Principle of responsible delivery through human-centred implementation, key considerations

"The demand for sensitivity to human factors should inform your approach to devising delivery and implementation processes from start to finish. To provide clear and effective explanations about the content and rationale of algorithmic outputs, you will have to begin by building from the human ground up. You will have to pay close attention to the circumstances, needs, competences, and capacities of the people whom your AI project aims to assist and serve.This means that context will be critical. By understanding your use case well and by drawing upon solid domain knowledge, you will be better able to define roles and relationships. You will better be able to train the users and implementers of your system. And, you will be better able to establish an effectual implementation platform, to clarify content, and to facilitate understanding of outcomes for users and affected stakeholders alike. Here is a diagram of what securing human-centred implementation protocols and practices might look like:Let us consider these steps in turn by building a checklist of essential actions that should be taken to help ensure the human-centred implementation of your AI project. Because the specifics of your approach will depend so heavily on the context and potential impacts of your project, we’ll assume a generic case and construct the checklist around a hypothetical algorithmic decision-making system that will be used for predictive risk assessment.Step 1: Consider aspects of application type and domain context to define roles and determine user needs
1. Assess which members of the communities you are serving will be most affected by the implementation of your AI system. Who are the most vulnerable among them? How will their socioeconomic, cultural, and education backgrounds affect their capacities to interpret and understand the explanations you intend to provide? How can you fine-tune your explanatory strategy to accommodate their requirements and provide them with clear and non-technical details about the rationale behind the algorithmically supported result? When thinking about providing explanations to affected stakeholders, you should start with the needs of the most disadvantaged first. Only in this way, will you be able to establish an acceptable baseline for the equitable delivery of interpretable AI.
1. After reviewing Guideline 1 above, make a list of and define all the roles that will potentially be involved at the delivery stage of your AI project. As you go through each role, specify levels of technical expertise and domain knowledge as well as possible goals and objectives for each role. For instance, in our predictive risk assessment case:- Decision Subject (DS)- - Role: Subject of the predictive analytics. - Possible Goals and Objectives: To receive a fair, unbiased, and reasonable determination, which makes sense; to discover which factors might be changed to receive a different outcome. - Technical and Domain Knowledge: Most likely low to average technical expertise and average domain knowledge.- Advocate for the DS- - Role: Support for the DS (for example, legal counsel or care worker) and concerned party to the automated decision. - Possible Goals and Objectives: To make sure the best interests of the DS are safeguarded throughout the process; to help make clear to the DS what is going on and how and why decisions are being made. - Technical and Domain Knowledge: Most likely average technical expertise and high level of domain knowledge.- Implementer - Role: User of the AI system as decision support. - Possible Goals and Objectives: To make an objective and fair decision that is sufficiently responsive to the particular circumstances of the DS and that is anchored in solid reasoning and evidence-based judgement. - Technical and Domain Knowledge: Most likely average technical expertise and high level of domain knowledge.- System Operator/Technician- - Role: Provider of support and maintenance for the AI system and its use.- Possible Goals and Objectives: To make sure the machine learning system is performing well and running in accordance with its intended design; to handle the technical dimension of information processing for the DS’s particular case; to answer technical questions about the system and its results as they arise.- Technical and Domain Knowledge: Most likely high level of technical expertise and average domain knowledge.- Delivery Manager- - Role: Member of the implementation team who oversees its operation and responds to problems as they arise. - Possible Goals and Objectives: To ensure that the quality of the automation- supported assessment process is high and that the needs of the decision subject are being served as intended by the project; to oversee the overall quality of the relationships within the implementation team and between the members of that team and the communities they serve. - Technical and Domain Knowledge: Most likely average technical expertise and good to high level domain knowledgeStep 2: Define delivery relations and map delivery processes
1. Assess the possible relationships between the defined roles that will have significant bearing on your project’s implementation and formulate a descriptive account of this relationship with an eye to the part it will play in the delivery process. For the predictive risk assessment example:- Decision Subject/Advocate to Implementer: This is the primary relationship of the implementation process. It should be information-driven and dialogue-driven with the implementer’s exercise of unbiased judgment and the DS’s comprehension of the outcome treated as the highest priorities. Implementers should be prepared to answer questions and to offer evidence-based clarifications and justifications for their determinations. The achievement of well-informed mutual understanding is a central aim.- Implementer to System Operator: This is the most critical operational relationship within the implementation team. Communication levels should be kept high from case to case, and the shared goal of the two parties should be to optimise the quality of the decisions by optimising the use of the algorithmic decision-support system in ways that are accessible both to the user and to the DS. The conversations between implementers and system operators should be problem-driven and should avoid, as much as possible, focus on the specialised vocabularies of each party’s domain of expertise.- Delivery Manager to Operator to Implementer: The quality of this cross-disciplinary relationship within the implementation team will have direct bearing on the overall quality of the delivery of the algorithmically supported decisions. Safeguarding the latter will require that open and easily accessible lines of communication be maintained between delivery managers, operators, and implementers, so that unforeseen implementation problems can be tackled from multiple angles and in ways that anticipate and stem future difficulties. Additionally, different use cases may present different explanatory challenges that are best addressed by multidisciplinary team input. Good communications within the implementation team will be essential to enable that such challenges are addressed in a timely and efficient manner.
1. Start building a map of the delivery process. This should involve incorporating your understanding of the needs, roles, and relationships of relevant actors involved in the implementation of your AI system into the wider objective of providing clear, informative, and understandable explanations of algorithmically supported decisions. It is vital to recognise, at this implementation-planning stage of your project, that the principal goal of the delivery process is two-fold: to translate statistically expressed results into humanly significant reasons and to translate algorithmic outputs into socially meaningful outcomes.These overlapping objectives should have a direct bearing on the way you build a map for your project’s delivery process, because they organise the duties of implementation into two task-specific components:
1. A technical component, which involves determining the most effective way to convey and communicate to users and decision subjects the statistical results of your model’s information processing so that the factors that figured into the logic and rationale of those results can be translated into understandable reasons that can be subjected to rational evaluation and critical assessment; and
1. A social component, which involves clarifying the socially meaningful content of the outcome of a given algorithmically assisted decision by translating that model’s technical machinery—its input and output variables, parameters, and functional rationale—back _into the everyday language of the humanly relevant categories and relationships that informed the formulation of its purpose, objective, and intended elements of design in the first place. Only through this re-translation will the effects of that model’s output on the real human life it impacts be understandable in terms of the specific social and individual context of that life and be conveyable as such.These two components of the delivery process will be fleshed out in turn.Technical component of responsible implementation: As a general rule, we use the results of statistical analysis to guide our actions, because, when done properly, this kind of analysis offers a solid basis of empirically derived evidence that helps us to exercise sound and well- supported judgment about the matters it informs.Having a good understanding of the factors that are at work in producing the result of a particular statistical analysis (such as in an algorithmic decision-support system) means that we are able to grasp these factors (for instance, input features that weigh heavily in determining a given algorithmically generated decision) as reasons that may warrant the rational acceptability of that result. After all, seen from the perspective of the interpretability of such an analysis, these factors are, in fact, nothing other than _reasons that are operating to support its conclusions.Clearly understood, these factors that lie behind the logic of the result or decision are not ‘causes’ of it. Rather, they form the evidentiary basis of its rational soundness and of the goodness of the inferences that support it. Whether or not we ultimately agree with the decision or the result of the analysis, the reasons that work together to comprise its conclusions make claims to validity _and can _as such _be called before a tribunal of _rational criticism. These reasons, in other words, must bear the burden of continuous assessment, evaluation, and contestation.This is an element especially crucial for the responsible implementation of AI systems: Because they serve surrogate cognitive functions in society, their decisions and results are in no way immune from these demands for rational justification and thus must be delivered to be optimally responsive to such demands.The results of algorithmic decision support systems, in this sense, serve as stand-ins for acts of speech and representation and therefore bear the justificatory burdens of those cognitive functions. They must establish the validity of their conclusions and operate under the constraint of being surrogates of the dialogical goal to convince through good reasons.This charge to be responsive to the demands of rational justification should be essential to the way you map out your delivery strategy. When you devise how best to relay and explain the statistical results of your AI systems, you need to start from the role they play in supporting evidence-based reasoning.This, however, is no easy job. Interpreting the results of data scientific analysis is, more often than not, a highly technical activity and can depart widely from the conventional, everyday styles of reasoning that are familiar to most. Moreover, the various performance metrics deployed in AI systems can be confusing and, at times, seem to be at cross-purposes with each other, depending upon the metrics chosen. There is also an unavoidable dimension of uncertainty that must be accounted for and expressed in confidence intervals and error bars which may only bring further confusion to users and decision subjects.Be that as it may, by taking a deliberate and human-centred approach to the delivery process, you should be able to find the most effective way to convey your model’s statistical results to users and decision subjects in non-technical and socially meaningful language that enables them to understand and evaluate the rational justifiability of those results. A good point of departure for this is to divide your map-building task into the means of content delivery _and the _substance of the content to be delivered.Means of content delivery: _When you start mapping out serviceable ways of presenting and communicating your model’s results, you should consider the users’ and decision subjects’ perspectives to be of primary importance. Here are a few guiding questions to ask as you sketch out this dimension of your delivery process as well as some provisional answers to them:- How can the delivery process of explaining the AI system’s results aid and augment the user’s and decision subject’s _mental models _(their ways of organising and filtering information), so that they can get a clear picture of the technical meaning of the assessment or explanation? What is the best way to frame the statistical inferences and meanings so that they can be effectively integrated into each user’s own cognitive _space of concepts and beliefs?While answering these questions will largely depend both on your use case and on the type of AI application you are building, it is just as important that you start responding to them by concentrating on the differing needs and capabilities of your explainees. To do this properly, you should first seek input from domain experts, users, and affected stakeholders, so that you can suitably scan the horizons of existing needs and capabilities. Likewise, you should take a human-centred approach to exploring the types of explanation delivery methods that would best be suited for each of your target groups. Much valuable research has been done on this in the field of human-computer interaction and in the study of human factors. This work should be consulted when mapping delivery means.Once you have gathered enough background information, you should begin to plan out how you are going to line up your means of delivery with the varying levels of technical literacy, expertise, and cognitive need possessed by the relevant stakeholder groups, who will be involved in the implementation of your project. Such a multi-tiered approach _minimally requires that individual attention be paid to the explanatory needs and capacities of implementers, system operators, and decision subjects and their advocates. This multi-tiered approach will pose different challenges at each different level.For instance, the mental models of implementers—i.e. their ways of conceptualising the information they are receiving from the algorithmic decision-support system— may, in some cases, largely be shaped by their accumulation of domain know-how and by the filter of on-the-job expertise that they have developed over long periods of practice. These users may have a predisposition to automation distrust or aversion bias, and this should be taken into account when you are formulating appropriate means of explanation delivery.In other contexts, the opposite may be the case. Where implementers tend to over- rely on or over-comply with automated systems, the means of explanation delivery must anticipate a different sort of mental model and adjust the presentation of information accordingly.In any event, you will need to have a good empirical understanding of your implementer’s decision-making context and maintain such knowledge through ongoing assessment. In both bias risk areas, the conveyance and communication of the assessments generated by algorithmic decision-support systems should attempt to bolster each user’s practical judgment in ways that mitigate the possibility of either sort of bias. These assessments should present results as evidence-based reasons that support and better capacitate the objectivity of these implementers’ reasoning processes.The story is different with regard to the cognitive life of the technically inclined user. The mental models of system operators, who are natives in the technical vocabulary and epistemic representations of the statistical results, may be adept at the model- based problem-solving tasks that arise during implementation but less familiar with identifying and responding to the cognitive needs and limitations of non-technical stakeholders. Incorporating ongoing communication exercises and training into their roles in the delivery process may capacitate them to better facilitate implementers’ and decision subjects’ understanding of the technical details of the assessments generated by algorithmic decision-support systems. These ongoing development activities will not only helpfully enrich operators’ mental models, they may also inspire them to develop deeper, more responsive, and more effective ways of communicating the technical yields of the analytics they oversee.Finally, the mental models of decision subjects and their advocates will show the broadest range of conceptualisation capacities, so your delivery strategy should (1) prioritise the facilitation of optimal explanation at the baseline level of the needs of the most disadvantaged of them and (2) build the depth of your multi-tiered approach to providing effective explanations into the delivery options presented to decision subjects and their advocates. This latter suggestion entails that, beyond provision of the baseline explanation of the algorithmically generated result, options should be given to decision subjects and their advocates to view more detailed and technical presentations of the sort available to implementers and operators (with the proviso that reasonable limitations be placed on transparency in accordance with the need to protect the confidential personal and organisational information and to prevent gaming of the system).- How can non-technical stakeholders be adequately prepared to gain baseline knowledge of the kinds of statistical and probabilistic reasoning that have factored into the technical interpretation of the system’s output, so that they are able to comprehend it on its own technical terms? How can the technical components be presented in a way that will enable explainees to easily translate the statistical inferences and meanings of the results into understandable and rationally assessable terms? What are the best available media for presenting the technical results in engaging and comprehensible ways?To meet these challenges, you should consider supplementing your implementation platform with knowledge-building and enrichment resources that will provide non- technical stakeholders with access to basic technical concepts and vocabulary. At a minimum, you should consider building a plain language glossary of basic terms and concepts that will include all of the technical ideas covered by the algorithmic component of a given explanation. If your explanation platforms are digital, you should also make them as user friendly as possible by hyperlinking the technical terms used in the explanations to their plain language glossary elaborations.Where possible, explanatory demonstrations of technical concepts (like performance metrics, formal fairness criteria, confidence intervals, etc.) should be provided to users and decision subjects in an engaging and easy-to-comprehend way, and graphical and visualisation techniques should be consistently used to make potentially difficult ideas more accessible. Moreover, the explanation interfaces themselves should be as simple, learnable, and usable as possible. They should be tested to measure the ease with which those with neither technical experience nor domain knowledge are able to gain proficiency in their use and in understanding their content.Substance of the technical content to be delivered: The overall interpretability of your AI system will largely hinge on the effectiveness and even-handedness of your technical content delivery. You will have to strike a balance between (1) determining how best to convey and communicate the rationale of the statistical results so that they may be treated appropriately as decision supporting and clarifying reasons and (2) being clear about the limitations of and potential uncertainties in the statistical results themselves so that the explanations you offer will not mislead implementers and decision subjects. These are not easy tasks and will require substantial forethought as you map out the content clarification aspect of your delivery process.To assist you in this, here is a non-exhaustive list of recommendations that you should consider as you map out the execution of the technical content delivery component of the responsible implementation of your AI project (This list will, for the sake of specificity, assume the predictive risk assessment example):- Each explanation should be presented in plain, non-technical language and in an optimally understandable way so that the results provided can enable the affordance of better judgment on the part of implementers and optimal understanding on the part of decision subjects. On the implementer’s side, the primary goal of the explanation should be to support the user’s ability to offer solid, coherent, and reasonable justifications of their determinations of decision outcomes. On the decision subject’s side, the primary goal of the explanation should be to make maximally comprehensible the rationale behind the algorithmic component of the decision process, so that the decision subject can undertake a properly informed critical evaluation of the decision outcome as a whole.- Each explanation should present its results as facts or evidence in as sparse but complete and sound a manner as possible with a clear indication of what components in the explanation are operating as premises, what components are operating as conclusions, and what the inferential rationale is that is connecting the premises to the conclusions. Each explanation should therefore make explicit the rational criteria for its determination whether this be, for example, global inferences drawn from the population-based reasoning of a demographic analysis or more locally or instance-based inferences drawn from the indication of feature significance by a proxy model. In all cases, the optimisation criteria of the operative algorithmic system should be specified, made explicit, and connected to the logic and rationale of the decision.- Each explanation should make available the records and activity-monitoring results that the design and development processes of your AI project yielded. Building this link between the process transparency dimension of your project and its outcome transparency will help to make its result, as a whole, more sufficiently interpretable. This can be done by simply linking or including the public-facing component of the process log of your PBG Framework.- Each explanation provided to an implementer should come with a standard Implementation Disclaimer that may read as follows:- Each explanation should specify and make explicit its governing performance metrics together with the acceptability criteria used to select those metrics and any standard benchmarks followed in establishing that criteria. Where appropriate and possible, fuller information about model validation measurement (including confusion matrix and ROC curve results) and any external validation results should be made available.- Each explanation should provide confirmatory information that the formal fairness criteria specified in your project’s Fairness Policy Statement has been met.- Each explanation should include clear representations of confidence intervals and error bars. These certainty estimates should make as quantitatively explicit as possible the confidence range of specific predictions, so that users and decision subjects can more fully understand their reliability and the levels of uncertainty surrounding them.- When an explanation offers categorically ordered scores (for instance, risk scores on a scale of 1 to 10), that explanation must also explicitly indicate the actual raw numerical probabilities for the labels (predicted outcomes) that have been placed into those categories. This will help your delivery process avoid producing confusion about the relative magnitudes of the categorical groupings under which the various scores fall. Information should also be provided about the relative distances between the risk scores of specific cases if the risk categories under which they are placed are unevenly distributed. It may be possible, for example, for two cases, which fall under the same high risk category (say, 9) to be farther apart in terms of the actual values of their risk probabilities than two other cases in two different categories (say 1 and 4). This may be misleading to the user.- Each explanation should, where possible, include a counterfactual explanatory tool, so that implementers and affected individuals have the opportunity to gain a better contrastive understanding of the logic of the outcome and its alternative possibilities.Social component of responsible implementation: We have now established the first step in the delivery of a responsible implementation process: making clear the rationale behind the technical content of an algorithmic model’s statistical results and determining how best to convey and communicate it so that these results may be appropriately treated as decision supporting and clarifying reasons. This leaves us with a second related task of content clarification, which is only implicit in the first step but must be made explicit and treated reflectively in a second.Beyond translating statistically expressed results into humanly significant reasons, you will have to make sure that their _socially meaningful content _is clarified by implementers, so that they are able to thoughtfully apply these results to the real human lives they impact in terms of the specific societal and individual contexts in which those lives are situated.This will involve explicitly translating that model’s technical machinery—its input and output variables, parameters, and functional rationale—back into the everyday language of the humanly relevant meanings, categories, and relationships that informed the formulation of its purpose, objectives, and intended elements of design in the first place. It will also involve training and preparing implementers to intentionally assist in carrying out this translation in each particular case, so that due regard for the dignity The human right to be valued and treated with respect because of one's personhood. of decision subjects can be supported by the interpretive charity, reasonableness, empathy, and context-specificity of the determination of the outcomes that affect them.Only through this re-translation will the internals, mechanisms, and output of the model become _useably interpretable _by implementers: Only then will they be able to apply input features of relevance to the specific situations and attributes of decision subjects. Only then will they be able to critically assess the manner of inference-making that led to its conclusion. And only then will they be able to adequately weigh the normative considerations (such as prioritising public interest or safeguarding individual well-being) that factored into thesystem’s original objectives.Having clarified the socially meaningful content of the model’s results, the implementer will be able to more readily apply its evidentiary contribution to a more holistic and wide-angled consideration of the particular circumstances of the decision subject while, at the same time, weighing these circumstances against the greater purpose of the algorithmically assisted assessment. It is important to note here that the understanding enabled by the clarification of the social context and stakes of an algorithmically supported decision-making process goes hand-in-glove with fuller considerations of the moral justifiability of the outcome of that process.A good starting point for considering how to integrate this clarification of the socially meaningful content of an algorithmic model’s output into your map of the delivery process is to consider what you might think of as your AI project’s content lifecycle.The content lifecycle: The output of an algorithmic system does not begin and end with the computation. Rather, it begins with the very human purposes, ideas, and initiatives that lay behind the conceptualisation and design of that system. Creating technology is a shared public activity, and it is animated by human objectives and beliefs. An algorithmic system is brought into the world as the result of this collective enterprise of ingenuity, intention, action, and collaboration.Human choices and values therefore punctuate the design and implementation of AI systems. These choices and values are inscribed in algorithmic models:- At the very inception of an AI project, human choices and values come into play when we formulate the goals and objectives to be achieved by our algorithmic technologies. They come into play when we define the optimal outcome of our use of such technologies and when we translate these goals and objectives into target variables and their measurable proxies.- Human choices and values come into play when decisions are made about the sufficiency, fit-for-purpose, representativeness, relevance, and appropriateness of the data sampled. They come into play in how we curate our data—in how we label, organise, and annotate them.- Such choices and values operate as well when we make decisions about how we craft a feature space—how we select or omit and aggregate or segregate attributes. Determinations of what is relevant, reasonable, desirable, or undesirable will factor into what kinds of inputs we are going to include in the processing and how we are going to group and separate them.- Moreover, the data points themselves are imbued with residua of human choices and values. They carry forward historical patterns of social and cultural activity that may contain configurations of discrimination, inequality, and marginalisation—configurations that must be thoughtfully and reflectively considered by implementers as they incorporate the analytics into their reasoned determinations.Whereas all of these human choices and values are translated in to the algorithmic systems we build, the responsible implementation of these systems requires that they be translated out. The rationale and logic behind an algorithmic model’s output can be properly understood as it affects the real existence of a decision subject only when we transform its variables, parameters, and analytical structures back into the human currency of values, choices, and norms that shaped the construction of its purpose, its intended design, and its optimisation logic from the start.It is only in virtue of this re-translation that an algorithmically supported outcome can afford stakeholders the degree of deliberation, dialogue, assessment, and mutual understanding that is necessary to make it fully comprehensible and justifiable to them. And, it is likewise only in virtue of this re-translation that the implementation process itself can, at once, secure end-to-end accountability and give due regard to the SUM values.The content lifecycle of algorithmic systems therefore has three phases: (1) The translation in of human purposes, values, and choices during the design process; (2) The digital processing of the quantified/mechanised proxies of these purposes, values, and choices in the statistical frame; (3) The translation out of the purposes, values, and choices in clarifying the socially meaningful content of the result as it affects the life of the decision subject through the implementation process. Here is a visualisation of these three phases of the content lifecycle: The translation rule: A beneficial result of framing the implementation process in terms of the content lifecycle is that it gives us a clear and context-sensitive measure by which to identify the explanatory needs of any given AI application. We can think of this measurement as the translation rule. It states that:What is _translated in _to an algorithmic system with regard to the human choices and societal values that determine its content and purpose is directly proportional to what, in terms of the explanatory needs of clarification and justification, must be _translated out.The translation rule organically makes two distinctions that have great bearing on the delivery process for responsible implementation. First, it divides the question of what needs explaining into two parts: (1) issues of socially meaningful content in need of clarification (i.e., the explanatory need that comes from the translation in to the AI model of the categories, meanings, and relations that originate in social practices, beliefs, and intentions) (2) issues of normative rightness in need of justification (i.e. the explanatory need that comes from translation in to the AI model of choices and considerations that have bearing on its ethical permissibility, discriminatory non-harm, and public trustworthiness). These two parts line up with what we have above called interpretable AI and justifiable AI respectively, and what we have also identified as tasks 2 and 3 of delivering transparent AI.Secondly, the translation rule divides the two dimensions of translation (translation in and translation out) into aspects of intention-in-design and intention-in-application. Translating in _has to do with _intention-in-design. It involves an active awareness of the human purposes, objectives, and intentions that factor into the construction of AI systems. Translating out, _on the other hand, has directly to do with _intention-in-application, or put differently, the intentional dimension of the implementation of an AI system by a user in a specific context and with direct consequences for a subject affected by its outcome.In human beings, intention-in-design and intention-in-application are united in intelligent action, and it is precisely this unity that enables people to reciprocally hold each other accountable for the consequences of what they say and what they do. By contrast, in artificial intelligence systems, which fulfil surrogate cognitive functions in society but are themselves neither intentional nor accountable, design and application are divided. In these systems, intention-in-design and intention-in-application are and must remain punctuation points of human involvement and responsibility _that manifest on either side of the vacant mechanisms of data processing. This is why translation is so important, and this is why enabling the implementer’s capacity to _intentionally translate out the social and normative content _of the model’s results is such a critical element of the responsible delivery of your AI project.It might be helpful to think more concretely about the translation rule by considering it in action. Let’s compare two hypothetical examples: (1) a use case about an early cancer detection system in radiomics (a machine learning application that uses high throughput computing to identify features of pathology that are undetectable to the trained radiological eye); and (2) a use case about a predictive risk assessment application that supports decision- making in child social care.In the radiomics case, the _translating in _dimension involves minimal social content: the clinical goal inscribed in the model’s objective is that of lesion detection and the features of relevance are largely voxels extracted from PET and CT scanner images. However, the normative aspect of _translating in _is, in this case, significant. Ethical considerations about looking after patient wellbeing and clinical safety are paramount and wider justice concerns about improving healthcare for all and health equity factor in as well.The explanatory needs of the physician/implementer receiving clinical decision support and of the clinical decision subject will thus lean less heavily on the dimension of the clarification of socially meaningful content than it will on the normative dimension of justifying the safety of the system, the priority of the patient’s wellbeing, and the issues of improved delivery and equitable access. The technical content of the decision support may be crucial here (Issues surrounding the reproducibility of the results and the robustness of the system may, in fact, be of great concern in the assessment of the validity of the outcome.), but the _translating out _component of the implementation remains directly proportional to the minimal social content and to the substantial ethical concerns and objectives that were _translated in _and that thus inform the explanatory and justificatory needs of the result in general.The explanatory demands in the child social care risk assessment use case are entirely different. The social content of the _translating in _dimension is intricate, multi-layered, and extensive. The chosen target variable may be child safety or the prevention of severe mistreatment and the measurable proxy, home removal of at-risk children within a certain timeframe. Selected features that are deemed relevant may include the age of the at-risk children, public health records, previous referrals, family history of violent crime, welfare records, juvenile criminal records, demographic information, and mental health records. Complex socioeconomic and cultural formations may additionally influence the representativeness and quality of the dataset as well as the substance of the data itself.The normative aspect of _translating in _here is also subtle and complicated. Ethical considerations about protecting the welfare of children at risk are combined with concerns that parents and guardians be treated fairly and without discrimination. Objectives of providing evidence-based decision support are also driven by hopes that accurate results and well-reasoned determinations will preserve the integrity and sanctity of familial relations where just, safe, and appropriate. Other goals and purposes may be at play as well such as making an overburdened system of service provision more efficient or accelerating real-time decision-making without harming the quality of the decisions themselves.In this case of predictive risk assessment, the _translating out _burdens of the frontline social worker are immense both in terms of clarifying content and in terms of moral justification. If, for example, analytical results yielding a high risk score were based on the relative feature importance of demographic information, welfare records, mental health records, and criminal history, the implementer would have to scrutinise the particular decision subject’s situation, so that the socially meaningful content of these factors could be clarified in terms of the living context, relevant relationships, and behavioural patterns of the stakeholders directly affected. Only then could the features of relevance be thoroughly and deliberatively assessed.The effective interpretability of the model’s result would, in this case, heavily depend on the implementer’s ability to apply domain-knowledge in order to reconstruct the meaningful social formations, intentions, and relationships that constituted the concrete form of life in which the predictive risk modelling applies. The implementer’s well-reasoned decision here would involve a careful weighing of this socially clarified content against the wider predictive patterns in the data distribution yielded by the model’s results—patterns that may have otherwise gone unnoticed.Such a weighing process would, in turn, be informed by the normative-explanatory need to translate out the morally implicating choices, concerns, and objectives that influenced and informed the predictive risk assessment model’s development in the first place. Again, the interpretive burden of the frontline social worker would be immense here. First, this implementer would have to deliberate with a critically informed awareness of the legacies of discrimination and inequity that tend to feed forward in the kinds of evidentiary sources drawn upon by the analytics. Such an active reflexivity is crucial for retaining the punctuating role of human involvement and responsibility in these sensitive and high-stakes environments.Just as importantly, the frontline social worker would have to evaluate the real impact of ethical objectives at the point of delivery. Not only would the results of the analytics have to be aligned with the ethical concerns and purposes that fostered the construction of the model, this implementer would have to reflectively align their own potentially diverging ethical point of view both with those results and with those objectives. This _normative triangulation _between the original intention-in-design, the implementer’s intention-in- application, and the content clarification of the AI system’s results is, in fact, a crucial safeguard to the delivery of justifiable AI. It again enables a reanimation of moral involvement and responsibility at the most critical juncture of the content lifecycle.Step 3: Build an ethical implementation platform:
1. Train ethical implementation. The continuous challenges of translation, content clarification, and normative explanation should inform how you set up your implementation training to achieve optimal outcome transparency. In addition to the necessary training to prevent implementation biases in the users of your AI system (discussed above), you should prepare and train the implementers to be stewards of interpretable and justifiable AI. This entails that they be able to:- Rationally evaluate and critically assess the logic and rationale behind the outputs of the AI systems;- Convey and communicate their algorithmically assisted decisions to the individuals affected by them in plain language. This includes explaining to them in an everyday, non-technical, and accessible way how and why the decision-supporting model performed the way it did in a specific context and how that result factored into the final outcome of the implementation;- Apply the conclusions reached by the AI model to a more focused consideration of the particular social circumstances and life context of the decision subject and other affected parties;- Treat the inferences drawn from the results of the model’s computation as evidentiary contributions to a broader, more rounded, and coherent understanding of the individual situations of the decision subject and other affected parties;- Weigh the interpretive understanding gained by integrating the model’s insights into this rounded picture of the life context of the decision subject against the greater purpose and societal objective of the algorithmically assisted assessment;- Justify the ethical permissibility, the discriminatory non-harm, and the public trustworthiness both of the AI system’s outcome and of the processes behind its design and use
1. Make your implementation platform a relevant part and capstone of the sustainability track of your project. An important element of gauging the impacts of your AI technology on the individuals and communities it touches is having access to the frontlines of its potentially transformative and long-term effects. Your implementation platform should assist you in gaining this access by being a _two-way medium of application and communication. It should both enable you to sustainably achieve the objectives and goals you set for your project through responsible implementation, but it should also be a sounding board as well as a site for feedback and cooperative sense-checking about the real-life effects of your system’s use. Your implementation platform should be dialogically and collaboratively connected to the stakeholders it effects. It should be bound to the communities it serves as part of a shared project to advance their immediate and long-run wellbeing.3. Provide a model sheet to implementers and establish protocols for implementation reporting. As part of the roll-out of your AI project, you should prepare a summary/model sheet for implementers, which includes summation information about the system’s technical specifications and all of the relevant details indicated above in the section on substance of the technical content to be delivered. This should include relevant information about performance metrics, formal fairness criteria and validation, the implementation disclaimer, links or summaries to the relevant information from the process logs of your PBG Framework, and links or summary information from the Stakeholder Impact Assessment.You should also set up protocols for implementation reporting that are proportional to the potential impacts and risks of the system’s use.4. Foster outcome understanding through dialogue. Perhaps the single most important aspect of building a platform for ethical implementation is the awareness that the realisation of interpretable and justifiable AI is a dialogical and collaborative effort. Because all types of explanation are mediated by language, each and every explanatory effort is a participatory enterprise where understanding can be reached only through acts of communication. The interpretability and justifiability of AI systems depend on this shared human capacity to give and ask for reasons in the ends of reaching mutual understanding. Implementers and decision subjects are, in this respect, first and foremost participants in an explanatory dialogue, and the success of their exchange will hinge both on a reciprocal readiness take the other’s perspective and on a willingness to enlarge their respective mental models in accordance with new, communicatively achieved, insights and understandings.For these reasons, your implementation platform should encourage open, mutually respectful, sincere, and well-informed dialogue. Reasons from all affected voices must be heard and considered as demands for explanation arise, and manners of response and expression should remain clear, straightforward, and optimally accessible. Deliberations that have been inclusive, unfettered, and impartial tend to generate new ideas and insights as well as better and more inferentially sound conclusions, so approaching the interpretability and justifiability of your AI project in this manner will not only advance its responsible implementation, it will likely encourage further improvements in its design, delivery, and performance.Leslie, 2019, p.54-68