In January 2024, I received an email from a journalist at The Times (London) asking to have a chat about machine translation. He was working on a piece about the Live Translate feature that was about to be released on the new Samsung Galaxy phone. The feature works like the AI version of a phone interpreter. It uses a voice agent – like Amazon’s Alexa or Apple’s Siri – to interpret for the caller and the other participant on the call. The journalist asked what language other than English I preferred because he wanted to speak to me using the feature. I told him ‘Portuguese’. When he called, the voice agent opened the conversation by announcing that the call was going to be interpreted with the use of AI. I could then carry on speaking Portuguese and the journalist could carry on in English.
Or at least that was the idea. The test was a complete failure. There was a delay in hearing the interpretations, and the journalist and I ended up speaking over each other. Our false starts and interruptions confused the AI model. We gave up after a few moments because we could not understand what each other was saying. At the time of writing more than a year later, it seems there is a function that enables callers to mute their interlocutor’s voice so that they only hear the interpreted version of what their interlocutor says. This function might not have been available at the time or perhaps the journalist did not select it. Such a setting may go some way towards addressing the problem we experienced, but the fact remains that, in that instance, the tool did not fulfil its intended purpose.
We had a debrief in English after the test call. The journalist told me that he had been calling hotels in different countries using the feature. Staff at these hotels had been hanging up on him because the test was going just as badly with them as it did with me. The journalist then wrote his piece, which struck a humorous tone about the tool’s mistakes.1 In our conversation, I remember thinking about the scenario selected for the tests: a tourist calling a hotel. This is in many ways a typical use for machine translation. Indeed, Samsung envisaged a similar use scenario for the feature. In a flashy marketing video, screen captions suggest: ‘Say you’re traveling abroad and you need to call a taxi. Not sure how to say it? It’s time to use Live Translate.’2 The Times journalist and Samsung tapped into perceptions of machine translation as a technology that can come in handy for travellers while performing unremarkable low-risk tasks. But what if instead of calling a taxi or a hotel, the user needs to call the police? Or an ambulance?
The developers would probably say that their tool could be useful in these scenarios too. But they might add, as Samsung does, that they do not ‘make any promises, assurances or guarantees as to the accuracy, completeness or reliability of the output provided by AI features’.3 Apple goes a step further and says that their machine translation app ‘should not be relied on in circumstances where you could be harmed or injured, in high-risk situations, for navigation, or for the diagnosis or treatment of any medical condition’.4 As mentioned in Chapter 1, this type of statement mostly exempts technology developers from liability. But it is also true that statements like Apple’s, and marketing suggestions like Samsung’s, reflect how the developers – or least their legal teams – expect their products to be used.
As the examples discussed so far demonstrate, many uses of machine translation do not fit these expectations. Irrespective of the plans that developers chart for their products, their technologies can always be taken down unintended paths. This possibility raises important questions around accountability, human agency and ethics. This chapter engages with some of these questions. I draw on moral-philosophical theories to examine uses of machine translation as ethically imbued communicative acts. The chapter has three focal points. First, it looks at the meaning of ‘use’, which I interrogate by drawing on post-phenomenology concepts linked to the notion of technological mediation. Second, it examines the potential for technologies themselves to influence human behaviour. Lastly, the chapter applies ethical theories to the question of machine translation use. I look briefly at the use of consequences and rules as guiding moral principles. I then focus on virtue ethics, a paradigm whose flexibility, I would argue, is well suited to the decision-making challenges posed by high-risk uses of machine translation.
Human–Technology Relations
The concept of usability reviewed in Chapter 1 underpins much previous work on the social, economic and cognitive consequences of machine translation. Scholars are often interested in how satisfied users are with machine translation and in how effective it is for different tasks. The type of user under analysis will often reflect the corresponding disciplinary field. In education, teachers and students, and in some cases students’ families, are common user categories. Studies of this nature have looked at the efficacy of using machine translation as a tool for coursework writing5 and for communicating with parents who do not speak the language of instruction.6 Uses of machine translation by doctors and patients have been studied in medical research,7 and similar discussions have taken place in relation to legal services8 and patent processing.9 As mentioned in the Introduction, human translators have to date been a common user category in translation studies. A sizeable body of work is available on what human translators think of machine translation and on how they use it.10 The relationship between machine translation and human translators in this type of work is often one of instrument and user. This relationship can also be described in terms of interaction,11 especially if translators use machine translation systems that learn from human edits on the fly.12
Describing uses of machine translation as interaction makes sense. Interactional relations are increasingly evident in cases where AI agents mediate machine-translated communication, whether in writing or with the use of speech. Editing the original message or the resulting translation output to improve the readability of the message can also be described as a type of interaction. However, the notions of interacting with or simply using an instrument are not enough to capture the complexity of machine-translated communication. The public service contexts discussed in this book call for a broader analysis.
In an article titled ‘Beyond interaction’, Peter-Paul Verbeek mentions how the concept of interaction often presupposes a relationship between two distinct entities.13 Usually, these entities are ‘human subjects and technological objects, between which there is a specific kind of activity’ (emphasis in original).14 As discussed by Verbeek, a closer examination will reveal that this relationship is not that simple. First, the difference between humans and technologies is not this clear-cut. Within the framework of post-phenomenology and technological mediation,15 humans and technologies are thought to feed off each other. Technologies frame our interpretation of the world, which in turn frames and shapes technologies themselves. Technologies are part of us. They augment our senses and cognition, so the ‘human subjects’ at one side of the interaction are not purely human or completely detached from technologies, just as the ‘technological objects’ are not detached from humans.
Moreover, interaction is ill-suited to describe what happens in some contexts of use. Consider a website with integrated machine translation. Those who avail themselves of the machine translation feature will often just browse the website as normal. They may be required to select a language on the website menu, but they do not directly operate a machine translation tool. Machine translation here acts as a sort of language filter. It is an interpretative window into the content, or a device that provides a ‘reading’ of the original. In the post-phenomenology framework, this type of reading can be classed as a hermeneutic (i.e., interpretative) human–technology relation.16
Post-phenomenology researchers have examined several other types of human–technology relations. Embodiment refers to uses of technologies as conduits for human action or extensions of the human body, such as a hammer or a digital hand-held device.17 Fusion refers to situations where humans and technology are integrated, for example through the use of brain implants.18 Immersion refers to technologies that are embedded in the environment in ways that can analyse and ultimately influence human behaviour, such as sensors that may remind us to close the fridge door or fasten our seatbelts.19 I have discussed these and other human–technology relations in previous work.20 These relations question simplistic notions of technology ‘use’.
This type of questioning matters for this book because use and usability frame much of the socially oriented literature on machine translation to date. The goal of researchers has often been to examine what users need, how they interact with machine translation and how satisfied they are with it so that the technology, or its deployment, can be improved or better understood. Some of this work is cited in Chapter 1. In a language industry context, research on usability and human–computer interaction has led to new methods of integrating machine translation into the human translation process and new computer-assisted translation tools (i.e., professional text editing software used by human translators in their work).21 This type of research is therefore crucial. But assumptions that there will always be a clear-cut distinction between the user and the object – or a direct interaction between them – understate the complexity of machine translation’s presence in public service contexts.
Take the example from Chapter 1 where police officers used Google Translate to communicate with a Spanish speaker at a bus station. At one end of the communication were the officers together with Google Translate. At the other end was the man, who had to assimilate (understand, consider) the translations provided by Google as well as what the officers were trying to convey with their use of gestures, simplified English or their limited knowledge of Spanish. There are multiple layers of ‘use’ in such a context. The officers were the first ones to deploy Google Translate, so I will call them primary users. The man being questioned engaged with machine translation through the officers’ use of it, so I will call the man a secondary user. In his case, the lack of a clear-cut distinction between user and object is particularly clear. The man did not interact just with Google Translate, but rather with a product of the interaction between Google Translate and the officers.
For secondary users, engaging with machine translation is not necessarily a deliberate decision. The technology in this case may instead be a tool that is deployed by others or one that is embedded in a specific environment, such as at a border check point or on a website. The notion of use is therefore highly complex. It does not mean just interaction and it does not presuppose strict dichotomies between human and non-human. Moreover, machine translation use is neither static nor unidimensional. Secondary users have a different type of relationship with machine translation tools than primary users, and all these relationships may have elements of immersion, hermeneutics and embodiment, among other types. Primary users do not make decisions in a vacuum either. When the primary users are public service professionals, many of the organisational factors discussed in Chapter 1 will weigh on their decision to use the technology. In these cases, the machine translation use starts as a top-down decision that cascades through other layers of use, thereby generating different types of human–technology relations involving service users and providers.
The Persuasive Power of Technology
As discussed at the beginning of the chapter, technology developers cannot unilaterally determine how their products are used. Their finite ability to determine the future of their products, and the power of users in shaping this future, are important aspects of the social construction of technology.22 Supporters of this paradigm argue that technologies are not static artefacts, but rather socially co-constructed entities that are a product of how they are used and interpreted by different groups of people.23 In other words, the way in which a technology evolves is both flexible and socially dependent. Recognising this process of social co-construction does not exempt the designer from accountability, but it foregrounds the different processes through which a technology becomes something of value in people’s lives.
The focus of the social construction of technology has originally been on different social groups and how they can influence the ways in which technologies change. Cognate sociological theories also posit that technologies themselves can embody specific worldviews and thereby exert their own influence over human beings.24 Sometimes a tool’s influencing power is central to its functionality. For example, the algorithms deployed by video streaming platforms are supposed to learn from our viewing habits and suggest content that has been automatically calculated to match our preferences.25 The tool is programmed to make selections that are supposed to influence our choices about what to watch. Extensive research is available on these types of algorithmic decisions, their swaying power over human beings and their short- and long-term consequences.26 For the discussion in this book, what matters most is the fact that technologies can have an influencing effect on humans. Although machine translation tools may not directly try to influence what we watch on TV or the type of products we decide to purchase, they have their own type of influencing power.
Despite the voices trying to claim that AI is now sentient,27 most people would probably agree that technologies are not capable of feeling, cannot think for themselves and cannot make conscious decisions. Technologies therefore do not have free will. But saying that technologies can influence our decision-making is not the same as saying that they can think. It is to say that technologies – through their use possibilities and physical presence – may predispose humans to behave in a particular way.
Consider a traffic police officer who in December 2018 was keeping watch on an interstate highway in Pennsylvania. The officer decided to stop someone who was driving too close to the vehicle in front. After being pulled over, the driver indicated that he could only speak Spanish. In a move that by this point sounds familiar, the officer decided to speak to the driver using Google Translate. The conversation went well until the point where the officer asked the driver about his itinerary:
Trooper Hoy [the officer] asked him [the driver] about his travel plans. During this portion of the [mobile video recording], Carmenates [the driver] is heard giving lengthy responses in Spanish to Trooper Hoy’s questions. However, many of Carmenates’ responses were not translated by Google Translate at all and Google Translate translated some lengthy responses as short, nonsensical English statements, including the statement ‘you already see the see a bear for the girl the suitcase with the coat over coat’.28
Since Google Translate was making obvious errors, what ensues is perhaps surprising even though it again sounds familiar. Although the officer had a consent form in Spanish, he continued using Google Translate to ask for consent to inspect the driver’s luggage. Google Translate here was not used for lack of a practical alternative. It was the preferred method of communication:
Trooper Hoy had copies of a written ‘consent to search’ form already translated into Spanish in his vehicle. Nevertheless, approximately 12 minutes into the traffic stop, Trooper Hoy chose to use Google Translate to obtain Carmenates’ consent to ‘see’ his luggage.29
What leads a police officer to insist on using Google Translate even when it is making obvious errors? A combination of factors is at play in situations of this nature, not least the individuals concerned and the quality of the training they may have received. But I would argue that the affordances of the technology – its characteristics and use possibilities – cannot be ignored. Convenience is persuasive, and consumer technologies are designed to be convenient. The point of a smartphone is to allow users to access a range of useful features on the go. Similarly, machine translation was invented in great part to provide fast translations that are easy to obtain.30 It is supposed to make users’ lives easier, so consciously avoiding it in situations where it seems convenient takes effort. It calls for a type of critical resistance that can go under the heading of AI literacy. In this example, the officer was already using machine translation when he decided to ask for consent for the search. Changing tack at that point would have required resisting not only the convenience of the technology but also the inertia of the status quo.31
The decision to continue using machine translation had consequences. The officer eventually found large quantities of marijuana in the driver’s luggage. Like in similar cases discussed so far, the driver’s legal defence team filed a motion to supress the evidence alleging a lack of informed consent for the search. The court granted the motion,32 but the state of Pennsylvania appealed this decision. A higher court then re-examined the case and again found in favour of the driver. The language barrier, among other reasons, affected the driver’s ability to ‘knowingly, intelligently, and voluntarily consent to the search’.33 The use of Google Translate therefore compromised the state’s ability to prosecute the driver as well as the driver’s right to understand the question he was being asked.
Philosophy of technology scholars use the term ‘intentionality’ or ‘technological intentionality’ to refer to the way in which a technology may predispose humans to act in a specific way.34 Intentionality in this context does not mean having a conscious plan or intention.35 Rather, it refers to the fact that technologies may be directed at some type of goal or activity. They exist for a reason and fit specific types of behaviour. I prefer the term ‘influence’ over ‘intentionality’ because ‘intentionality’ can be easily confused with ‘purposeful action’, which technologies are not capable of. But this notion of being ‘directed towards’ something36 is central to the potentially persuasive power of technologies and thereby of machine translation. In the example above, it can be speculated that had a machine translation tool not been readily available, the officer would have tried to communicate with the driver via other means, possibly by using the consent form in Spanish or seeking assistance from a human interpreter. As discussed in Chapter 1, these possibilities are not risk-free. But the important point here is that the existence of an accessible, cheap and easy-to-use technology enabled the interaction to unfold in the way that it did even if the officer was free to make different choices. Moreover, the fact that the technology is accessible, cheap and easy to use is not an accident. It is, rather, central to the business model of corporations which rely on our use of their products, whether it be to generate data, to hone their advertising strategies (e.g., by using machine translation to allow us to consume more content online) or more broadly to normalise their technologies as a feature of everyday living.
One other way of framing a similar argument is to say that technologies are value-laden.37 The motto ‘Guns don’t kill people, people kill’38 is widely rejected by researchers who argue that there is something inherent to guns and their function which, when coupled with a human, gives rise to a lethal combination. In the technological mediation framework, it is this combination, rather than just the gun or just the human, that kills in the act of shooting.39 This example is well worn, but it does bring home the fact that this is not an abstract discussion. Failing to recognise that technological artefacts can influence human action may mark the difference between loosening or strengthening gun control laws, for instance.
It goes without saying that machine translation systems are in many ways unlike guns. But the technological mediation approach to the question of technological influence, while not flawless,40 is instructive. The fact that human existence is technology-mediated means that humans and technology cannot be completely disentangled. If, however, technologies evolve through a process of social co-construction, and if they can exert their own type of influence over humans, a reasonable question to ask is, who is accountable when something goes wrong? At a practical level the answer to this question is as dynamic as the communicative contexts in which machine translation is used. But if technologies cannot be disentangled from human existence, it follows that humans are in no way exempt from responsibility. Technological agency – by which I mean technology’s capacity to influence human action – exists because of human agency.41 Recognising the persuasive convenience of machine translation does not mean that its designers, deployers or users are unaccountable.
Designers, for one, are not detached from political and economic value systems. They will imbue their products with specific values even if they cannot unilaterally control how the product is used. Deployers, users and consumers will also be subject to specific value systems, which will influence their relationship with the technologies. Political and economic environments where efficiency is a core value may be more vulnerable to the appeal of cheap and convenient tools. It is by recognising and anticipating the persuasive power of technology in these cases that human agency is responsibly exercised. Designers need to anticipate the paths their products are likely to take, as argued in many discussions of technology ethics.42 Institutional and professional deployers of machine translation also need to anticipate the powerful appeal of the technology, which they can do in many ways including through education and workplace training. It is by recognising the influencing power of technologies that this power can, where necessary, be resisted.
There is therefore nothing inevitable about the way in which machine translation permeates the communicative contexts discussed in this book. These tools are not an outer force against which humans are powerless. They are a product of us. Their availability can influence human decisions, but it does not determine human behaviour.43 It instead calls for considered approaches to multilingual communication that are cognisant of how individuals may instinctively call on machine translation tools in ways that may be unsuitable for the context.
Machine Translation and the Virtues
Ethical theories can shed light on important factors to consider in anticipating the influencing power of machine translation tools. Most discussions of normative ethics to date have revolved around three groups of theories: consequentialism, deontology and virtue ethics. Some aspects of these theories have been discussed in the translation studies literature, mostly in terms of how they may apply to translators’ and interpreters’ decision-making.44 Implications of machine translation for professional language services, and for meaning making and communication, have also been examined from an ethical viewpoint.45 Analyses of how existing ethical frameworks may be applied to machine translation use in the public service contexts discussed here, on the other hand, are for the most part missing. I briefly examine consequentialism, deontology and virtue ethics in the sections below to identify principles that can guide discussions of machine translation’s ethical dilemmas. I note that covering all conceptual intricacies of these theoretical frameworks is not my objective. My goal is rather to highlight how prescriptive models of behaviour on their own are often unfit to support the nuanced decisions that machine translation use will usually require.
Consequentialism
Ethical theories focused on the consequences of human decisions can be collectively referred to as consequentialism. The most well-known group of consequentialist ethical theories is utilitarianism.46 Several examples have been used in the literature to illustrate utilitarian approaches to decision-making. The widely discussed trolley problem47 describes a situation in which a tram can either run over five people if it continues on the same track, or it can be diverted to a different track where it will hit one person. It is not possible to stop the tram, so the choice is between actively causing one death by changing tracks or allowing the tram to run its original course and kill five people.48 This extreme scenario has become a cliché, especially in debates about driverless cars.49 The technology ethicist Cennydd Bowles calls the trolley problem a red herring because it can divert attention to hypothetical scenarios while neglecting real problems such as data bias50 and, I would add, the privacy implications of technology use, which I discuss in Chapter 5. The trolley problem nevertheless does a good job of illustrating choices where none of the possible outcomes is inherently desirable. It also helps to illustrate different ways of framing the rationale for specific actions. One of the possible utilitarian ways of looking at this problem would be to divert the tram because this decision would save the largest number of lives even if it would require actively causing one death.51
Not all utilitarians would agree with this decision. There are different versions of utilitarianism depending on the values that are considered more important in seeking resolutions to ethical dilemmas and on how these values are measured.52 For example, utilitarianism can be associated with specific actions directly, known as act-utilitarianism, or with specific actions as well as their broader guiding principles or rules, known as rule-utilitarianism.53 Preferential utilitarianism places no emphasis on intrinsic values but rather on the preferred outcomes for most of the people involved.54 For example, if most individuals on earth want to protect rainforests, then one of the possible interpretations of preferential utilitarianism would be that protecting rainforests is good because the outcome of doing so is desired by the majority and not because this outcome is associated with happiness or some other value.55 Hedonist utilitarianism, on the other hand, favours actions that lead to happiness.56 A much deeper discussion would be necessary to evaluate all different facets of consequentialism, and of utilitarianism more specifically. My intentions here are simply to illustrate some of the challenges of adopting a strictly consequentialist approach to the risks of using machine translation.
In the context of machine translation use, consequentialism can be examined in relation to the inherent inconstancy of language and of machine translation systems. In a study published in 2021, volunteers were asked to assess Google translations of hospital discharge instructions.57 English was the original language, and translations in seven languages were assessed: Spanish, Chinese, Vietnamese, Tagalog, Korean, Armenian and Farsi. Bilingual volunteers assessed the translations using criteria such as fluency (if the translations were grammatically correct and easy to understand and to read) and adequacy (how much of the original meaning was retained). The results showed that ninety per cent of translations into Tagalog, an Austronesian language spoken in the Philippines,58 were considered accurate.59 For the sake of argument, let us assume that these assessments indicate how well this machine translation system is likely to perform in future translations of discharge instructions into Tagalog at the same hospital. Let us also assume that alternative or complementary language assistance has been difficult to obtain. For example, the hospital might have been unable to locate professional Tagalog translators, and no member of staff can speak the language. The hospital might therefore see in Google Translate an opportunity to address this issue.
If we focus strictly on outcomes with the aim of generating the greatest amount of good to the greatest number of people, adopting Google Translate could be seen as favourable in this case. If, again, it is assumed that the assessment ratings have some predictive value, only ten per cent of the Tagalog translations will be inaccurate. Most Tagalog translations are therefore likely to be beneficial even if this might involve giving inaccurate advice to some patients. Moreover, leaving the instructions untranslated could lead to negative health outcomes for a larger number of patients than those who would be negatively affected by mistranslations.
Consequentialist approaches to ethics may therefore lead to the conclusion that some harmful interventions can be justified if they lead to some type of greater good. This rationale is applied even in medical contexts where lives may be at risk. Controlled trials often report adverse reactions to medical interventions, although in medical settings the benefits need to clearly outweigh the risks for the risks to be considered acceptable. Take the Oxford–AstraZeneca vaccine developed during the Covid-19 pandemic. This vaccine was associated with a severe type of blood clotting disorder. The blood clot events were ‘very rare’, a specific medical classification that means their chance of occurring was less than one in ten thousand.60 There was wide consensus in the scientific community that the risks posed by the Covid-19 virus itself were much greater than the risk posed by the vaccine.61 While some individuals suffered fatal blood clots after vaccination,62 the vaccine went on to save 6.3 million lives in the first year of its rollout.63 The fact that the vaccine’s benefits outweighed its risks is therefore not in question.
However, many factors would need to be considered in such an analysis if it were to be applied to translation, including the quality of alternative or complementary communication methods, whether alternatives were available in the first place and the severity of the translation errors. Some errors may just cause unnecessary alarm and prompt further action that would eventually clear up the confusion. Other errors may omit or distort critical information and lead to eventual harms. Additionally, language is difficult to control. The fact that a machine translation system performs well ninety per cent of the time during a test – in any event an arguably low level of accuracy for medical settings – does not guarantee that this performance will be sustained. The content being translated is likely to change significantly, as is the translation system itself if it is an online tool. The risk of providing untranslated instructions is also difficult to calculate. Some patients may have a limited level of language proficiency that would allow them to read the untranslated instructions, perhaps with the help of bilingual dictionaries and other tools. Others may be able to rely on family and friends, which is not ideal but probably safer than simply not understanding the instructions or, worse, being misinformed by inaccurate translations.
Risk-benefit ratios will therefore be subject to significant fluctuation, so making overarching risk-benefit assessments to be rolled out at an institution is not straightforward. While not necessarily all mistranslations would lead to an adverse health event, doctors who are familiar with the risks of inaccuracies, and with the dynamic nature of language, may argue that the uncertainty of machine translation is too high to be considered acceptable for use in discharge instructions. If machine translation is used in these cases without a critical consideration of the risks, some doctors could see this as a breach of duty.
Deontology
A duty to a specific set of norms or rules is central to the concept of deontology.64 In deontological ethics, doing what is right is the main goal even when this might not bring about the greatest amount of good for the people involved.65 For example, when faced with the trolley problem described above, one may favour leaving the trolley on the original track because causing someone’s death could be argued to be intrinsically wrong. Deontology is usually associated with Kantian moral philosophy.66 In the Kantian tradition, the concept of a ‘universal law’ is often invoked in proposals for how individuals can assess the morality of their actions. In this paradigm, individuals should only behave in a manner consistent with what they would wish to be a universal law applied to all.67
As implied by the mention of rule-utilitarianism in the previous section, there is more crossover between ethical theories than it would perhaps seem on first impression. Sometimes different theories may lead to similar outcomes. Different theoretical traditions can also draw on similar concepts, such as rules or virtue.68 But in the contexts discussed here, public service providers could be said to be following a predominantly deontological approach to ethics if, for instance, they shaped their machine translation use behaviour on prescriptive policies or professional codes of conduct that attempted to spell out what their duties are and the norms according to which they should act.
Primary care guidance issued by the UK’s National Health Service (NHS) in England has seven items under its ‘translation of documents’ principle. The seventh item under this principle states: ‘Automated on-line translating systems or services such as “Google-translate” should be avoided as there is no assurance of the quality of the translations.’69 Although this document was first published in 2018, this same item is quoted in more recent official guidance issued by the UK government, for instance in instructions last updated in 2021.70 For UK healthcare practitioners, refraining from using machine translation because of this guidance would be a rule-oriented approach to the question. Avoiding machine translation in this case can be considered right because translation quality cannot be assured and because this is the approach recommended by the relevant national body.
As this book makes clear, guidance of this nature is not stopping UK health professionals from using machine translation. Policies can be extremely useful, as I will discuss later, but there are several issues with this type of instruction. For one thing, in this specific example, the guidance lacks detail. Urging healthcare workers to avoid machine translation is not the same as forbidding it, so the guidance leaves open the question of what can be classed as a circumstance where machine translation cannot be avoided, and its use therefore justified. Moreover, this type of guidance can clash with other commitments which practitioners may be expected to honour. The oath made by new doctors graduating from the University of Exeter in the UK states that new doctors ‘will work towards a fairer distribution of health resources’.71 A similar oath used at the University of Edinburgh states that doctors ‘will treat all patients equally and without prejudice’.72 Could healthcare workers see a role for machine translation in ‘working towards a fairer distribution of resources’ and in ‘treating all patients equally’ if it means providing non-critical information to linguistic minorities in a timely and accessible manner? As will be discussed later in the book, many healthcare professionals would probably say so. The challenges posed by machine translation are therefore too complex to be fully addressed by prescriptive statements that simply try to curb its use.
Even flexible guidance can suffer from a similar problem. The Immigration Partnership for the Canadian Region of Waterloo has produced a language access checklist to be followed by local organisations.73 The Immigration Partnership is a collaborative initiative involving public service providers, businesses and local government.74 It is therefore unlike a central government agency, and the guidance it provides may well be seen as more in touch with the grassroots reality of the local community. The checklist starts with statements that seek to confirm that organisations have an interpreting policy in place and are committed to language services provision. The checklist then seeks to confirm that such a policy provides guidance on ‘When it is acceptable and not acceptable to use machine translation (e.g. Google Translate)’ (emphasis omitted).75 The policy therefore does not define what is acceptable. It just asks organisations to consider the question and adopt a clear position.
The importance of initiatives of this nature is not in dispute. Having a translation and interpreting policy that includes statements on machine translation will often be instrumental in ensuring that an organisation’s use of the technology is given due consideration. But in contexts marked by high stress and time pressure, it can be difficult to define ‘acceptable’. Even if organisations implement the guidance flexibly by listing principles to follow rather than specific uses of machine translation to avoid, the reality on the ground is likely to be dynamic. A policy may state, for instance, that machine translation should only be used in emergencies and if all alternative methods of communication have been exhausted. But what if a service provider has had long waits for professional interpreting in the past and is now in a situation where waiting could be detrimental? They may bypass attempts to use other methods in this case and move straight to machine translation. Even if it later becomes apparent that this was the wrong decision, the circumstances of the interaction may be unique and simply weigh more heavily on decision-making than the clauses of a workplace policy. The use of rules to decide when to use machine translation will therefore be limited by the challenges of proposing overarching definitions of what is right for communicative contexts that are marked by change.
Virtue Ethics
As moral principles, both rules and consequences have limitations. No ethical theory, in fact, is ever perfect. These theories tend to intertwine and bifurcate in ways that, for my objectives in this book, would not justify strict allegiance to any one of their specific versions. Nevertheless, recent discussions of the notion of virtue try to address the dynamic challenges posed by technology use in ways that resonate with some of the dilemmas posed by machine translation.76 Western versions of virtue ethics date back to Plato and Aristotle.77 Eastern traditions can be linked to Buddhism and Confucianism.78 Rather than focusing on adherence to a set of norms or on how actions may lead to a greater good, virtue ethics prioritises individuals’ ability to make the most appropriate decision for each context. Three concepts are central to virtue ethics: (1) virtue itself, (2) practical or moral wisdom and (3) happiness, which is also referred to as flourishing.79
The ‘virtue’ in virtue ethics is an individual’s disposition to act in a way that ultimately leads to happiness or flourishing.80 This is not the same as simply doing something virtuous or having a virtuous feeling since both of which are possible for a non-virtuous person.81 Virtue is not accidental either. It is a trait developed over time through experience and moral habituation.82 The technology philosopher Shannon Vallor describes moral habituation as dependent on a series of practices including ‘repeated moral practice of right (or nearly right) conduct’ (emphasis omitted) and ‘a practice that gradually accustoms the individual to actions which were previously seen as painful or unattractive’ (emphasis omitted).83 ‘Nearly right’ conduct is described as a goal of moral practice because there is always room for improving one’s virtues.84 A virtue is therefore a process. In the ethical-philosophical sense of the word, individuals do not happen to be virtuous through instinct or without trying. Being virtuous takes effort and moral training.
The second important concept within virtue ethics, moral wisdom, is underpinned by a distinction between an instinctive type of virtue, on the one hand, and the deeper, cultivated type of virtue described above, on the other. If virtue were taken to mean just an instinct to act in a given way, then this same instinct could lead to undesirable conduct.85 For example, an instinct to be helpful may lead someone to cave in to questionable requests to please the requester. Moral wisdom is what distinguishes instinctive or natural virtues from full virtue.86 Moral wisdom can also be broken down into a set of more specific practices. One of these practices is referred to as prudential judgement: ‘the cultivated habit of deliberating and choosing well, in particular situations, among the most appropriate and effective means available for achieving a noble or good end’ (emphasis omitted).87 I would argue that prudential judgement is central to decisions on whether and how to rely on machine translation in the contexts discussed in this book. But ‘a noble or good end’ also needs to be defined. This definition is linked to the concept of happiness.
Definitions of happiness are the source of much debate among philosophers.88 Analysing the nuances of different ethical-philosophical positions on the meaning of happiness would be a significant undertaking that would not fit within the discussion provided in this chapter, but a few brief points are worth emphasising. First, in a virtue-ethics sense linked to the Greek notion of eudamonia, happiness is not the same as pleasure or fulfilment of desires or needs.89 If that were the case, then this would make for a short-termist and self-centred version of happiness since many of our pleasures and desires can be ultimately detrimental not only to ourselves but also to others around us. Happiness concerns individuals, but it also has important social and collective components.90 Second, happiness is not just about having positive feelings either since this would preclude happiness in the face of adversity. Happiness instead comports the possibility of living well with life’s misfortunes.91 Third, happiness is not a static goal. It is closer to an activity92 – to living well while seeking the best for others and for ourselves.93 This layered understanding of happiness differs from a modern understanding of the word associated with happy states, moments or feelings, so the term ‘flourishing’ is often used as a substitute for happiness.94 Flourishing can be more readily associated with notions of activity and with living well.
In brief, virtue ethics concerns the process of cultivating the virtues necessary for living well in society. It is about developing the capacity to problem-solve rather than the capacity to execute instructions. Compared to consequentialism and deontology, virtue ethics can be criticised for not providing clear guidance on how to act, and for being too centred on individuals and their traits rather than on how they should behave.95 But living well with rapidly evolving technologies, and in a context of multifaceted social and political challenges, requires flexibility.96 For the professionals involved in the machine translation use examples discussed so far, the use of machine translation involves complex factors including the attractiveness of the technology itself, as discussed earlier in this chapter, as well as the organisational factors discussed in Chapter 1. The decision to use machine translation will also have arisen out of circumstances that in many cases will be unique to each human–technology encounter. Individual discernment in the form of moral wisdom and prudential judgement is therefore always required. Individual discernment is also central to accountability. In the same way that interpreters and translators should not follow rules or codes of conduct uncritically,97 it is important for public service professionals too to challenge ineffective rules in the process of advocating improved models of behaviour.
Moreover, recognising the value of a virtue ethics framework does not require rejecting all principles of consequentialism and deontology. As mentioned, all these theory groups have limitations. When pushed to their extremes, they may lead to different conclusions, but when adapted to the realities of each context they may approximate each other’s goals even if they do not share the same ways of framing the problem. The broader challenge, therefore, is that a dogmatic adoption of any given theory is unlikely to equip individuals with the level of critical discernment necessary to navigate machine translation’s risks. Consequence-driven decisions which seek the good of the majority, even if at the expense of a minority, may well be what is called for in some contexts. Such a decision would not be incompatible with a virtue-oriented stance. What would be incompatible would be to disregard all aspects of virtue and assume that the good of the majority should be pursued at all costs. Similarly, it would be the assumption that a prescriptive deontological code should always prevail that could be problematic. Many facets of deontology are in fact desirable, if not necessary, to raise awareness of the risks of machine translation. Although the contents of a policy might not reflect best practices for all scenarios, just the process of describing approaches to the use of machine translation may prompt reflection on its risks. A policy may also be a tool in instigating prudential judgement. The policy may define broad principles and a mechanism for reviewing practices and keeping on top of technological developments. It may specify the need for staff training. Norms do not need to prescribe whether and how machine translation should be used. They may instead outline how best to ensure that individuals and groups can cultivate the moral wisdom to deal with the challenges of each communicative context.
Some of the main competencies to develop in this process are already operationalised in different formulations of AI literacy. For public service professionals, the process of cultivating moral wisdom also involves cultivating sensitivities that are widely discussed in the practitioner training literature, especially in healthcare. Cultural humility is one aspect of this training that is particularly relevant to machine translation use. Cultural humility is defined as ‘a process of openness, self-awareness, being egoless, and incorporating self-reflection and critique after willingly interacting with diverse individuals. The results of achieving cultural humility are mutual empowerment, respect, partnerships, optimal care, and lifelong learning.’98
Diverse cultures and some type of power imbalance between them are some of the antecedents that call for cultural humility in a public service setting.99 While cultural humility is referenced particularly in relation to healthcare and social work, it is relevant to any professional who encounters language barriers in their work, especially those in public-facing roles. Medical education research has emphasised how cultural humility is not a competence which one does or does not have. It is, rather, a commitment to lifelong learning,100 much in the same way that the virtues necessary for living well with technology are never final.
The value of cultural humility is illustrated particularly well by the role of professional interpreters. In interpreter-mediated interactions, interpreters often compensate for cultural blind spots. They are not neutral meaning conveyors.101 The interpreting and communication scholar Claudia Angelelli has provided a series of vivid examples of interpreter intervention in healthcare. Among other metaphors, she refers to interpreters as diamond connoisseurs.102 Some cultures are more prone to indirect or verbose communication, so interpreters may need to identify critical information (the ‘diamonds’) in long stories that may not be relevant to a healthcare provider. She quotes an interaction where, when asked about the severity and duration of a headache, a patient started to describe an accident. The patient believed the accident was associated with his symptoms. The interpreter kept the story to himself while working out whether the unsolicited information was relevant to the question.103 The decision to skip the story carries grave risk for all interactants, including the interpreter. But controlling the flow of information may be necessary in the process of brokering effective understanding between all parties.
This example shows the level of responsibility held by interpreters when they navigate this type of decision-making. If language professionals are not available to bridge cultural or communicative gaps of this nature, this only reinforces the need for moral wisdom among those who call on machine translation to overcome a language barrier. Living well with machine translation is often about living well with others.
Open access