As shown by examples discussed earlier in the book, asylum seekers may, and sometimes do, have their cases rejected because of machine translation errors.1 Immigration officers themselves use machine translation, and the needs of the civilians they interact with, including unaccompanied children, can be misunderstood.2 Crime suspects may not understand what information police officers require from them,3 while language needs within an organisation can go undetected and lead to life-threatening miscommunication.4 As AI translation continues to advance, its benefits will grow. But even as the technology advances, some uses of it will continue to pose significant risks. This chapter is about these uses and the circumstances that motivate them.
I note that the project data does not include examples where machine translation directly caused someone’s death or something similarly catastrophic. The project participants were in fact highly satisfied with their uses of machine translation.5 As expressed in other words elsewhere in the book, the challenge with machine translation is not that it is always dangerous, but rather that its dangers can be difficult to predict with the level of confidence that many public services require. Moreover, given the technology’s increasingly idiomatic and plausible translations, the real nature and extent of misunderstandings may not be apparent to non-bilingual users. By the same token, even when there are no misunderstandings, it is difficult for users to be certain of the technology’s efficacy. The lack of catastrophic machine translation uses in the data should not, therefore, be deemed to imply that machine translation is safe. Many of the use contexts discussed in this chapter raise significant concerns about service users’ poor access to information and service providers’ poor access to resources.
I start the chapter by discussing different examples that illustrate the high-risk circumstances in which machine translation tools are currently used in UK public service contexts. I then home in on the issue of informed consent, an area of concern in service encounters which many project participants highlighted. I discuss two types of consent. The first relates to the need for service users to consent to the service they are receiving or to any procedures they may be subjected to – for example, consenting to a medical examination or to a police search. The second type of consent relates to machine translation itself – whether it is used with the consent of the individuals involved. Later in the chapter, I examine circumstances where uses of machine translation may seem unavoidable. These circumstances involve service users’ habits and preferences – which can influence service providers’ decisions – as well as time pressure and conceptions of person-centred care. I end the chapter with a discussion of how human language services can also be unreliable. In my research, deficient provision of these services was particularly noteworthy among the reasons why machine translation was used and why sometimes service users did not receive the standard of care that they deserved.
* * *
While some uses of machine translation discussed in Chapter 3 hovered between ancillary communication and core care, the uses I discuss here are unequivocally central to the service being provided. In medical contexts, machine translation was often used for history taking and for deciding the type of care that the patient required: ‘In A&E [the emergency department] when taking history from a patient or when explaining plan of care to a patient’ (Paediatric Nurse); ‘To understand a patient’s care needs when she was distressed’ (Inpatient Healthcare Assistant). In social work, machine translation was used in safeguarding – that is, when trying to protect individuals from harm, health threats or abuse:6 ‘To communicate with foreign nationals over safeguarding issues’ (Social Worker); ‘Planning care. Dealing with safeguarding. Dealing with complaints’ (Registered Manager, Healthcare and Social Assistance). In law enforcement contexts, it was used in front-line tasks to speak to suspects, victims and witnesses: ‘Speaking to victims of crime to establish what has taken place, who what where etc. Speaking to foreign drivers who do not have a UK licence and are committing offences’ (Police Officer). In legal services, professionals used machine translation to understand documents, to translate statements and to deliver critical information: ‘I would have to translate legal letters to the client’s own language using Google Translate’ (Legal Secretary); ‘Mostly translating vouching docs to understand the case better and build up evidence’ (Solicitor).
When I started collecting data for this project, I expected to come across machine translation uses of this nature that have a direct bearing on individuals’ livelihoods and well-being. These uses of the technology have made it into the public domain in some cases as discussed earlier in the book. They can be alarming but are not surprising. What I do find surprising is the stark disconnect between institutions’ current high-level public messaging and what is happening in practice when the institutions’ staff meet, interact and serve members of the public.
As mentioned in Chapter 2, guidelines published by the National Health Service in the UK say that machine translation should be avoided.7 In legal settings, too, the Standards for Language Access in Courts, first published by the American Bar Association in 2012, cautioned United States courts against machine translation.8 Yet a previous review I worked on uncovered case law passages like the following issued by a court in New York: ‘Because Plaintiffs provided no translation of any Polish documents submitted in support of their motion, the Court used the free “Google Translate” service, available at translate.google.com, in order to confirm certain statements contained in those documents.’9 Although this case is from 2017, five years after the publication of the American Bar Association Standards, machine translation was not and still is not reliable enough – whether in 2017 or at the time of writing in 2025 – to confirm statements in a court of law.10
There are contexts where machine translation is just too risky. One such context involves medication dosing and prescription. Responses submitted to this project included the following machine translation use descriptions: ‘A patient who speaks very limited English requiring medical help, post op [i.e., after an operation], with medication counselling, prescribing and dosing information’ (Pharmacy Technician); ‘Taking a drug history, counselling patients on medication, understanding the presenting complaint’ (Pharmacist); ‘Writing medication instructions’ (Doctor). Even the most up-to-date large language models available at the time of writing make mistakes with numbers and maths.11 Dedicated translation models also make mistakes of this nature involving ‘deviation in numbers/time/units/date’.12 Using machine translation for medication advice can pose a direct threat to an individual’s health if the model mistranslates dosage and frequency instructions.
Communication between pharmacists and members of the public poses particularly complex challenges. In the case of pharmacists working in a community setting (i.e., outside of hospitals and medical centres), communication is likely to be ad hoc in cases where individuals walk into a pharmacy seeking over-the-counter advice. The impromptu nature of these encounters is likely to make it harder to obtain professional language assistance in a timely manner even though language services are available to pharmacies in many health systems.13 Moreover, while in some countries, including the UK, pharmacists can request to view an individual’s medical record, tourists and newcomers to the country may not have such a record, at least not one that the pharmacy can access. It is up to the individual in these cases to provide information about allergies, chronic conditions and any medications they may be currently taking. The use of machine translation in encounters of this nature can be highly consequential even if the interactions are brief.
Another use context described in the project involved ‘translating medical reports in order to put the relevant information into someone’s NHS record’ (Patient Experience Coordinator, Healthcare and Social Assistance). An individual’s medical record can influence several aspects of the care the individual receives in the future, so mistranslations here could have a long-lasting effect. Errors could also go unnoticed unless the healthcare professional adds a note to the record declaring that machine translation has been used and that the content could be inaccurate, a level of transparency that is not guaranteed.
If the material being translated includes medical jargon and specialist terminology, the risks are even higher. A nurse explains, ‘We would use Google Translate for more complex medical jargon when screening patient fitness.’ The nurse adds, ‘if someone has a grasp of English, but rarer complex medical jargon is beyond them, it [machine translation] can help’. This use of machine translation resonates with the mixed uses of written and spoken language discussed in Chapter 3. The expectation here is that machine translation may help to clarify key parts of the message by delivering the information in the patient’s native language, thereby complementing a potentially unreliable monolingual exchange.
This strategy may indeed have value, but it poses significant risk when machine translation is reserved for the parts of the message that are more difficult to understand. Previous research has shown that some machine translation inaccuracies are associated with text passages that humans too find difficult to translate.14 Research from the early 2000s identified textual features that tend to be associated with machine translation inaccuracies, such as ambiguous words or long sentences.15 Although a lot of this work was done before neural machine translation and large language models, the general principle that simpler language leads to more accurate translations still stands.
In a 2024 Australian study, machine translations of texts that had been simplified were preferred by native readers over machine translations of non-simplified texts.16 This simplification involves a process known as pre-editing – for example, changing ‘make a U-turn’ to ‘go back’ to make the phrase more culturally agnostic and translation-friendly.17 The process is analogous to so-called plain language rules.18 Texts written in plain language are easier for any user to understand, including native and non-native speakers who, in an emergency, may all have their reading ability impaired due to high stress or anxiety.19
The use of plain language is likely to be particularly beneficial in communicative contexts marked by specialist terminology. A study from 2023 showed that Google Translate left French and Spanish medical abbreviations untranslated, especially if the abbreviations were not accompanied by the full term.20 An English-to-Lithuanian study from 2024 reported similar findings: eight per cent of the medical terms included in the study’s sample were mistranslated by the tools used.21 For everyday low-stakes purposes, eight per cent may seem like a low error rate. But as discussed in Chapter 2, this is not a straightforward assumption when the content has the potential to affect someone’s health. In the professional sectors discussed in this book, machine translating jargon will rarely be a good idea. Yet many of the project participants needed to explain jargon or clarify technical terms to service users or civilians.
Like police officers quoted earlier in the book, some of the officers who took part in the present project used machine translation during arrests: ‘[To] explain to people what is happening, being arrested or searched’ (Police Officer); ‘To tell someone they are under arrest or to gain information from them about why they have called the police’ (Police Officer). Similarly, data from my previous work included an example where Google Translate had been used to read a patient their rights under the UK Mental Health Act.22 The patient was being sectioned – that is, being legally forced to stay in hospital. One of the aims of the Mental Health Act is to protect an individual, and those around them, from harm even if this means forcing the individual to be hospitalised against their will.23 Because of the very nature of the act’s enforcing powers, it can be difficult to challenge the section.24 But sectioned individuals have a series of rights, including the right to information, to receive messages from their solicitor and to appeal the section through the appropriate legal channels.25 Mistranslations can stop many of these rights from being fulfilled. The mental health advocate who took part in that previous study said the following: ‘It is important they [patients] have accurate information and an interpreter told me once that the written information I had been given by Google Translate was almost unintelligible.’26
English-to-Portuguese tests demonstrated that, in late 2024, Google Translate still failed to capture the meaning of ‘sectioned’ in a medico-legal context. When given a text passage about the Mental Health Act, it translated ‘to section you’ as ‘to dry you’ (secá-lo).27 Both DeepL and Microsoft Bing Translator, two well-known systems currently available, translated ‘sectioned’ as to be ‘cut’ or ‘divided into sections’ (ser seccionado).28 When provided with the same passage, GPT-4, the large language model that powered ChatGPT during the test, translated ‘sectioned’ into Brazilian Portuguese as ‘to be involuntarily admitted to hospital’.29 This translation is superior to those provided by DeepL, Google Translate and Microsoft Bing Translator, but there are differences in Brazilian law between an ‘involuntary’ and a ‘compulsory’ hospital admission. Involuntary admissions are requested by other individuals (e.g., family members) whereas compulsory admissions are mandated by a judge.30 Patients or service users may not themselves be aware of this distinction. But without explanations that are sensitive to potential differences between jurisdictions, a non-English-speaking Brazilian patient detained in a UK hospital would be unlikely to understand the full details of their circumstances based on these machine translations alone.
Languages considered to have a lower level of digital resources than Portuguese are likely to have even more problematic translations. During the tests outlined above, GPT-4 translated ‘to section you’ into Swahili as ‘to place you under special supervision’ or ‘special care’.31 A professional Swahili translator I consulted for this project described this translation as simply ‘not correct’. In the translator’s assessment, there is no single Swahili word that would capture the legal meaning of being sectioned in a UK context, so a full explanation would be required. Artificial intelligence would here again fall short of informing Swahili speakers of the full circumstances in which they might find themselves. While Swahili is not as prevalent in the UK as languages such as Romanian or Polish,32 there are over ten thousand Swahili speakers in the country according to the last census.33
As translation models continue to evolve, inaccuracies like these will become rarer. But the terminological precision of fields like medicine and law are inherently opaque in cases, so the unreliability of AI-mediated communication will continue to be an issue in these contexts. Additionally, interlocutors are likely to continue to have limited means to check whether messages have been accurately understood. For the mental health advocate quoted earlier, it was only when they spoke to an interpreter that it became clear that information provided to patients was unintelligible. Even if the mental health advocate were to try to check understanding – for example, by asking the patient to describe in their own words what they had been told – these attempts may be compromised by the individual’s state of mind and the distressing circumstances in which they find themselves.
Checking understanding can be even harder when disseminating written information. A common written use of machine translation reported in the project was to translate letters and information leaflets. If translations are not provided, service users can – and indeed are likely to – use machine translation themselves to read the material. It was nevertheless clear that in some cases it was service providers who machine-translated letters at source before they were sent: ‘I have used it [machine translation] to write letters to distribute to patients’ (Trainee Clinical Psychologist); ‘When I have sent outcome letters to patients’ (Senior Physiotherapist); ‘Sending letters to patients’ (Management Assistant, Healthcare and Social Assistance). These professionals were clearly trying to ensure that those under their care could understand important information. As discussed in Chapter 1, however, asynchronous uses of machine translation can be particularly problematic if the technology is used covertly. If providers have some knowledge of the language translated into, they may be able to identify machine translation errors. But most project participants did not speak the languages for which translations were needed. The trainee clinical psychologist quoted above who machine-translated letters to be sent to patients was, for instance, a native speaker of English. They also had some knowledge of French. Yet they often needed to use machine translation for Vietnamese and Polish. Machine translation in these cases can both inform and misinform service users, and there may be no quick way of knowing which one is more likely.
Machine-Translated Consent
Providing a service to an individual in any way that requires access to information about this individual, or to their person, usually requires the individual’s consent. In the context of policing, consent is one of the mechanisms that ensure the legality of a search. A warrant is another one of these mechanisms, as is suspicion by a police officer, based on defined reasonable grounds, that the search is likely to uncover incriminating evidence.34 The legal basis for a police search will vary between jurisdictions, but especially when this basis is consent, machine translation can go to the heart of whether the search is considered lawful. As discussed in previous examples, if upon later examination machine translation is not considered effective enough to allow the officer and suspect to understand each other, the search can be voided, and any gathered evidence may need to be dismissed.
In medical contexts, patient consent is usually what allows a healthcare professional to administer medication or make any type of intervention in the patient’s health. There are exceptions to the need to seek patient consent,35 but in most cases obtaining consent from the patient is ethically and legally required.36 If machine translation is used to seek consent, any mistranslations that potentially distort or prevent understanding will affect the patient’s right to make decisions about their own care. Misunderstandings can also have consequences for the healthcare professional. Treating a patient without valid consent can be considered a civil or criminal offence, and in any event may lead to formal complaints and accusations of negligence.37
Consent is therefore central to most contexts discussed in this book, not only in policing and healthcare, but also in legal and social services. Machine translation mediates the process of seeking consent in many of these contexts. Of the project’s 828 survey participants who confirmed that they had previously used machine translation at work, twenty-one specifically mentioned consent when describing how or for what purpose machine translation had been used. Participants were not asked any questions about consent, so the unprompted focus on this subject is noteworthy.
Some uses of machine translation to seek consent took place in contexts where the environment’s ecology of action (see Chapter 3) is likely to give clues to service users about what needs to be consented to. For example, a doctor explained that they had used machine translation when ‘explaining a procedure and obtaining consent, for example taking [a] blood test’ (Doctor). Taking blood from a patient who is conscious and in a sane state of mind will usually require the patient to behave compliantly by sitting still and stretching out their arm. Once the patient gets into position, the doctor may start preparing the equipment that will be used to draw the patient’s blood. The patient may recognise some of this equipment (e.g., a needle and vials) and may understand what is about to happen just by watching the doctor and following body language cues. It may be tempting to argue that the patient will have an opportunity to object to the procedure even if they do not understand what the doctor is saying. The patient may cross their arms, walk away, shake their head or do all the above. It could be assumed, therefore, that using machine translation to gain consent from the patient in this case just adds a verbal layer of communication to an interaction that does not depend entirely on verbal language.
This assumption is problematic for several reasons. The first reason is that, as mentioned in Chapter 3, body language is culturally dependent. Not necessarily all patients would express disapproval in the same way, not to mention the possibility that some patients – because of culture or personality – may be less inclined to overtly demonstrate disapproval in front of doctors, police officers or other authority figures. A verbal exchange can make a difference in these cases. A verbal exchange is also what would allow the doctor to explain what will be done with the blood sample, why a sample is required, and the different courses of action that may be taken based on the results. It may be that all these details will have been explained beforehand, and the practitioner performing the test does not need to revisit the reasons why the test is necessary. But what if the rationale for the test has not been understood or the type of test requested takes the patient by surprise? Some blood tests, such as those used for sexually transmitted diseases, can be considered taboo. A presumption of non-verbal consent may be ill-advised in these cases.
The UK’s National Health Service recognises non-verbal consent. In its explanation of how consent is given, it states, ‘Someone could also give non-verbal consent, as long as they understand the treatment or examination about to take place – for example, holding out an arm for a blood test.’38 The key part of this statement concerns the need for the patient to ‘understand the treatment or examination about to take place’. As discussed widely in the medical literature, consent is about informing patients and not just about obtaining permission for the specific actions of touching their arm or using a needle to take their blood.39
If we take this broader view of consent into account, the number of project participants who referred to consent when explaining how they had used machine translation is in fact much larger. Many professionals did not use the word ‘consent’ but described how machine translation had been used for purposes such as ‘explaining eye conditions, treatments and medical terms to patients’ (Optician), ‘speaking with a Polish patient and their family to explain their treatment’ (Mental Health Nurse), or ‘to explain the procedure and what to do about the outcome [and] to explain the post care instructions’ (Nurse). Outside of healthcare, machine translation was used for purposes such as ‘to explain criminal law’ (Sergeant), ‘to speak to and explain a situation to an unaccompanied asylum-seeking child’ (Social Worker), or ‘to explain how to get help, to explain an offence to someone, to ask how I can help them’ (Police Officer). On the one hand, machine translation may be precisely what allows service providers to use verbal language in these cases. On the other, the inherently unreliable nature of machine translation tools may overemphasise non-verbal aspects of the interaction in ways that could induce service users to agree to procedures that they do not fully understand.
This risk needs to be considered in many of the performative interactions described in Chapter 3 involving uses of body language and the physical environment. But the ways in which machine translation is used to inform service users, and thereby to obtain their consent, do not necessarily involve physically guiding them or handling objects or specialised equipment. As seen above, it may also involve explanations of abstract concepts such as legal rights or a criminal offence, and here machine translation is particularly problematic.
In the cases above, the project participants did not mention retaining a written record of consent, which is not always necessary. But machine translation was also used for written consenting procedures – to translate ‘patient information leaflets and consent forms’ (Research Quality Officer, Medical/Healthcare) and ‘email/text to patient/client, information leaflet to client, consent form to client’ (Rehabilitation Case Manager, Healthcare and Social Assistance). A healthcare managing director used machine translation ‘to translate our informed consent forms and to ensure that the customer understood and could give informed consent’. Here the use of machine translation is not impromptu. Obtaining professional translations of consent forms is clearly possible in these cases. Especially for static documents that are used multiple times, procuring professional translations may be both possible and desirable. But here too, cost efficiencies can come in the way of reliable information access. Asking speakers of the language to check the machine translations can be a reasonable compromise in these cases: ‘In order to cut costs, we use automatic translators to translate … resources and letters which are then proofread by speakers of the languages’ (Health Mentor).40 A minority of participants referred to this type of quality control, however.
Consenting to Machine Translation Use
One other aspect of consent concerns whether the use of machine translation itself is consensual. This type of consent is central to the principle of user autonomy, which can be considered a cornerstone of ethical technological development.41 The concept of autonomy is notorious for its many definitions. A review of the human–computer interaction literature has concluded that autonomy is an umbrella term rather than a uniform concept.42 Autonomy does in any case have a few prominent interpretations, such as ‘the implicit experience that our actions are responsible for outcomes’ or ‘the material influence we have on a situation’ (emphasis in original).43 Both of these notions of autonomy are related to decisions about whether to consent to the use of machine translation. The project participants mentioned this type of consent too. When asked what advice they would give to colleagues about machine translation, a healthcare assistant mentioned ‘Ensure the patient has the ability to understand what you are doing and can consent to this.’ A social worker said, ‘You should ask the person what they prefer to use as well.’
User preferences are central to guidelines published in 2024 by the Stakeholders Advocating for Fair and Ethical AI in Interpreting (Interpreting SAFE-AI) Task Force, a group of stakeholders who issued guidance for the use of AI in multilingual communication.44 User autonomy is the first ethical principle in these guidelines. The principle states that end users should be involved in the process of designing and testing AI technologies and that ‘AI products that are procured and utilized for interpreting services must include explicit informed consent to accept/decline the use of AI’.45 The guidelines also establish that users should accept or decline the use of AI ‘with confidence that a human interpreter will be provided in a timely manner’.46 The guidelines allude to contexts where organisations officially procure AI products to be used in interpreting. Machine or AI translation tools are procured by many institutions, including hospitals and,47 as mentioned in Chapter 1, governments.48 However, most public service professionals who took part in this project had used freely available machine translation tools on an ad hoc basis.49 These were not tools that had been vetted and formally procured. In ad hoc contexts of this nature, some aspects of the Interpreting SAFE AI Task Force guidelines, albeit desirable, are not realistic.
For one thing, human interpreters are often not provided in a timely manner. Any attempt to make such a guarantee would at best be aspirational, whether because of rarer languages for which it is harder to find interpreters or because of the ad hoc nature of the context. Framing consent as a choice between AI and human interpreters can therefore be idealistic.
Defining the user can also be a problem. As mentioned in Chapter 2, primary users who deploy the tool may give rise to indirect, secondary uses of it. What if a patient insists on the use of AI but a doctor refuses on the grounds that it could cause harm? The question of which user should be granted more autonomy in this case is fraught with complexities. If the doctor believed that using AI could be harmful, it could be argued that the doctor would have a right to refuse to treat the patient in the absence of a human interpreter. Similar arguments can be made about other aspects of medical care. Malpractice aside, patients are not usually able to demand medication from their doctor, for instance. Even if patient requests need to be considered, the decision of what to prescribe is ultimately the doctor’s, so there are limits to patient autonomy. On the other hand, it is also true that, unlike medication, AI-mediated communication has not usually been considered part of the doctor’s sphere of expertise – at least not so far. The patient may be more familiar with the technology than the doctor, so it could also be argued that service users’ preferences about AI need to be respected even when service providers disagree.
Best practice in these cases is likely to fall somewhere in the middle. Service users and providers both have a right, I would suggest, to influence how the interaction takes place. While the prospect of this type of conflict is an important one to consider, the project data provided no evidence of a previous event where a doctor or other professional declined to offer a service on grounds that the service user insisted on using an AI tool. Service users’ preference for AI, on the other hand, was extremely common. This was often a reason why machine translation had been used.
Direct Reasons for Using Machine Translation
As mentioned earlier in this chapter, existing guidance on machine translation use and the front-line reality described by participants in this project do not necessarily align. This disconnect concerns not only the framing of consent but also the use of machine translation itself since the technology is used even where official guidance advises against it. I discussed some of the reasons for this disconnect in Chapter 1 and Chapter 2. Cost, the efficiency imperative and machine translation’s persuasive convenience will often speak louder than official guidelines. Testing and designing guidelines also take time. The speed with which new technologies become available, and their use normalised, may simply outpace attempts to design realistic best-practice guidance that reflects the challenging decisions professionals are expected to make. But over and above these broad incentives to use machine translation, my research found that there were often more direct reasons why users resorted to it. I discuss some of these reasons below under four key groupings: service user preferences, person-centred care, urgency and unreliable language services.
Service User Preferences
It will not come as a surprise that, for many service users, machine translation was the preferred method of communication. Using machine translation tools is now second nature to many. When completing the project’s survey, professionals from the selected sectors were invited to declare, based on their most typical experience, who had decided to use a machine translation tool. Often it was service providers themselves who had made the decision, but fifteen per cent of the 828 project participants who had used machine translation at work selected an option that said, ‘Someone I was speaking to started using it and I continued interacting with them in that way.’50
A midwife reports: ‘Women decline. Want to use their partners/phones for convenience [and] privacy (don’t want to use a [professional] translator).’ It is well known that service users may prefer ad hoc solutions to professional language mediation. A 2005 study of users of interpreting services in the UK found that users preferred the assistance of friends and family when they felt they might not be accurately represented by professionals, or when professionals were perceived as a risk to confidentiality.51 Other studies reported similar findings52 and previous research has also shown that some users value machine translation precisely because it eliminates the involvement of other individuals.53
Examples described in this book suggest that service users favour machine translation particularly in private or intimate contexts – for example, giving birth, as in the case reported above by the midwife. Service users who have some knowledge of the relevant language may also feel more capable of getting by without the assistance of language professionals. A sexual health adviser reports:
I was talking to a patient about testing for infections and explaining about window periods for testing. His English was good but there was some confusion as to what exactly I was trying to explain. I offered a telephone interpreter, which he declined, asking instead to use Google Translate to translate a small amount of information to help gain complete understanding.
In this case, not only does the communication involve an intimate subject (sexual health), but the service user’s knowledge of English is described as good. Machine translation here is complementary. The need for it may also be unexpected. Patients’ level of language proficiency may allow them to perform everyday tasks without language assistance. Communication difficulties may therefore take them by surprise if they overestimate their ability to understand specialist sexual health vocabulary. Machine translation tools are increasingly likely to stand out to service users (and providers, as I discuss below in the section on Person-Centred Care) as a first port of call in these cases. Especially when the subject is sensitive, seeking external help from a language professional may seem awkward and undesirable.
As discussed in previous research, however, the assumption that machine translation offers more privacy than a human mediator is problematic.54 In avoiding the involvement of human third parties, service users may in fact jeopardise the confidentiality of the communication by sharing it with the machine translation provider. The risk to individual service users is probably low. Technology providers would arguably be unlikely to target individuals with the intent of divulging their information to others. Leaks and information disclosure are nevertheless possible and have happened.55 The internet connection itself can pose security threats, as can the devices on which machine translation is used. While professional language services are also technology-mediated, the technical infrastructure used by these services is more likely to have been assessed and deemed safe in terms of privacy, confidentiality and information security. This type of assurance is unlikely to be in place when machine translation is used ad hoc. The convenience of just reaching for the smartphone in our pockets may in any case be more appealing, especially if users are unaware of the technology’s privacy implications.
Person-Centred Care
Sometimes the machine translation use arose out of service providers’ desire to offer a more personal service. A social worker explains: ‘The local authority I work for is diverse. I work with a high number of people for whom English is not their first language. As a social worker I need to be able to share and gather information … We follow a person-centred approach and the tools we use allow us to do this.’
A person-centred approach – or person-centred care – is central to many types of services. It is often discussed in the context of health and social care. According to the World Health Organization, ‘Integrated people-centred health services … is about building a long-term relationship between people, providers and health systems where information, decision-making and service delivery become shared.’56 It is also about ‘engaging and empowering people to have a more active role in their own health’.57 Communication is central to achieving these goals. Healthcare research has identified a series of enablers of person-centred care, and ‘cultivating communication’ is one of them.58 Cultivating communication involves ‘sharing of information regarding [a] patient’s condition and their own impact/influences on their condition’.59 It also involves ‘discussing and building capacity of patients for self-management and self-care’.60 Most individuals will be familiar with machine translation. They will use it in their personal lives and are likely to feel comfortable with the idea of using it to access critical services too. For the service providers, machine translation may seem like an obvious way of putting users at the centre of the service.
A different social worker shared a similar opinion: ‘I used an automatic translator to translate a social work assessment on a disabled child from English into Polish. I also used the [machine] translator to be able to understand their written feedback following the completion of the assessment.’ When asked what advice they would give to colleagues who were thinking of using machine translation, this social worker replied, ‘I would actively encourage them to use it, as I felt that it had a positive impact on the assessment process and felt more person centred.’ Other professionals agreed, although not necessarily by referring to person-centred services. Machine translation was not necessarily a technology of choice in these cases, but it could be a welcoming tool when language professionals were unavailable, as described by a trainee pharmacy technician: ‘[I used machine translation when] giving patients some basic advice during one of their appointments when we didn’t have a translator there in person. Using their native greetings and basic questions to make them feel welcomed.’
Some of these professionals were aware that using machine translation involved risks. When asked what advice they would give to colleagues, the trainee pharmacy technician said the following: ‘Do not use it [machine translation] for important or detailed information that could affect their [patients’] treatment/diagnosis etc (a properly trained translator/fluent staff member should be used for that) but it can be helpful for relaying some simple information (such as ‘are you thirsty?’) or using greetings to make them feel comfortable.’
This pharmacy technician used machine translation for the types of ancillary communication discussed in Chapter 3. They emphasised the person-centredness of these exchanges. Other professionals made it clear that they were not supposed to use machine translation, but that sometimes they judged this to be in the interests of the service user:
You are not really meant to use Google Translate. However if [it’s] just for conversational use such [as] ‘you are in hospital, we are trying to help you feel better’ instead of more personal details it is acceptable, depending on [the] circumstances. I use it when deescalating very distressed patients who otherwise would need safe holding and sedation.
This dementia support worker makes a complex judgement call. Although they are not meant to use Google Translate, they use it when it is likely to calm down a distressed patient. Waiting for professional language assistance in this case may involve subjecting the patient to more invasive and potentially unpleasant treatments, so the person-centredness potential of machine translation here is clear. While inaccuracies could further confuse the patient, if the patient is already in crisis, it is possible that the risk posed by mistranslations may be lower than the risk posed by prolonged distress. Although this dementia support worker did not make this reasoning explicit in their answer, they show tacit awareness of the need for this type of risk-benefit consideration.
Urgency
It goes without saying that in many of the examples discussed so far, the communication between service users and providers could be considered urgent. Many of the project participants were first responders (e.g., paramedics and police officers) or worked in emergency healthcare. Similarly, when psychologists and social workers spoke to service users in their homes or on the streets, time pressure was usually a feature of these interactions. Healthcare workers mentioned language assistance needs that arose in the middle of the night: ‘It is mainly on a night shift in the acute ward. Formal appointments really should have a human translator. Sometimes when this is not possible, we have used automatic translators as a tool’ (Infant Feeding Support Worker, Healthcare and Social Assistance). The cardiac physiologist quoted in Chapter 3 in relation to uses of a heart monitoring device mentioned using machine translation in ‘desperate’ situations. As mentioned in Chapter 3, this cardiac physiologist used machine translation when ‘translators weren’t available or booked and patients arrived’. They added:
I think in healthcare it’s always better to use a real-life translator, as I would believe they communicate better with the patient, and it feels more personal. I only use automatic translators when desperate, as I believe it takes away from the patient experience. It is helpful in times of need, especially as I’m part of an emergency team now and there are times during the night that patients arrive and don’t speak English … in times like that, an automatic translator would be beneficial to communicate with the patient. I think ensuring the patient is happy by asking them via the automatic translator that they understand is key too.
Unlike professionals previously quoted, this cardiac physiologist believed that machine translation was detrimental to the service user experience. They used it nonetheless as a last resort at unsociable hours. Like nurses, doctors and other critical workers, interpreters too may be expected to work night shifts. Many interpreting roles advertised online include night work, but positions vary greatly in terms of pay and conditions. In late 2024, a United States language company was looking for a night-time Haitian/French Creole interpreter. The interpreter could live in a range of different states and would be offered a starting rate of 0.40 US dollars per minute.61 At the same time, a hospital in Palo Alto, California, was looking for a full-time Spanish interpreter to work day shifts. The minimum salary offered by the hospital in California was 81,806.40 US dollars per year.62 The hospital was looking for a member of staff while the language company was looking for an independent contractor. Although the Haitian/French Creole interpreter was going to work nights, the level of pay offered for this post was significantly lower compared to the hospital position. Even if the Haitian/French Creole interpreter were to work a full-time schedule of forty hours per week – which in any event is unlikely – their annual earnings would still be far below those of the Spanish interpreter.
Several factors are at play in the differences between these two positions. Unlike freelance work, direct employment offers more stable income and more predictable working hours, and if these conditions are on offer it is usually because demand is also likely to be stable and predictable. Spanish is the United States’ second language.63 French is in the country’s third-largest language group based on the 2019 census, but the US population of French speakers is almost twenty times smaller than its population of Spanish speakers.64 The population of French/Haitian Creole speakers specifically – a subset of the broader French category – is even smaller.65 Budget holders at a hospital may baulk at the idea of hiring a French/Haitian Creole interpreter as a member of staff because the need for French/Haitian Creole may be intermittent. The hospital may therefore rely on intermediary companies that will themselves need to make a profit. So, in a paradigm where costs must always be reduced, the convenience of on-demand access to a French/Haitian Creole interpreter may come at the expense of precarious interpreting jobs.
Even if more attractive work packages were on offer – and in many cases they may be depending on the language and the country – it is arguably unrealistic to expect qualified interpreters and translators to always be available irrespective of the language. Emergencies therefore offer a compelling rationale for machine translation use. Guidelines prepared for the Victorian government in Australia have proposed a tiered approach to machine translation use considerations based on different levels of urgency.66 The guidelines state that unedited machine translations should not be used for emergency communication, except as a last resort when certain precautions have been taken, such as ensuring the material to be translated is clear, post-editing the translations if possible, or contacting previously selected individuals from the community who may be able to assist.67 All these measures are likely to reduce risk, but not all of them will be feasible on all occasions. In the same way that it can be difficult to locate interpreters in the middle of the night, contacting other human mediators could be a challenge. Post-editing the machine translations would in turn be best reserved to those with some level of proficiency in the source and target languages, which may not be the case of the machine translation users. The ‘desperate’ situations described by some of the project participants can therefore create dilemmas. There is a broader debate to be had in this respect in terms of raising standards for language professionals rather than normalising the use of AI. But it is also important to consider the current circumstances affecting public service workers, who may not be able to wait for language assistance in contexts where complementary or safer communication methods are not available.
Unreliable Language Services
In most contexts, timely services provided by a qualified human linguist are safer than informal uses of machine translation. But a factor that is easily overlooked in discussions of this subject is that professional language services can also be unreliable. The efficiency pressures which often drive uses of machine translation have had significant impacts on the language industry. A trend towards outsourced working models can be noted for both translation and interpreting services.68 Hospitals, courts, government departments and other public institutions often have contracts with large agencies that establish the link between language service users and providers. Sometimes this outsourced model does not go to plan:
Unfortunately, sometimes when we outsource interpreters to provide language support to our clients in a counselling session, they don’t show up, without any notice. This has happened several times. In some cases, I have been able to verbally explain to the client. In other cases, the client has zero English and therefore I have had to use Google Translate to type out a message to explain the situation to them.
The project participants were asked about their employer’s standard procedures for dealing with a language barrier. In almost all cases, these procedures involved the use of language professionals, but many participants were dissatisfied with the language services they had received. A psychological well-being practitioner said that human interpreters were ‘useful but sometimes the interpreters are not up to standard or don’t translate word for word’. Some professionals, like the counsellor above, mentioned instances where language professionals had not turned up: ‘Not appropriate when translators do not turn up or there is a [long] wait’ (Therapist, Medical/Healthcare); ‘Fairly satisfied [with my employer’s procedures] but the translators don’t always turn up’ (Audiologist). More commonly, dissatisfaction was due to urgency and language services that were too slow: ‘Not really satisfied as sometimes you need it urgently’ (Nurse); ‘It is normally good, but it obviously has to be booked in advance and can’t be accessed with short notice’ (Physiotherapist).
Participants also referred to the reliability of technical infrastructure. Audio quality, signal strength and hardware portability were all mentioned as reasons for dissatisfaction with remote human interpreting: ‘We needed to help a lady position on the table correctly and used it [machine translation] for that as we couldn’t sign to her and the language line [professional remote interpreting connection] we had had crashed’ (Theatre Scrub Sister, Health Care and Social Assistance). Similarly:
It’s incredibly time consuming. The phones in the department don’t have great signal and often aren’t clear to hear, and sometimes it takes ten minutes just to find a cordless handset with which to make the call. It’s also a really unnatural way of communicating and makes it difficult to clarify points or speak about sensitive information.
Poor signal or audio quality may also call the faithfulness of the interpreting into question. A police detective says, ‘Sometimes they [civilians] can’t hear them [interpreters] well enough.69 You’re trusting the translator is saying it correctly.’
On first impression, these comments do not point to anything new. The drawbacks of remote interpreting are well known.70 However, there is a risk that arguments against the use of machine translation might rest on assumptions that professional language services are always available and effective. In many cases, they are not.
Over and above the issue of irregular demand discussed in the previous section, one important reason why professional language assistance may fail to meet expectations is the underlying conception of language itself. The expectation that effective interpreting means relaying information ‘word for word’ was mentioned by the psychological well-being practitioner quoted above as well as by several other participants. Although such an expectation is often motivated by the pursuit of certainty, it risks missing the complexities of how languages relate to each other. Relaying information word for word would be akin to the so-called conduit model of interpreting, where the interpreter is seen as a neutral link through which information gets across.71 The conduit metaphor does not necessarily equate interpreters to machines.72 This is the prevailing model of interpreting in many contexts.73 But especially in medical and community settings, previous research shows that interpreters are much more than an objective channel.74 They instead play an active role in brokering understanding, and their ability to do so through a potentially unreliable phone connection is at best compromised (see the discussion of accuracy in Chapter 5). Logistical and infrastructural challenges can therefore be combined with expectations of objective linguistic equivalence, in turn leading to dissatisfaction with language professionals. Additionally, machine translation is easily available and almost invariably fast. In the words of a nurse, ‘Google Translate is at the touch of a button.’ In terms of speed and convenience, it would be difficult for language professionals to match machine translation’s appeal.
All these factors contribute to public service workers’ instinct to use machine translation when needed, even where this instinct should be resisted in the interests of safety or ethics. Working out when this type of resistance is necessary is one of the great challenges posed by this technology and an important question that cuts across the discussion provided in this book. Addressing this question is not easy, but in any event it requires addressing the issue of working conditions in the language industry. The uses of machine translation discussed in this chapter are associated with broader questions concerning the type and availability of professional language services and the value attached to the work of linguists within institutions and in society at large. Human and machine translation are intricately linked. I come back to this link in my discussion of trust in the next chapter and again at the end of the book.
Open access