In October 2017, a child referred to as ‘Baby T’ tragically died at a London hospital. Baby T was just under eleven months old. Baby T’s mother, who was from Vietnam, was a victim of religious persecution and was seeking asylum in the UK. Since arriving in the UK in 2015, Baby T’s mother had to change address six times in just over a year, so it was difficult for her to build social connections. It was also difficult for her to access financial support. At one point, an automatic teller machine (ATM) ‘swallowed’ the debit card she used to collect the asylum seeker’s allowance provided by the government. Baby T’s mother did not speak English. She could not understand what was written on the ATM screen and relied on donations to feed herself and Baby T until the card could be replaced. About five months after the ATM incident, Baby T’s mother decided to look for work. She found a job at a nail bar and started leaving Baby T with a babysitter during the day. She had found the babysitter on an online Vietnamese messaging group. On 5 October 2017, an ambulance was called to the babysitter’s address because Baby T was severely unwell. After examining Baby T in hospital, the doctors found an internal bleed between her skull and the surface of her brain. She died later that day. The medical team suspected that Baby T had been physically abused. The babysitter was tried and found guilty of manslaughter. The court concluded that Baby T had been shaken and possibly thrown.
The local authority where Baby T lived with her mother commissioned a review of her death to examine whether similar tragedies could be avoided in the future. Baby T’s story is described in minute detail in the review report.1 Baby T and her mother had been in contact with multiple services including primary healthcare, refugee charities, the immigration department and the children and family social work teams at the different local authorities where they had lived. Baby T’s mother experienced recurrent difficulties in her interactions with government agencies. The case review says that ‘On 25th October 2016 the PIP [parent–infant partnership] key worker began attempting to contact [Baby T’s] mother by text, using Google Translate, which proved unsuccessful.’2 Thanks to a neighbour, the key worker eventually managed to visit Baby T’s mother, but ‘Google Translate again proved unsatisfactory and it was decided that an interpreter would to be used for all future visits.’3
Despite the intention to always use an interpreter, the language barrier continued to be a problem. Baby T’s mother spoke Vietnamese, but on one occasion a Mandarin interpreter was booked by mistake.4 On another, Baby T’s mother did not have an interpreter for an urgent healthcare appointment – the doctor’s practice asked patients to bring a friend to interpret for them.5 Baby T’s mother had suffered from postnatal depression in the past.6 She was offered counselling after giving birth to Baby T and requested a female interpreter to take part in the counselling sessions. Only a male interpreter was available, however, so Baby T’s mother decided to forgo the treatment.7
One of the findings from Baby T’s case review was that ‘barriers to communication with mother had the potential to increase her vulnerability and as she was the mother of Baby T, put the child at risk as well’.8 The report cites several instances where doctors and members of social service teams attempted to use machine translation. These attempts tended to be unsuccessful, so most of them had a negative framing in the review. But the review also paints Google Translate in a positive light. One of the review recommendations was to ensure ‘that advice to parents on caring for crying and sleepless babies is accessible in all community languages’.9 The review leaves open the question of how to make this information accessible but mentions the use of Google Translate as an example:
In [one of the Boroughs involved in Baby T’s care] advice on how to manage crying and sleepless babies is available online and a Google translate facility on the website allows the information to be accessed in Vietnamese. This translate facility does not appear to be available in [two other Boroughs that had also been involved in Baby T’s care].10
Baby T’s story is in many ways emblematic of the discussion I engage with in this book. Baby T and her mother had been in contact with many of the services selected for data collection in the project, including social services, the police, emergency services, healthcare and legal services, which supported Baby T’s mother in her dealings with the immigration department.11 While all these sectors play a critical role in society, the state’s responsibility to protect children usually falls to social workers, who often find themselves at the intersection of all these services. Social workers’ central position within a broader network of state support is particularly illustrative of the challenges examined in this project. Their job requires sensitivity and cultural awareness but also a firm hand, so social work communication can be particularly complex.
This chapter homes in on some of this complexity. The chapter is based on interviews I carried out with eighteen UK social workers. Although the project survey included social workers too, the interviews went a step further. They allowed me to speak directly to public service professionals and hear detailed accounts of their stories. In the sections that follow, I provide some background about the social workers I spoke to. I then look at the many communication challenges they faced in their work. Like in the rest of the book so far, the use of machine translation is my subject of interest, but I often had to consider broader aspects of language access provision to ensure that the circumstances leading to uses of machine translation were fully understood.
Social Workers
Suppose that you have a neighbour who you are concerned about. Perhaps the neighbour is an older individual who lives alone and is struggling to look after themself. Or maybe it is a child who you suspect is being mistreated. In the UK, local government authorities – the local government bodies that employ most state-funded social workers – have a system in place for you to report your concerns and refer your neighbour for an assessment. The relevant social work team at the local authority would then decide what to do next.
Social services tend to be grouped into specialist areas, such as younger adults, older adults, children and families, or individuals experiencing mental health difficulties. The collective experience of the social workers I spoke to covered all these fields. They were a diverse group in terms of their roles and levels of professional seniority. But they also had many things in common. For one thing, they all had strong opinions about the experience of working with service users who had limited English proficiency. Some of them had grown up in multilingual homes and had a lived experience of navigating cultural difference, which often influenced their professional attitudes. Working within tight financial constraints was also an experience they all shared, which often had a negative effect on their caseloads.
They were all extremely busy. One of them spoke to me in the morning straight after a night shift. Another had to delay our conversation because something unexpected had come up in their working day. One other mentioned overseeing 110 child protection cases at the point of joining her team – an amount of work she described as ‘crazy’. I am profoundly grateful to them for taking time away from these hectic schedules to contribute to this project. I introduce them below without using their real names.
Anesu and Deborah specialised in adult social services. Anesu worked with older adults ensuring that their care needs were met. Deborah worked with those between 18 and 25 years of age. The support she provided focused on preparing for adulthood, a type of service that tends to benefit neurodiverse individuals, such as those who live with autism or attention deficit disorder.
Lucy, Salma and Patricia worked in mental health. Salma’s focus was on providing ongoing support to children who were experiencing mental health difficulties, while Lucy and Patricia conducted mental health assessments. Lucy carried out her mental health work in addition to a role in children services.
Farah, Ife and Joanne also worked in mental health, but their experience cut across several areas of specialisation. Farah and Ife had worked with both children and adults at different points in their careers. In Joanne’s case, her role itself involved multiple sectors. She was part of the out-of-hours team, which handles any urgent issues that may come up irrespective of the field. Being a generalist is unusual in social work. Joanne described generalist social workers as a ‘rare breed’.
Maya and Hannah worked in social work education. They trained new social workers, and both had front-line experience in child protection. They also engaged with real cases while supervising and supporting students.
All the other social workers specialised in children services. Ayanda and Rudo managed referral, assessment and intervention teams. These are the teams that investigate referrals to assess the case and decide on any intervention needed to protect or support a child.
Gabrielle managed a child exploitation team that predominantly looked after children above eleven who were being coopted for criminal purposes.
Gisella and Jane were child protection coordinators. Their main role was to chair or manage child protection conferences, a type of meeting attended by professionals, parents/carers and, where relevant, interpreters to discuss the types of support or safeguarding measures that a child requires.
Bethany, Josie and Megan worked in several roles within child and family services including, in Bethany’s case, supporting children who were the subject of ‘child-in-need’ or ‘child protection’ plans. These plans are specific packages of support. In the case of child protection, the plans are put together following child protection conferences. Child-in-need plans are designed following child-in-need planning meetings which, like conferences, are attended by parents/carers and professionals connected to the child.
Although these specialisms may require different working methods, communication is central to all of them. And like in many parts of the world, in the UK this communication is often multilingual and multicultural. Communicating across language barriers was a common experience for all the social workers. The fact that it was common did not, however, make it easy. All of them had experienced difficulties when trying to bridge language gaps. Most had made some use of machine translation. Patricia, Ife and Gabrielle were the exceptions. They had not used machine translation in their work but had important insights to share about what they saw as the potential of AI and why they had not used it. In Gabrielle’s case, machine translation use was common for those she managed even though she had not used it herself.
Social Work Communication
The role of a social worker is to support and protect others. To fulfil this role, social workers often need to intervene in individuals’ private lives and enforce measures that may be unwelcome and sometimes feared. In extreme circumstances, protecting a child may require taking the child away from their parents, which the parents are likely to resist. Offending parents may try to hide important facts or downplay the severity of a problem. Even if the parent has done nothing wrong, they may for different reasons omit relevant details of their story. They may also try to protect others who may put their child at risk. To remain alert to all these possibilities, social workers need to be tactful communicators.
Gisella gave an example of these communication dynamics in the context of domestic abuse. She recalled taking part in a training session where attendees were split into two groups. One group watched a video of a mother and a father having an argument where the father shouted and broke objects in the house. The other group only listened to the audio of the argument without looking at the screen. The argument came across as more violent to those who only listened to it. The lack of any visual reference amplified the intensity of the altercation for the listeners, which was supposed to mimic the experience of a child who can hear their parents arguing in a different room. In Gisella’s experience, parents can underestimate the effects that arguments can have on a child. They may say that their child was sleeping, or that the child was too young to understand what was happening. Ensuring that the parents consider the consequences of their actions is likely to require a stern communicative tone which, in Gisella’s view, is difficult to replicate using tools like Google Translate.
Aspects of prosody – that is, characteristics of spoken language such as rhythm and intonation – can be difficult to reproduce in machine translation-mediated interactions. Some AI tools can be put to this task by allowing the user to adjust how spoken messages are delivered. But it is difficult to judge the efficacy of prosody based on the unfamiliar sounds of a different language. If the message is delivered in written form, then prosody is likely to be lost altogether. In synchronous interactions that take place in a shared physical space, the disruptive presence of a machine translation tool can itself minimise the severity of a message. Having to input the message first (either by typing or speaking it), then obtaining the translation, then showing the translation on a screen or generating a spoken version of it can all take attention away from the message itself – especially, according to Gisella, from the message’s tone, a word she emphasised at different points in our conversation. Ayanda said something similar while flagging the difficulties of providing reassurance to service users. She said: ‘Most families from the global majority have a negative view of social care and they already [have] that barrier or that worry around children being removed. So trying to have that conversation and trying to reassure parents that you’re there to support using Google Translate is very challenging.’
It was not just the process of operating machine translation tools that was described as disruptive. The use of human interpreters too was sometimes problematic. One source of frustration about speaking through interpreters was not knowing the content of additional interactions that took place between the interpreter and the service user. Bethany said the following:
Sometimes I’ve had a situation where I say something and then there seems to be like a whole conversation between the client and the interpreter in their own language and I’m like, ‘What’s going on here? This is so stressful. I have no idea what conversation is being had.’ And sometimes it’s just that there’s clarification happening, which is good but quite often if a family are struggling to understand something, it would be interesting and relevant for me to know that they’re struggling to understand it … As a social worker, it’s important for me to understand sort of how the family are reacting to this information and what seems surprising to them and what is easy for them to grasp and what isn’t.
Assessing individuals’ reactions is important in many professions. For social workers, it is central to providing adequate levels of support. The way in which a parent reacts to a message may hold clues about whether that parent is likely to heed advice or comply with specific requirements. The safety of a child may therefore depend on subtle aspects of communication which language and cultural barriers can easily conceal.
The practical difficulties of using interpreters in social services are well known. The translation and interpreting literature has covered several important aspects of interpreter-mediated interactions in social work, including the value of trust relationships12 and the scarcity of ethical support for language professionals.13 Previous research has also highlighted the potential for interpreters to interfere with social work communication methods,14 and this came up in the project interviews too. Bethany mentioned specific strategies that she had been taught for eliciting information from service users. But those strategies were often introduced with the caveat that they were unlikely to work if interpreters were involved. She found this type of advice unhelpful because it just identified a problem without suggesting how to solve it.
General frustrations of this nature about training or the broader process of dealing with language barriers are an important backdrop to the uses of machine translation that the social workers described. In this chapter, I discuss their encounters with machine translation tools while keeping in sight details of their broader experience, including their interactions with interpreters and their requests for more resources and advice. They faced many challenges that took unique shapes in social work, especially in relation to the multifaceted nature of their role as supporters, protectors but also assessors and enforcers. At the same time, their stories added colour and detail to communicative contexts previously described. Organisational factors (Chapter 1) incentivised some communication methods over others, for instance, while the persuasive convenience of technology (Chapter 2) often weighted on their decision-making. Urgency, unreliable human services and other direct motivators of machine translation use (Chapter 4) permeated their stories. In analysing our conversations, my objective was not to revisit all these points but rather to look at what these different issues meant for them personally – how communication difficulties affected their work and where they saw machine translation in their professional futures.
Plugging Gaps
Rudo described a case where a parent had hit their child and left a significant bruise. When a child is bruised by physical punishment, the police treat it as a case of assault, so an arrest operation was under way. Rudo was summoned to accompany the police to the parent’s address. His role was to ensure that the child had appropriate care arrangements in place while the parent went into custody. The parent did not speak English, and the arrest came up unexpectedly. Machine translation allowed Rudo to provide the parent with basic information while they waited for an interpreter:
We went out with the police because they have to do their bit, but we had to do our bit in terms of, ‘Okay, what’s the plan now? You’re going to be arrested, is there a friend or family member who can come and look after your child, rather than your child being temporarily in foster care?’ … Because it happens so quickly, we had to all go out whilst we’re trying to arrange Language Line [professional phone interpreting], but it took about thirty, forty minutes later for Language Line to find an available interpreter and then we could have those conversations. But in the interim, there was a lot of Google Translate to try and in some ways give the parent some information to let them know what’s going on. But at the same time, it calms people down I suppose when you at least know what’s happening.
Rudo reflected on the situation the parent had found themself in: ‘It’s quite worrying, isn’t it, when you have a police officer and a social worker in your house looking like they are going to do something and then you don’t know what’s going on and it causes panic.’ I asked him how he would have handled this situation if machine translation tools did not exist. He replied that he would have found it very difficult. Asking a family member to interpret occurred to him as a possibility, as did sending out a social worker who spoke the language. But none of these options is particularly convenient or realistic. Family members too might not speak English and, even if they did, sharing details of the case with them could be inappropriate. Sending out a social worker who speaks the language is preferable to using family members, but it would be unlikely that a social worker with the right language skills would always be available.
For interim communication of the kind described by Rudo, machine translation plugs a gap. Relying just on body language, or on a service user’s limited understanding of English, would be likely to heighten the panic Rudo describes, so machine translation here is a de-escalation tool. It is used to explain the basics of what is happening and to let service users know that an interpreter is being contacted. Rudo remembered having to wait two hours with a family in a different case while trying to find a Vietnamese interpreter. Two hours is a long time to fill. He used Google Translate for similar purposes on that occasion. Like in the emergency calls discussed in Chapter 3, machine translation in these cases replaces what would possibly be a silent wait.
The fact that language support may be needed with very little notice is a factor to consider in these interactions, as is the levels of linguistic diversity within a country or specific region. Over 300 different languages are spoken in London according to the Greater London Authority, for instance.15 Covering all the language assistance needs of such a diverse region can be challenging. Additionally, any member of society can make a referral to their local authority. The referral can be incomplete and unreliable. Referrals can also be anonymous, which makes probing for more detail difficult if not impossible. Hannah refers to the arrival of each case as a ‘plate of food’: ‘You’re given this kind of plate of food and you have to like sort it and get into it and find out, you know, what’s true, what’s not true, what’s needed and what happened.’ It is easy for questions of language to go unnoticed in this process. Deborah recalls: ‘I had a case where I didn’t know that the mum couldn’t speak English very well, so I hadn’t put anything in place for additional support. So it was a very uncomfortable visit.’ I asked her why she was not aware of the need for language support. She explained that the child’s father spoke fluent English and it was the father who had been communicating with the local authority. But when Deborah went out on her visit, it was the mother who was looking after the child. Deborah said that whoever submits the referral would ideally answer standard questions about potential language needs, but she is not confident that her authority asks questions of this nature. In any case, the individual who submits the referral is not necessarily aware of all details of a case.
Even when social workers have all the information that they need and arrange language support in advance, they may still have to improvise. Bethany describes:
I can sometimes use the telephone interpreting service as a sort of on-demand alternative if the in-person interpreter doesn’t turn up, although that can be a bit awkward because then you can end up sort of on hold for like twenty minutes sitting in the family’s home, like on hold to the interpreting service … I think what I’ve done – because I’ve just felt like it’s not going to work to try and have a whole conversation using Google Translate – I’ve sometimes, if there’s sort of a key word that is important that I feel that they’re not getting, then I’ll use it [Google Translate] for that.
Here Bethany finds herself in a similar situation to Rudo’s. She had to sit with a family and wait. Bethany did not put this down to urgency but rather to interpreters who are supposed to be there in person and do not turn up. Just like the professionals who completed the project survey, many of the social workers had been in situations where interpreters failed to attend appointments. Here too machine translation addresses a gap, but this time the gap is avoidable. It is created by human language services that are not always reliable.
Some interpreters were unreliable not just for missing appointments but also due to low levels of interpreting ability. Ayanda described using machine translation in a case involving medical terms in English that were difficult to translate. Even though an interpreter mediated the interaction, she felt the need to use Google Translate to find simpler medical terms in Bengali to ensure no important information was missed. Josie sometimes used machine translation to double check that interpreters had understood the seriousness of a case:
When we’re talking about the risk assessment we’re talking about all the restrictions and asking that parent to ensure that, you know, this person is not left alone with the children … and any breach of that would potentially mean that that person could be recalled. And also, at the same time they were liable for a deportation order. So, you know, real-time technology, Google Translate, [is] highly effective in that situation … whilst you might have an interpreter on the phone or in that meeting, they might not fully absorb the seriousness of that.
I was somewhat surprised by Josie’s comment. I wondered what would have led her to question the interpreter’s ability to convey the seriousness of the message. I asked her to elaborate on that, and the explanation served as a reminder of the poor state of ethical support available to interpreters.16 Josie gave me an example. She said that some topics, especially those around sexual abuse, are taboo and risk being sugar-coated or only glossed over by interpreters. She also referred to the fact that some languages may have no standard way of directly expressing sexual acts such as ‘penetration’. Ensuring that important information is delivered in these cases may require phrasing the subject in ways that could be considered vulgar or extreme. Josie expressed a general lack of confidence in interpreters’ ability to handle situations of this nature without softening the message. She therefore took matters into her own hands where she felt the need to alert service users to what was at stake.
Working out the value of using machine translation in the context Josie described is not straightforward. Interpreters are being paid to provide a service, and this service risks being undermined by attempts to correct or complement the message with the use of Google Translate. But Josie’s distrust of interpreters’ ability to deliver the message is not coincidental. Regarding her specific concern about the word ‘penetration’, sexual abuse and rape are different crimes in English law. Penetration is what distinguishes them.17 Cases of rape, where penetration has taken place, carry a longer sentence.18 A 2023 BBC investigation uncovered the case of an Arabic-speaking rape victim in the UK who had her case temporarily dismissed by the police because, among other errors, the interpreter had used the phrase ‘sexual assault’ in English rather than rape.19 Although machine translation is not the most appropriate method of cross-checking an interpreter, Josie is right to have concerns.
Distrust of human linguists was a common theme in the interviews. Most social workers had negative comments to share about the experience of working with interpreters. Jane remembers at least two child protection conferences where minute takers who happened to speak the language broke protocol and interrupted the conference because the interpreter was failing to convey important information. This type of interjection is in principle not allowed. But Jane said that she allowed it as chair because failing to convey the information correctly could have had legal implications. I heard similar stories from the other social workers about interpreter-mediated interactions where service users themselves said they could not understand what the interpreters were saying. Sometimes the answers provided by service users did not correspond to the questions that the social workers had asked.
Machine translation therefore often addressed gaps left by existing communication methods, but these gaps were not all the same. Some of them, like those discussed in Chapter 3, would be unlikely to be addressed by professional linguists. This applies particularly to contexts where service users need to be told that professional human assistance is being arranged. Most social workers who used machine translation for this purpose were thankful for the technology. They also used it to send short written messages and for other types of brief communication which could be impractical to delegate to a professional linguist. Other gaps they described were created by human interpreting services that fell short of expected standards. Here the use of machine translation is more problematic because it is hoped to address deficiencies that should not exist in the first place. Some gaps cannot be addressed by ad hoc machine translation use. They require more systematic and well-funded solutions.
Compounding Barriers
Most of those who need assistance from social workers will be going through some type of difficulty. They are likely to be vulnerable in some way whether because of their age or because of health conditions. They may also experience financial problems that make it harder for them to look after themselves or others under their care. Individuals who are new to the country may experience language barriers on top of all these difficulties. The social workers I spoke to gave vivid accounts of what these overlapping difficulties meant in practice. Some of them helped individuals who had ‘no recourse to public funds’, an immigration condition that restricts access to certain types of government support, such as financial benefits or housing assistance.20 Service users’ needs may also relate to religious or cultural factors. They may prefer to deal with professionals of the same gender, for instance. Like in the case of Baby T’s mother, it might be impossible to meet their gender preferences, which can effectively block access to appropriate assistance – an issue I heard in this project too.
These different barriers often compounded each other. The reliance on machine translation was a feature of this broader context, and two compounding difficulties stood out as particularly problematic: the question of languages or dialects for which language services are not widely available and the question of service users who had other accessibility needs in addition to having low proficiency in the local language. I look at these two questions separately below.
Low-Resource Languages
I have referred to ‘low-resource languages’ at different points in this book. Although there is no standard definition of a low-resource language,21 in the context of machine learning it is generally agreed that these are languages of low digital representation – that is, languages for which textual data and automatic text processing tools are less likely to be available. This is not a new problem.22 Researchers have been trying to address linguistic disparities in language technology performance for some time. But the lower quality of machine translations for low-resource languages persists.23 Without significant amounts of texts or speech records in a language, it is difficult to get AI models to learn the language’s vocabulary or reproduce its common patterns. So the scarcity of linguistic resources is what operationally causes the problem. But why are linguistic resources scarce for some languages and not others?
The work of the French sociologist Pierre Bourdieu is relevant to answering this question. As famously stated in one of his articles, ‘a language is worth what those who speak it are worth’.24 Bourdieu here is alluding to the notion of linguistic capital, or in other words the social value attributed to a language or to specific ways of using it.25 The social workers I spoke to often interacted with service users from the Global South. Many of these service users spoke languages that historically have not enjoyed the same level of visibility and prestige as the languages of countries that are perceived to be politically and economically more influential. Take the example of Bengali. With over 280 million users, Bengali is estimated to be the world’s seventh most spoken language.26 It has a larger user population than languages such as Japanese, Italian or standard German.27 Bengali is often described as a low-resource language.28 Japanese, German and Italian are not.29 Japanese, German and Italian are predominantly spoken in countries with large economies whose populations have high levels of purchasing power. There is a strong financial incentive to ensure that products and services are available in these languages, which are also spoken in countries of significant political influence with direct representation in international blocs such as the G7 or,30 in the case of German and Italian, the European Union.31 The number of users of a language is therefore not enough to explain linguistic disparities in technology performance. Asymmetries of both financial and social capital, as well as the political status of the language regionally and internationally, can influence the value attributed to the language on the world stage as well as its levels of digital support and internet presence.
Through no fault of their own, speakers of low-resource languages are more likely to be ill-served by AI models. Speakers of regional accents of a language may have a similar experience. Asymmetries of social and financial capital are built into machine learning by reason of how, for the most part, this technology just mimics representations of the world with all their biases. Limited digital resources are therefore a symptom of a deeper problem. If a language has a low level of linguistic resources, this language may be disadvantaged in other ways, including in terms of the size of its pool of qualified linguists.
In Chapter 1, I mentioned the fact that there was a single Igbo interpreter on the UK’s National Register of Public Service Interpreters. The difficulty of finding professional linguists for low-resource languages is something I have experienced myself. Take again the example of Swahili. While Swahili is estimated to have over 200 million users across several African countries including the Democratic Republic of Congo, Kenya and Tanzania,32 it is often described as a low-resource language.33 It took a whole month for a language company to find a Swahili translator who could assist with the passages about the Mental Health Act I discussed in Chapter 4. My timeframe for this book was flexible, so I could wait until a translator became available. Social workers, on the other hand, rarely have the luxury of time.
Low-resource languages are thus languages of low human resources too. While the difficulties reported by the social workers were not restricted to low-resource languages, unequal technological performance was sometimes a personal experience for them. Lucy said: ‘Sometimes, Google thinks I’m saying something very different because Google itself cannot understand me so … It should be for Google to be able to pick up different pronunciations … because we are different.’ More often, it was the low availability of professional services for certain languages that posed a problem. Interpreter impartiality, for instance, can be compromised if a language has a small local group of users. Maya described a case where the presence of an interpreter from the service users’ community made a family feel uncomfortable:
Once, I had an interpreter with a … family and then I wondered why is mother suddenly clammed up and then upon talking further I came to know that actually the interpreter was from the same community and they saw each other in the temple … She [the mother] felt very, very, very embarrassed and it was about her child [and] sexual abuse and of course this brought a lot of shame.
In public service settings, it would arguably be incumbent on the interpreter to declare any factors that may stop them from being perceived as impartial. Megan mentioned always asking interpreters and service users if they knew each other and ‘if at any time they then realise, for example, [that] they are from the same village et cetera, can they let us know so that we then know that there is some connection there’. But what happens if interpreters or service users declare an existing connection? Booking a different in-person interpreter at that point would at best involve a significant delay, assuming that it is at all possible given the small size of the community.
Service users who belong to small linguistic communities may be reluctant to speak through interpreters in the first place due to concerns about privacy and confidentiality. Bethany says: ‘I’ve also had experiences where … people come from very small countries or like countries where there’s quite a limited number of people … they would worry about confidentiality if working with an interpreter from that particular culture.’ The use of machine translation here may on first impression seem beneficial, but it will be precisely speakers of minority or socially disadvantaged languages that will experience the greatest risks of being mistranslated by AI.
Megan had experienced problems with Sudanese Arabic, a version of Arabic spoken in Sudan and nearby regions.34 She said: ‘Sudanese Arabic is harder to find so you end up just using the Arabic version.’ She clarified that by ‘the Arabic version’ she meant the standard Arabic option that is usually available in machine translation tools. She added:
Those languages that are less common, that’s where you then find the issue … You might ask a simple question and then from their response you then realise that they haven’t fully understood. Like you can just ask a yes or no question. ‘Is it okay for me to see you tomorrow instead?’ And then they cannot respond, so you’re not sure as to what is happening.
Confidentiality, machine translation accuracy and access to qualified linguists are all likely to be issues of greater concern for those with low level of linguistic capital. When Google Translate fails to support specific dialects and when interpreters are also unavailable, the result will be a lower standard of care provided to the service user. These linguistic asymmetries cut through every aspect of multicultural societies, from the limited language options available on guided tours at a museum to the small group of additional languages that are usually taught in schools and universities. Languages are socially unequal. This is not an easy problem to solve, but unless low-resource languages get the attention they deserve, inequalities will continue to feed warped perceptions of AI based on its superior performance for privileged languages.
Additional Needs
Although some of the points raised by the social workers had been previously mentioned in the project’s survey, the interviews also helped to bring several less prominent topics into focus. One of these topics concerned accessibility needs that involved more than just a barrier between verbal languages.
When speaking to Jane, she mentioned difficulties with getting written reports translated for families. These were reports provided to families ahead of child protection conferences. Timescales are tight and sometimes families arrive for these conferences without knowing the content of reports that have been written about their child. When translated reports were unavailable, Jane normally interrupted the conference and asked the interpreter on duty to sit with the family and tell them what the report said. This practice is akin to what translation and interpreting scholars would call ‘sight translation’ – reading a text and translating it orally at the same time. I asked Jane whether she knew if families themselves used machine translation tools to read reports provided to them in English. She did not know if they did but hastened to add: ‘We do also have parents who can’t even read as well … That is another layer to it.’
Low literacy, in the standard sense of being able to read and write, was a significant additional barrier reported by some of the social workers. In Gisella’s view, calls for greater availability of written translations sometimes overlook the fact that some service users cannot read or write in their native language. Using machine translation in these cases can be particularly risky. Non-readers and non-writers can use machine translation so long as the tool has a speech function and can recognise what they say and generate spoken versions of any translations. But these users will be subject to speech recognition itself as an additional source of errors. In other words, further to the possibility that a tool might mistranslate logical and complete inputs, for non-writers the input itself can get corrupted – for example, because the tool was not trained to understand their accent or because they are speaking in a noisy environment and their speech is misrecognised. The risk of speech recognition errors is the reason why travellers into the US are allowed to correct transcriptions of what they say to officers at the border, as discussed in Chapter 1. But individuals who cannot read are unable to carry out this type of check. For non-readers and non-writers who do not speak the local language, the interaction with social services is particularly reliant on third-party human mediators.
Gisella said that she often asked service users whether they wanted written translations. If she just assumed that translations were required, waiting for the translation would risk causing unnecessary delays for individuals who could not read the document independently anyway. She remembered a specific case of this nature: ‘There was this one Romanian mum and she was just like, “Don’t bother like – even though I can speak Romanian, don’t bother [to] get it translated. Just give it to me in English and I’ll get my daughter-in-law to read it to me.” Because she couldn’t read.’ Leaving the document untranslated may save time but it has its own risks. Information access in this case depends entirely on the daughter-in-law and her level of bilingual proficiency. A professional Romanian translation of the document would probably make it easier for the mother to seek help from others. In any case, there is no getting away from the fact that this mother was unable to engage directly with a system that is heavily reliant on written records of important assessments and decisions.
Additional needs often act as additional obstacles, but it is also true that not all accessibility requirements are the same in terms of institutional perceptions and resource availability. At the end of the interviews, I asked the social workers if there was anything else that they would like to share that had not been mentioned. Anesu at this point remarked that her local authority had different attitudes to the use of technology depending on the nature of service users’ needs:
I’ve worked with citizens who have learning disabilities and their language would be technology because they’re using, you know, the wheelchairs where it’s got a voice … As a council we’re more prone to, ‘Oh wait they’ve got learning disabilities, let’s give them as much technology as we can to try and see if we can get them to communicate with us.’ But then when it’s someone who’s not been through that or doesn’t have a learning disability, I’ve noticed it’s just been like, ‘Yeah, let’s shy away from technology a little bit.’
Anesu here is referring to a tendency for uses of machine translation to be officially discouraged, while uses of assistive speech technologies may be encouraged and widely accepted. This is not a straightforward comparison. Speech synthesisers – like those famously used by the British physicist Stephen Hawking – inevitably involve different types and levels of risk compared to uses of spoken machine translation. But it is true that these technologies are not dissimilar. Speech synthesisers have limitations too. The voice Stephen Hawking made famous was initially described by the New York Times as ‘usually understandable’ in a nod to how some aspects of communication were lost by using the device.35 Experimental systems that convert brain activity into speech are also unsurprisingly imperfect. Listeners have judged one recent such system to be between seventy-five and eighty-four per cent accurate, meaning that at least sixteen per cent of the synthesised output was liable to be misunderstood – for example, because of confusingly similar pronunciations of the words ‘back’ and ‘left’.36 Widely available commercial tools are also limited. In her cultural history of voice technologies, Sarah A Bell reminds us of how seemingly simple human sounds such as laughter can be extremely difficult to synthesise and may thus be omitted by voice assistants such as Apple’s Siri.37 Laughter is important for communication and relationship-building. If synthesisers fail to reproduce it effectively, this is bound to be felt by those who rely on this type of technology.
Lucy raised a similar point to Anesu’s. As discussed elsewhere in the book, most types of communication will be subject to some level of risk. Machine translation can be seen as just one additional communication method that is not fundamentally different from other methods which individuals may favour or disfavour for different reasons. Lucy says the following in this respect:
You’re looking at alternative ways of communicating with your client. It’s like some patients don’t want to talk; they just want to write down or they just want to draw. So it does have different forms of communication. Google translation is not very much different from that. If that is the preferred way of communicating, then why not?
Spontaneous pictorial messages are likely to be inherently subjective and potentially ambiguous, so professionals may instinctively question whether their interpretation of a drawing is correct. This means that drawings , despite their subjectivity, may in fact be less consequential overall than potentially misleading machine translations. So Google Translate is not quite just another communication method like many others that are already in use. But there is a point in saying that some assistive technologies are more established than machine translation and may be less controversial and more readily accepted by public institutions. The risks posed by these different technologies will depend on factors including community integration and continuity of care.
Users of formally endorsed assistive tools will in most cases have accessed other services by the point they start using an assistive technology. They will have needed to undergo assessments, attend appointments and interact in some way with professionals who have overseen the process of requesting and deploying the tool. If the assistive tool is funded by the government, the service user may also need to be allowed recourse to public funds which, as mentioned, newcomers to the country may not have. An individual’s experience of any type of social barrier is likely to be unique, so there is little point in comparing the severity of different barriers. But when significant barriers overlap, the experience of those affected is unlikely to be positive. Like Baby T’s mother, those who are new to a community and do not speak the local language may have limited support networks. They may be unable to count on informal advocates and thereby struggle to go through preliminary steps that would grant further access to other services. If on top of being a newcomer they have other specific needs because of mental health difficulties or because they might not be able to read or write or speak, they are likely to be further ostracised. Their additional needs may go unmet or undetected, which can in turn exacerbate language barriers they may already be experiencing.
The compounding nature of these needs were particularly clear in the accounts of mental health social workers. Lucy said the following: ‘In some instances … they can’t speak English, they are unwell, and we get the interpreter and the interpreter will say, “I can’t understand what they are saying.”’ I asked Lucy why the interpreter would not understand the service user. She replied:
Because of their [the service user’s] disorder, they are [saying] something that is totally not making sense and the interpreter cannot add up. But if it’s somebody that speaks English … or understands me very well, I’m able to make sense of it myself, probably from what they are speaking and picking that, ‘Oh they’re saying this because they are unwell.’
Lucy was referring to cases where service users struggled to communicate because of their mental health. Not even interpreters are able to help in these cases. Mental health professionals may be able to use their experience to figure out what the problem is, but this is difficult to do if the individual is new to the service or if they have low proficiency in the local language. Moreover, if a professional interpreter could not understand incoherent speech, it would be unlikely for AI to offer any help. Human professionals are expected to be transparent about their limitations. If they do not understand something, professional guidance requires them to say so.38 The same cannot be expected of AI. For the time being at least, AI models are not good at admitting what they do not know.39 When their use overlaps with circumstances that make communication harder, whether it be learning disabilities, mental health conditions or low levels of literacy, these models are likely to be particularly unreliable.
On the Strengths and Weaknesses of Humans and Machines
Almost all the social workers emphasised that their use of machine translation was limited to ad hoc contexts like those described above, such as while waiting for an interpreter, when arranging visits or when interpreters did not turn up for an assignment. Something I heard often was that machine translation could compromise the admissibility of information in court, so it could not be used to carry out formal assessments or investigations. For Patricia, who had not used machine translation at work, legal admissibility was one of the main reasons for avoiding it. She explained: ‘In terms of doing a Mental Health Act assessment, I and my colleagues would have a lot of queries about using something like that … Say we see an incident through the courts? We’re liable, so …’
It is expected that the social workers would have refrained from using machine translation for interactions that would be scrutinised in court. But non-use of machine translation was due to more than just fear of inaccuracies. First, not all social workers were familiar with the technology. Ife, another non-user alongside Patricia and Gabrielle, had not used machine translation because she had only just heard about it: ‘No, I’ve never heard of it to be honest. I don’t know this Google translator.’ She then laughed and suggested technology was not her forte. But later in our conversation, she said that she had tried a speech recognition tool together with a colleague two weeks before. The tool could not pick up what she had said. If she depended on this type of tool during an assessment, she explained that she would probably insist on using a human interpreter or interrupt the assessment.
Perceptions of the social norm (see Chapter 5) were another reason why the social workers avoided machine translation in formal contexts. For some of them, the direct concern was not that the technology would make significant mistakes but rather that judges could take issue with the fact that it had been used at all. Gabrielle said the following:
I can see good aspects for it [machine translation] as well because we are working in a bit of a technology age and I think even for parents it’s instant for them if I say something, they can just give me the phone and I’ll say it and they understand it … but I feel like say, for example, we were taking a case to court. I’m not too sure how well Google Translate would be for the judge if we said that’s how we shared really serious information.
Gabrielle’s concern was mainly what the judge would think about it – how official or professional the use of machine translation would be perceived to be. Similarly, Megan seemed concerned about the possibility that families could allege having misunderstood critical information: ‘I have been in court a number of times. The judge will always ask you, “How did you explain to the family that this is what was going to happen?” And if you are to say, “I used Google Translate”, it’s very easy for them [the family] to say they did not understand.’ Here again what is called into question is not the accuracy of machine translation but rather that its use can be leveraged to serve specific interests.
It was common for the social workers to anticipate service users’ attempts to exploit different aspects of the system to achieve their own objectives. Patricia recalled a case where she needed to assess an individual under the UK Mental Health Act. The individual’s family had told her that an interpreter was not required for the meeting. During the appointment, the family started interpreting for the individual and Patricia realised that an interpreter was, in fact, required. I asked her why the family would try to hide or downplay the need for language assistance. ‘They felt the person needed admission; that was their agenda’, she said. Patricia explained that the family probably wanted to elicit from the individual the types of answers that they presumed would prompt social services to admit the individual to hospital. The assessment was rescheduled in this case so that the individual could speak through a professional interpreter. Insisting on the use of interpreters is important to ensure that individuals are given a ‘fair chance’ to tell their story, Patricia added.
Gisella also described contexts in which friends and family members acted as informal interpreters. I asked her what she preferred: Google Translate or this type of informal assistance. She was emphatic about preferring Google Translate. She thought that family members were unreliable because they could intentionally misrepresent other interlocutors: ‘The other person could say, “Tell her to leave my house right now”. And they might say, “Oh yeah, my mum said that’s fine.”’ Compared to some of the communication methods the social workers had to employ, machine translation could therefore be preferred because of its inability to be deliberately misleading.
But overall the social workers were quite nuanced about whether they preferred humans or machines. Human and machine mediators were not that different for them because all communication methods required some level of trust. They thought machine translation was clearly more efficient than humans and could sometimes lead to a better standard of care by avoiding delays, particularly if they already knew the service user and had an ongoing relationship with them. Humans were preferred for their emotional intelligence and ability to empathise. But too much empathy was also a problem. Many of them referred to experiences where the conduct of human interpreters was considered too personal and thereby unprofessional. Farah said:
The client might be explaining their personal issue to them [interpreters] and they [interpreters] tend to get carried away and are not professional in that aspect and, although they are maybe trying to show empathy, they get carried away and [will] then be discussing things that are different. Whereas I, the one that initiated [the interaction], will be left out.
Megan said something similar. Between facing a long wait for a human interpreter and promptly using a machine translation tool, she preferred the tool:
Waiting for a face-to-face interpreter does take time … So if there was an app that I could use, I would much prefer that than having to wait … and at least you would know that they [apps] are quite objective, whereas sometimes when you are working with an interpreter, you can tell that they’re either being so empathetic with the family, that they feel sorry for them and they are trying to help them, but that’s not their role. Because sometimes in them trying to help them, they’re making things worse for the family.
As discussed in Chapter 2, technologies are not neutral. The unequal performance of machine translation across languages, for one example, is far from impartial, so the assumption that machine translation is objective is at best precarious. But it is true that machine translation tools do not have consciously motivated intentions. They do not feel or think in the way that humans do, and for that they can give the impression of being more trustworthy.
Machine translation can be called upon even in formal contexts where its use is unexpected. Megan had been involved in carrying out age assessments. These assessments are necessary when an individual’s age is in doubt and needs to be determined for some legal reason. Often this is the case when individuals arrive in the country and claim to be minors without being able to show proof of their age.40 Determining the individual’s age is important because, if they are under eighteen, they will be supported by social services. Adults, on the other hand, are placed in general accommodation and may in some circumstances be removed from the country.41 Megan was the first among the social workers I interviewed to mention age assessments. Given how consequential these assessments are for the individual, she described age assessments as a typical context where machine translation was not used. But when speaking to Joanne, it became clear that not even age assessments were entirely off limits.
Joanne was sometimes called to conduct age assessments at airports. She described challenging scenarios where individuals would go to a toilet right after landing and wait there to allow other flights to arrive and make it harder for their departing country to be identified. She recalled a specific airport-based age assessment: ‘At one time, we didn’t have any type of [human language assistance] so we had to use the phone, you know, the Google Translate.’ I asked her whether she thought Google Translate had worked on that occasion. She replied: ‘I think it worked pretty okay … I mean for the purposes of getting basic demographics and whatever the person was saying. Obviously we did need to do further assessment to determine the authenticity of their claim, but in terms of us understanding what they were claiming, it did work.’
When describing this case, Joanne mentioned that unless there was evidence that someone was over-age, that they had to be treated as a minor. So just getting a sense of what the individual is trying to claim about their own age can be helpful. But what if translations of initial accounts provided by the individual are misleading? The topic of age assessments can be highly charged. A finding that an individual is over-age can be seen as favourable to governments that are under pressure to control immigration.42 Based on my conversation with Joanne, there was no evidence that machine translation was being used to stop individuals from accessing critical services or to confirm findings that they were above eighteen. That was reassuring, since it would arguably be preferrable for machine translation to spuriously benefit someone than to spuriously deny them support. But what Joanne did make clear was that machine translation was present even in contexts where its use may be presumed unacceptable.
In my previous work on how language professionals use technology, I have often come across strong views on machine translation and on its financial consequences for translators.43 The use of machine translation can be highly controversial in the language industry. It is clearly controversial in social work too, but not for the same reasons. Like most professionals who took part in the project survey,44 the social workers tended to be highly satisfied with machine translation, but they were not dogmatically pro or against it. What they were invariably concerned about was the standard of care received by their service users. Some of them, like Salma, framed almost our entire conversation around the idea that different tools can be suitable for different circumstances, and that having the option to draw on a range of tools was a good thing. Especially when some of these tools are new, and when their use may have a bearing on perceptions of professional conduct, confidently deciding which one to use is likely to require guidance. Most social workers had strong opinions to share and requests to make in this respect.
Requests for Support
Towards the end of our conversations, I asked the social workers what topics they would like to see covered in multilingual communication training and what resources they would ask from the government if given the chance. Unlike some of those who participated in the project survey (see Chapter 5), all the social workers interviewed thought that more training would be helpful. Just the opportunity of reflecting on the subject during our conversations may have alerted them to the complexity of the problem: ‘It’s been good actually reflecting on it as well from my end’, Jane said at the end of our interview.
Much of what they asked for could be addressed by more funding. For example, being able to employ interpreters in-house would allow the social workers to build trust with the same group of interpreters. It would provide social service providers and users with more consistent and continuous language support. They also asked for more training for interpreters, and for interpreting providers to offer a more tailored service to social work. Although social work, healthcare, prisons or the courts all involve different vocabularies and communication objectives, often the same interpreter can be assigned to any of these sectors depending on the job and on availability.
Solving problems of this nature would have significant cost implications, but the social workers also made suggestions that were more financially modest. Echoing some of the survey participants, Anesu referred to a culture of ‘whispers’ about machine translation in her local authority:
I think it’s just like a company whisper going around … I even used it [Google Translate] like two weeks ago and then I messaged my colleague, ‘I just sent this text message, I hope it’s okay?’ But yeah, so they [colleagues, managers] just whisper it around. They don’t really advise – they don’t give any further information.
Giving further information requires acknowledging that machine translation tools exist and that they might be used. Managers may find it risky to acknowledge this openly. If they issue an outright machine translation ban, they might end up harming individuals who cannot wait for human interpreters. If they say that it is okay to use machine translation in certain circumstances, they might fear that this could open the floodgates to widespread use of it and that they would be held accountable for data protection breaches and mistranslations. So often the social workers ended up in a liminal space where some uses of machine translation were probably fine, even though they were not certain of that. Many of their requests therefore boiled down to wishing for a franker discussion about the challenging circumstances in which they often found themselves – what the privacy implications of using AI translation were, but also what to do in emergencies or when interpreters did not turn up.
They also called for service users to be involved in discussions about language support.45 Service user involvement can be seen as part of the humanitarian principle of ‘accountability to affected people’.46 Their involvement is particularly important in the process of drafting specific guidelines or bringing new policies into effect. It is a way of checking that any new procedures to be implemented within an organisation are indeed beneficial to the individuals who stand to be affected by the procedures, including service users and individual service providers.
Some of their requests were broader. Josie called for stronger emphasis on language learning at a national level. She had relatives in Germany who spoke multiple languages, but she was disappointed by the few weekly hours of French her daughter had at school in the UK. In a nod to the concept of linguistic capital, she admitted that even if her daughter learnt more languages at school, these would probably not be the languages that social workers need assistance for in their work. The languages she often needed help with were not usually taught in formal education. But the broader point that Josie was making, which all the social workers mentioned in some form, was that more attention needed to be paid to diversity – to the importance of recognising how multicultural and multilingual the country is.
This type of recognition can be difficult to achieve. Opinions on the type and extent of support that should be available to non-speakers of the local language can be polarised in some countries. These are, in many ways, political questions that can be easily dismissed by governments that do not see language access as a priority for their electorates. In the UK, it would indeed be unlikely for questions of multilingual communication to be high on the electorate’s agenda because many of those who most require language assistance are not part of the electorate in the first place. They are often visitors and newcomers to the country, and many of them do not have the right to vote. But attending to their needs, I would argue, is a moral question.
The fact that the social workers emphasised some of these broader questions is not a coincidence. Many of them had direct experience of supporting individuals who had been denied access to important services and information. Like many of the project’s other participants, they often found themselves between conflicted managers, on the one hand, and service users with urgent communication needs, on the other. They called for more discussion of these circumstances. Although this book cannot address the problems that they brought to the fore, I hope that by sharing their stories it will help to promote the type of open dialogue they wished to see.
Open access