≡ ∃𝑣𝑖𝑜𝑙𝑎𝑡𝑒-2.𝑁𝑎 ships are the main purpose of the RAIN Graph and how 𝑛𝑐 ∶𝐴𝑠𝑠𝑒𝑠𝑠𝑚𝑒𝑛𝑡-3 ≡ ∃𝑣𝑖𝑜𝑙𝑎𝑡𝑒-3.𝑁𝑎 it helps to bridge the abstraction gap. Since each of the RAIN nodes identifies a particular 3.2.5. Operational semantics algorithms threat, it can be accompanied with a corresponding set of requirements alleviating that threat. In this manner, The RAIN Decomposition Algorithm encodes a policy a context-sensitive assessment of how a given AI ap- into the graph. A second algorithm described here, the plication complies with one or several policies can be RAIN Expansion Algorithm, fills out the missing areas expressed as the degree of which it fulfills the require- of concern and expands the policy with consideration of ments selected by its features. Some types of the technical a potentially new area of socio-technical context. requirements can be verified in an automated manner; Algorithm 1, the decomposition algorithm or back- in other words, the RAIN Graph fulfills the interpreta- wards algorithm goes from policy and provides a RAIN tion stage of the Glass Box by identifying in which ways graph encoding of its content. it is relevant to monitor an application with regards to 1. Start with a policy document. Ex. GTAI. ethical concerns. Other requirements, concerning or- 2. Identify the top values which are directly im- ganisational features, design choices, or documentation, pacted or taken into consideration by the policy. require a wider socio- intervention by stakeholders. Such Ex. Privacy. requirements instead lend themselves to manual assess- 3. Consider the characterization of each of these ment procedures. Formally we can represent these RAI high level values in order to break down each of norms and their accompanying assessment rules as their these high-level concepts into singular areas of concern. The key here is to be parsimonious. One Algorithm 2: RAIN Expansion Algorithm concern or concept per item. Data: 𝐺(𝑉 , 𝑆, 𝐹 , 𝑁 ), a RAIN Graph Ex: Right to Privacy, Data protection, Data Governance Data: 𝐹𝑛𝑒𝑤 a set of new socio-technical features ⊏ Privacy ∈ GTAI begin 4. merge the resulting derived top values or top merge 𝐹𝑛𝑒𝑤 into 𝐹 norms form the values of the Graph with respect for (𝑓 , 𝑣, 𝑠) ∈ 𝐹 × 𝑉 × 𝑆 do to this policy. for Issues 𝑖 which 𝑓 threaten 𝑣 with respect 5. If Stakeholders or particular socio-technical fea- to 𝑠 do tures are explicitly mentioned in the policy, re- 𝑉𝑖 ← Values 𝑣 ⊂ 𝑉 impacted by Issue 𝑖 peat the previous step for them as well in the 𝑆𝑖 ← Stakeholder concerns impacted same manner. Ex. End Users and Developer are by Issue 𝑖 mentioned stakeholders in GTAI. 𝐹𝑖 ← Socio-technical features which 6. Go through each Issue raised by this policy and must be present for Issue 𝑖 to state it in connection to at least one Value and threaten 𝑉𝑖 with regards to 𝑆𝑖 at least one Stakeholder. In addition determine merge norm 𝑛(𝑉𝑖 , 𝑆𝑖 , 𝐹𝑖 ) into 𝑁 and 𝑁𝑐 what socio-technical feature which must be in merge concerns 𝑆𝑖 into 𝑆 place for the issue to exist. It might be a way of merge features 𝐹𝑖 into 𝐹 dealing with data, a particular technology, or a particular use case, for example. Ex: Handling if ∅ = {𝑛|𝑛 ∈ 𝑁 relates to 𝑓 } then of the personal data of End Users is required for remove 𝑓 from 𝐹 issues of GDPR in GTAI. 7. The identified issue or problem can now be stated as one or more RAI norms which state that this Feature threaten the identified Value with respect Issues identified in this manner is treated just as to the identified Stakeholder. These RAI norms in Algorithm 1 in order to add new RAI norms are added to the Graph in their own context, as and RAIN-norm contexts to the Graph. Ex. Re- described above. mote processing interacts strongly with the GTAI Ex. personal data𝑐 ∶ 𝑁𝑝 𝑑 ⇒𝑐 (Personal data ∧ End values already in the graph, regarding Privacy, Ro- User ∧ Data Governance). bustness and Transparency. Each interaction give 8. When all the issues mentioned in the policy are rise to RAI Norms. treated in this manner one can consider the con- 3. When all features are considered, features which tent of the Graph to contain the explicit parts can not be connected to both a value and a stake- of the policy. However since most policy have holder are removed. Ex. If features were added at selected some level of abstraction and scope, it step 1 which were of no consequence, then they are is likely that many intended issues related to AI removed here. Ethics are not yet mentioned in the Graph. These are captured using Algorithm 2. 3.2.6. Multiple policies and coverage Algorithm 2 consider each intersection of identified features, concerns and values in order to fill in the blanks. Algorithm 1 & 2 are both idempotent and can be used It can also be used for considering a particular set of repeatedly to merge several policies into a single RAIN socio-technical features in the light of the values and Graph. If policies overlap in values and issues, parts of stakeholder concerns in the Graph. In this manner the the graph might be unchanged by such additions. A spe- Graph can be extended to cover new particular contexts. cial case of interest is when policies have values which are defined differently. That ethical values lack a univer- 1. Add any socio-technical features to be consid- sal definition is a common mentioned problem [5]. This ered to the Graph. Ex. For the example in the is actually not a problem for the RAIN Graph as differing next section, features common to home automation, definitions mean their component values differ. In this voice control and human interaction such as e.g. manner it is possible to combine even apparently conflict- Remote Processing and Passive Recording. Socio- ing policies in the same RAIN Graph. Untangling these technical features such as Vulnerable End Users are possibly conflicting viewpoints is handled by their dif- also relevant to the elderly care example below. ferent semantics and the activation of different contexts, 2. For each intersection between a value, a stake- and the result projection mentioned below. holder concern and a socio-technical feature, con- As policy are combined into the RAIN Graph in this sider the possible ways in which the value is chal- manner it is possible to define a RAIN Graphs coverage. lenged with regards to the stakeholder concern. 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 re- duce 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 appli- safety 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, An- thropomorphic 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 stakehold- embedded into a pipeline with the following three steps: ers. These identified in the context layer of the pipeline. • Context layer capturing the socio-technical con- 3.3.2. Assessment layer text and identifying stakeholders, top level values and policies. The output of this layer is context Given that a certain set of features and stakeholders have features and activated values. been asserted by the context layer, some of the contexts • Assessment layer providing context-specific of G will be active, and their statements will apply. The testable requirements, satisfying the identified purpose of the assessment layer is to see if the rules (7) norms on a five-step scale of compliance. with regards to these contextualized norms have been • Result layer aggregating the result of the indi- violated or not. Each of these assessment statements vidual norms onto the high-level values as well are connected to an appropriate type of test (e.g quiz, as projections upon compatible policies of choice. 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 assess- of 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. Be- public-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 effects 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. Anthropomor- phic language is a Transparency concern especially due to The features of the Graph are, as per the Algorithms 1 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 particu- lar socio-technical domain, the set of features the Graph 3.3.3. Results layer and projections of the emerging ecosystem of stakeholders: developers, procurers, users, regulators, and policymakers. When assessments have been performed, the result can The RAIN Graph shifts the guidelines and assessment be evaluated in several ways. A straightforward way is criteria from abstract values to contextualised features to enumerate the Values in set 𝑉 and determine what and requirements. In contrast to high-level AI Ethics, level of violation and thus maturity score which applies software development is already apt at working with to each Value. This would be a RAIN Graph-specific such feature requirements. Expert knowledge embedded result. Another straightforward way is to look to the in the graph decreases the overhead of local practical- context of a particular policy and similarly enumerate its philosophy and policy expertise and complicated organi- particular HLVs together with the aggregated maturity sational containment-strategies, thus increasing both the level. A less straightforward but highly effective way is availability and impact of any policy. to provide a set of statements on the contents of 𝐺, where As more AI products are marketed, complex software each statement maps to a particular requirement of an containing multiple AI modules developed by multiple external assessment covered by 𝐺. For instance if a RAIN developers procured by yet other public or commercial Graph covers GTAI, a set of statements on the graph can organisations will become more common. Applying a map the results to each of the assessment questions in RAIN Graph based assessment from the module level up, the guidelines. This way a particular high-level policy and from the top-organisational level down facilitates can be assessed in a context-aware manner even if the full-chain modular policy compliance checking and chart- policy itself is not constructed for the RAIN framework. ing of responsibility. Cross-application of local organisa- Given the structure embedded in the RAIN graph, re- tional policies and national and international guidelines sults can also be aggregated on particular stakeholders or allows procurers to set their own terms and requirements socio-technical features, giving a valuable and detailed on their suppliers, enabling each layer of the chain to take description on how ethical compliance is distributed over responsibility [7, 8]. This structured approach applied the socio-technical landscape of the Application. on a top-policymaking level enables top-down discus- Ex. While the system get high maturity levels on na- sions focused on socio-technical hot-spots rather than tional safety standards, the aggregated GTAI scores are nebulous and hard-to-define AI. strongly violated due to the Remote Processing, especially Continuing on alleviating the overhead on the AI with regards to Privacy. The local Procurers internal guide- ecosystem, our future work includes adding a functional lines are also found violated and the system is rejected. The model of AI systems to the graph representation which developer of the system could adapt for on-site processing would extend the scope from high-level principles to a of recorded data to gain a higher Privacy maturity level. more direct multi-level treatment of explainability, con- Such adaption is a concrete technical and business problem, testability, and trust. Finally, our future work also in- not an abstract ethical concern. After a switch to local cludes field testing of tools and methodologies building processing, a less anthropomorphic language-use might on the presented framework. further raise maturity level. Here the combined interests of all actors contribute towards an application with fea- tures in line with applied policies, driving towards more Acknowledgments ethically full-featured applications promoting sustainable and responsible development. This work was supported by the Wallenberg AI, Au- tonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation. 4. Discussion and Future Work Brännström, Theodorou and Dignum thank the Knut and Alice Wallenberg Foundation for grant RAIN (2020:2012) In this paper, we presented the RAIN framework; a struc- that supported their efforts. tured methodology for translating high-level policy to concrete normative requirements and features. Using the count-as operator, we can formally represent the socio- References technical contexts where a policy is relevant for an appli- cation. Formal representation of value-context relations [1] R. Vinuesa, H. Azizpour, I. Leite, M. Balaam, allows us to trace requirements and features to the values V. Dignum, S. Domisch, A. Felländer, S. D. Lang- they represent in both a verifiable and transparent way. hans, M. Tegmark, F. F. Nerini, The role of artificial The framework allows a structured discussion and com- intelligence in achieving the sustainable develop- munication about AI systems in a low-overhead manner; ment goals, Nature communications 11 (2020) 1–10. enabling effective policy making, compliance checking [2] G. D. R. Castro, M. C. G. Fernández, Á. U. Colsa, Un- and ethics in a concrete way. RAIN considers all levels leashing the convergence amid digitalization and sustainability towards pursuing the sustainable de- velopment goals (sdgs): A holistic review, Journal of Cleaner Production 280 (2021) 122204. [3] L. Sartori, A. Theodorou, A sociotechnical perspec- tive for the future of ai: narratives, inequalities, and human control, Ethics and Information Technology 24 (2022) 1–11. [4] P. Gehl Sampath, Governing artificial intelligence in an age of inequality, Global Policy (2021). [5] A. Theodorou, V. Dignum, Towards ethical and socio-legal governance in ai, Nature Machine Intel- ligence 2 (2020) 10–12. [6] B. Mittelstadt, Principles alone cannot guarantee ethical ai, Nature Machine Intelligence 1 (2019) 501–507. [7] D. S. Rubenstein, Acquiring ethical ai, Florida Law Review 73 (2021). [8] G. Falco, B. Shneiderman, J. Badger, R. Carrier, A. Dahbura, D. Danks, M. Eling, A. Goodloe, J. Gupta, C. Hart, et al., Governing ai safety through independent audits, Nature Machine Intelligence 3 (2021) 566–571. [9] S. Bird, M. Dudík, R. Edgar, B. Horn, R. Lutz, V. Mi- lan, M. Sameki, H. Wallach, K. Walker, Fairlearn: A toolkit for assessing and improving fairness in ai, Microsoft, Tech. Rep. MSR-TR-2020-32 (2020). [10] B. Friedman, P. H. Kahn, A. Borning, A. Huldtgren, Value sensitive design and information systems, in: Early engagement and new technologies: Opening up the laboratory, Springer, 2013, pp. 55–95. [11] I. Van de Poel, Translating values into design requirements, in: Philosophy and engineering: Reflections on practice, principles and process, Springer, 2013, pp. 253–266. [12] D. Grossi, Designing invisible handcuffs: Formal investigations in institutions and organizations for multi-agent systems, volume 16, 2007. [13] A. Aler Tubella, V. Dignum, The glass box ap- proach: Verifying contextual adherence to values, in: AISafety 2019, Macao, China, August 11-12, 2019, CEUR-WS, 2019.