22
Public Services Committee
Corrected oral evidence: Interpreting and Translation Services in the Courts
Wednesday 27 November 2024
12 pm
Members present: Baroness Morris of Yardley (The Chair); Lord Bach; Lord Blencathra; Lord Carter of Coles; Lord Laming; Lord Mott; Lord Prentis of Leeds; Lord Shipley; Baroness Stedman-Scott; Lord Willis of Knaresborough
Evidence Session No. 6 Heard in Public Questions 86 - 93
Witnesses
I: Professor Richard Susskind OBE; Daniela Ford, MSc Translation & Technology Director, Centre for Translation Studies, University College London; Professor Sabine Braun, Professor of Translation Studies and Director of the Centre for Translation Studies, University of Surrey.
Professor Richard Susskind, Daniela Ford and Professor Sabine Braun.
The Chair: Welcome to this public session of our inquiry into interpreting and translation services in the courts. I start by thanking our witnesses for coming to talk with us today and asking them to introduce themselves. Daniela, I will start with you.
Daniela Ford: Thank you very much for inviting me. I am the MSc Translation and Technology Director at the Centre for Translation Studies at the University College London.
I am a practising translator. I started my career in 1994. I am also teaching machine translation and generative AI, and I supervise a lot of dissertations related to machine translations all over the world.
I speak three languages fluently and have learned another 10 languages. What I want to say by that is that I love languages. I love technology related to languages. So this is my background.
My final point: I am not an interpreter. I have done some interpreting in my degree course, but I am mainly a translator, which means written translation. I do that daily.
The Chair: Thank you very much.
Professor Sabine Braun: Thank you for inviting me as well. I am a professor of translation studies at the University of Surrey. I am the director of the Centre for Translation Studies at Surrey and the co-director of the Surrey Institute for People-Centred Artificial Intelligence. I am a translator and interpreter by training, but I am now a full-time academic. I teach and research mainly interpreting and the use of technology in interpreting.
Professor Richard Susskind: I am president of the Society for Computers & Law and from 1998 to 2023 was technology advisor to the Lord Chief Justice. I have been working in law and technology for 40 years. I wrote my PhD on AI and law in the early 1980s—horrendous thought—but that was a long time ago.
AI is advancing enormously just now. I have just completed a book called How to Think About AI: A Guide for the Perplexed. What I can help you with today is past problems that we have had with court technology. I can help you think and plan in relation to the fast-moving field of AI. I will also maybe want an opportunity to say something about new forms of dispute resolution so that we are not focusing simply on courts as they are today.
The Chair: Thank you very much. John, I will ask you to start off with the first question.
Q86 Lord Shipley: Welcome and thank you very much for being with us. I want to ask you, first of all, about remote interpreting and specifically how effectively is remote interpreting currently utilised in court settings and what are the barriers to effective implementation?
Daniela Ford: I will have the shortest answer to that because I know a lot of interpreters, I have been around for a long time and I speak to interpreter colleagues. I know that one of the big barriers is sound quality and audio quality from a technical point of view. I know that if interpreters are asked to interpret via Zoom, they do not always see everyone in the courtroom so that makes interpreting really difficult, actually, when you do not know who is speaking.
Preparation is a big issue as well, when interpreters are asked to go into a courtroom without having been briefed on what the case is about so that there is no time for preparation.
That is my short answer. Sabine will be able to say a lot more about that.
Professor Sabine Braun: That is one of the topics that I have researched quite a lot. I have actually written my own PhD thesis about video interpreting but that is also a long time ago. What we see now in court, I think, is still a very mixed picture in how effectively remote interpreting is used.
Remote interpreting in the narrow sense, where the interpreter is separated from everybody else, was used in English courts but started only during the pandemic. This kind of interpreting is used in American courts, for example, and in other places. However, there is another kind of interpreting that we could also cover under remote interpreting. That is when you use videolinks in court to connect to prison or to police stations. This is very common and is also a situation where interpreters are very often needed.
Challenges arise there as well. The challenges arise when the interpreter is separated from the person, who very often does not speak the language of the court. As Daniela has said, there are often sound problems, visibility problems, depending on where the cameras are in court and how people position themselves in front of the screens. As an interpreter, you may or may not have a good view of the remote defendant. That is problematic because spoken communication also means that you need to see people’s non-verbal communication. All of that leads often to greater uncertainty.
We talk about a lack of presence in videolinks. We learned a lot about that in the pandemic, working online, but still it is not so easy to overcome that lack of presence. It often leaves a slight uncertainty over whether the message you have rendered actually arrives at the other side. That also means that many interpreters feel that remote interpreting is more tiring. You have to concentrate a lot more. It is more stressful. We also have some evidence that it could lead to a loss of quality.
Then, in particular, when a defendant is remote and the interpreter is in court, many of the current videoconferencing platforms and systems do not allow the interpreter to interpret simultaneously. Normally in court, when the legal practitioners speak in English, when they do not address the defendant directly, the interpreter would do simultaneous interpreting. They would whisper the interpretation to the defendant so as to not take too much time and so on. That is, of course, not possible when you have a normal videoconferencing system and your defendant is remote.
In English courts I think it works in Crown Courts now but I believe there is still no good solution in these court-prison videolinks. So that is not very effective. That is one of the main challenges because then the interpreter needs to wait and interpret consecutively, but the legal practitioners do not always give that time because courts do not normally have a lot of time, so the interpreter does not always get a chance to interpret, to have the time to interpret. So what interpreters do—we have seen that in our research—they try a mix. They sometimes do simultaneous by talking over the others in court. Then they sometimes slip in little bits of interpreting in quick pauses. You can imagine that is quite a disorderly way of interpreting, which means that there is quite a high risk that some information is literally lost in translation.
The main barriers, I think, are that the video systems and platforms that are used are still—well, they have not been designed with bilingual interpreted court hearings in mind. They would still have to be redesigned, I think, to become more effective. Some of the technology is outdated. The technology is also inconsistent. We see at least three different systems being used. I know this is changing, but I think it is still the case that there are several systems.
The third barrier is that there is, I think, also a lack of understanding of some of these issues that videolink communication brings. There would also have to be probably more training around that to make sure that everybody knows how you communicate in a videolink and how you recognise when things are about to go wrong in the communication, for example.
Lord Shipley: Thank you.
Professor Richard Susskind: I will not overlap or repeat. One point I wanted to draw out, related to the use of remote courts or video courts during Covid, is that around the world, and certainly here too, this is not an uncommon way of conducting a court hearing, and I think we will see much more of it.
The Government and the Society of Computers & Law set up a resource that you might find interesting, called Remote Courts Worldwide, where we monitored developments across the world during Covid. More than 170 countries embraced remote hearings.
We also gathered together and undertook quite a lot of research into the levels of satisfaction because if we are going to have remote hearings, obviously we will have to have remote interpretation at the same time. The truth is that the feedback is very mixed. If I can generalise hugely, by and large, lawyers and judges prefer to be in physical courtrooms, and one can understand that—that is how they have learned their trade and always worked—but there is a quite high level of satisfaction among the users of courts. They found it less forbidding. They found it more convenient. It meant that they did not need to take a day off work. One of the conclusions a number of us draw is that if we are trying to make our court system more proportionate, if we are trying to offer value for money, it might make sense for more hearings to be held on this basis. Certainly, among senior lawyers and judges, there is a view for quite a large class of procedural hearings, maybe not final hearings, to be held in this way.
The point I want to emphasise is that if we are going to move away, in some cases, from physical courts to different ways of resolving disputes, this by necessity will require us to provide translation and interpretation services in new ways. All the signs around the world are that if we are going to provide a more affordable, accessible service, we will have to look beyond people congregating at one point in time and space to resolve often quite low-value disputes.
Q87 Lord Shipley: The evidence that we have received from some has suggested that the main driver behind remote working would be reducing travel time and travel costs but also increasing the pool of available interpreters. For the court system, which has a lot of problems with managing the number of cases it has in front of it, that would help a bit to increase the speed of dealing with cases because there would be more people involved.
The other side of that might just be outcomes; in other words, my question really is this: is there any evidence as to what the impact on the quality of outcomes is? Do we know? Is there research that says that we get better outcomes if we are remote or that we think we get worse outcomes, or is it simply that people do not quite know?
Professor Richard Susskind: Are you asking about the outcomes of court hearings themselves or about the process of interpreting?
Lord Shipley: About the quality of decision-making.
Professor Richard Susskind: Yes, there are lots of studies on that. The most interesting one, I think, is in British Columbia in Canada, where they have been delivering services in a variety of different ways. I was an expert witness relating to a dispute that arose over whether or not that service was constitutional, so I had the opportunity to look at a lot of the work globally. You would like a crisp answer but the feedback is mixed, I am afraid. For certain types of cases, both parties and lawyers will say that kind of disposal is more appropriate.
To take a step back, there are physical court hearings, there are video hearings, there are things that a number of us call online courts where cases are decided on the papers alone. They are three different categories of hearing.
We have done a lot of research on video hearings. I would say that overall— and it is going to be a difficult policy decision for any Government to make–there will be strong evidence in certain categories of cases that the quality of service will be higher when you use some kind of online technique. I have written a whole book on this called Online Courts and the Future of Justice. My instinct is to be very supportive of this use of technology so to some extent my comments need to be taken in that light. But I analysed the concept of justice under seven different headings and came to the conclusion that there were many categories of case for which justice would be better served by not being in a physical courtroom and using alternative methods. So I do not think this is a fringe interest. I think it is going to be fundamental to the future of justice.
Professor Sabine Braun: Can I add to that? There is evidence. I want to point to a piece of evidence on the use of videolink use in remand hearings. A House of Commons Justice Committee report in January 2023 looked at the role of adult custodial remand in the criminal justice system. There were only two pages about the videolinks, but the report pointed out that videolinks are potentially difficult for vulnerable prisoners. It does not say anything about people speaking another language, but I think we always have to bear in mind that when we look at prisoners with additional needs and prisoners who do not speak the language—and that also, for me, is closely related to this category—that we need to be extra careful because if somebody cannot understand for whatever reason, if somebody needs an interpreter, there is always an extra layer of complexity in that communication. I think we need to look at those cases separately if you are asking how effective videolinks are. Videolinks where these demands do not exist may very well be effective, but videolinks where we have an interpreter involved tend to be more complex.
In research, we do have evidence for what happens to the interpretation. That evidence is mixed. Some research shows that there is an impact on the quality of the interpretation in legal settings. When interpreters are in the videolink, we also have research that shows that the interactional dynamic changes. For example, the normal turn-taking that you have in more or less dialogue situations, that you have in many legal settings, with witness examination, for example, it is more complex to organise that when you are not in the same place. The small non-verbal signs that we normally use to agree—When am I finished speaking? When do you start speaking?—that is all more difficult to co-ordinate, which impacts on the cognitive load that ultimately an interpreter has to carry because they often have to organise that.
So it is very difficult to say that we have exact evidence that shows us that the quality goes down every time but we have evidence that shows that there is a risk that the quality of interpreting is affected. If the quality of interpreting is affected, ultimately the quality of justice is also affected. I can provide all that evidence if you like?
Lord Shipley: Do you want to add anything, Professor Ford?
Daniela Ford: I am thinking of an example of one of my interpreter colleagues. She was offered an interpreting job at the last minute. She was called at 5 pm for a job at 9 am the next morning. It was not made clear how the interpreting was supposed to be done. In the end, it was done on the mobile phone by the solicitor, who was just showing the mobile phone around the room to show the interpreter who was in the room. The meeting started late. There was supposed to be a one-hour break in between. There was supposed to be another meeting afterwards. The interpreter was never called back for the next meeting. I think the cognitive load and stress on the interpreter is enormous. Also, I underpin everything that Sabine just said because you are outside the room. My friend also said that she had to interrupt the solicitor because there was no clear protocol about what the interpreter was supposed to do. There are lots of issues where the interpreter almost becomes an add-on but is not in the room. Therefore, there is no recognition. There is no support. That is what I wanted to add.
Lord Shipley: Thank you very much.
Q88 Lord Mott: I have a question in two parts. Perhaps I can start with Richard and then move along.
What are the possibilities of introducing AI tools into interpreting and translation services in the court? What steps would need to be taken to begin this process? I am thinking, in light of what you have already said, about timelines and potential investment because the investment side, considering what is there at the moment, feels as though it is a potential barrier.
Professor Richard Susskind: I have to take quite a long run-up at this question because the concept of AI is so widely discussed and so impressive are the recent developments that we do need to take a step back and put it in context.
At a very high level, my view is that many, if not most, of the short-term predictions about AI are overstating its impact, not just in courts but right across our community and society. However, and more fundamentally, I think most of the long-term predictions are understating its impact. If you are asking me whether we could transform interpretation and translation in the court system over the next two or three years using AI, I would pretty emphatically say no. If you are asking me to look forward to 2035—and I am unclear about the timescales that are your focus here—there is a strong view across the AI community that by 2035, we will have achieved quite remarkable developments. I think we need to distinguish between advances in technology on the one hand and the adoption of technology on the other; that is, we are seeing a fairly clear exponential growth in many aspects of the technological advance but adoption is a far more jagged affair, which depends on all sorts of factors, cultural, regulatory, and so forth.
What I want to say to you—and this is really coming out of my recent research and thinking—is that there are many different views among AI scientists about how quickly the technology is developing. The three that I think are of interest are as follows.
The first view is what I call the hype hypothesis, which is that actually all of this is just being overstated and overblown, that the systems are superficially powerful, but actually they make too many mistakes and we will never really be able to rely on them and the bubble will burst. That is one school of thought.
A second school of thought is what I call the generative AI-plus school of thought. You will all be familiar with ChatGPT. It is an example of a technology known as generative AI, where these systems generate text and sounds and video. They can even write code. Many people say that what they do today is phenomenal, and I am one of them. But I also point out that the data on which they rely suffers from bias and these systems often make mistakes, which are metaphorically called hallucinations. Those who believe in the generative AI-plus hypothesis say that we accept all of that but there are billions of pounds being invested in this technology, there is clear market demand and that it is not unreasonable to say that by 2030, the current shortcomings of generative AI will be overcome and these systems will be, by and large, reliable—maybe not perfect, but better than many humans in many respects.
The more challenging hypothesis is the third one, and that is called the AGI hypothesis. I am sorry to pause on this because there is simply no point in talking about AI on its own and I want to give you some context.
The AGI—artificial general intelligence—hypothesis is the view that at some stage these systems will be able to perform all the cognitive tasks that human beings can. If you had asked most AI scientists, maybe three years ago, about the prospects for AGI, they would have said it is 20 to 40 years away, if ever. If you ask most today, they will say it is five to 10 years away—not all, but most. Within that group, there is a group that supports something called the scaling hypothesis, which says that to get to AGI, we do not need any new technological inventions; all we need is to have more data and more processing power. So it is massively uncertain, and I cannot sit here and say to you, “Here is how it is going to unfold”.
My day job is helping big organisations think about this and we do scenario planning. You say, “Here are three realistic scenarios” and you ask the question, “What if?” What if the hype hypothesis proves to be true? What does this mean for interpretation and translation? What if generative AI and what if AGI? I think the defining question of our age, actually, is what if AGI because if AGI comes about, it will constitute the biggest and most fundamental change to humanity, actually, I think, over the last 300,000 years. But it may not come about, so it is little like an asteroid coming towards us. We have to devote some time to diverting it or at least understanding it, but it may not happen.
In the broadest of terms, if I were to be pinned down, I would say, “Expect generative AI, the technology we have today, to be pretty problem-free by 2030 and at least plan for AGI by 2035” but this calls, therefore, for an incremental approach. We have to think of what the road map will to look like because we cannot automatically assume that by 2035, we will have these systems. We talked about this briefly outside. I think we will have to move at a modest pace over the next few years to see how the technology unfolds but part of the policy would be having in place contingencies to allow you to move in that direction.
Now, what I am saying is, to some extent, entirely unhelpful. There is no evidence from the future. I cannot tell you what the future is going to look like. But I think that there are two pretty likely futures, and policymakers will have to get their heads around this possible bifurcation. I think we will know within five years how the future is likely to unfold.
What I discourage people from as a mindset is what I call technological myopia, and that is evaluating the future potential of technology in terms of today’s limitations. People will say to you that these systems are inherently unreliable, that they hallucinate and all the rest of it. That is, of course, true but, according to my research, we used to see breakthroughs in AI every five to 10 years. We are now seeing them every six to 12 months.
Finally, I just thought I would paint the picture by way of response, and this is the hardest thought of all, that the most powerful technologies have probably not yet been invented. We are only 75 years into AI, and in the grand arc of human history we are at the foothills just now. We are feeling our way. A lot of what is going on is experimental. It is in that spirit, I think, that you need to think about the impacts of AI. There is a world in which not just for interpreters and translators, but for most professional people, their work will be replaced by AI. There is another world where, actually, it will constitute a remarkably powerful tool for workers but will not replace them. I cannot tell you sitting here today which scenario is likely to unfold, but I think we need to plan for both and not just in this Room in this building.
The Chair: Does anybody else want to come in on that?
Professor Sabine Braun: I was not going to talk a lot about timescales. You have talked a lot about timescales. I would like to remind you that six years ago, the then Lord Chief Justice said that in a few years, we will not need interpreters any more. That prediction, unfortunately, is simply not true. I hear what you are saying, that of course, technology is speeding up that development but I think we need to be extremely cautious about these time predictions. I cynically sometimes notice that people—I do not mean you—make these predictions for around, say, in five years because that is probably a long enough time for us to have forgotten. I think we need to be extremely cautious. I agree with what you are saying, that we also need to plan, but it is very difficult to plan and we need to take these things one step at a time.
Here is another thought I want to bring in. I also think it depends very much on what kinds of AI uses we are actually talking about. So if we are really talking about full automation, machine interpreting, I do not think these systems are ready and I think that would be very difficult to say. Will they ever be ready or is interpreting something that is actually too complex for a machine to tackle, or is it so far away that we cannot plan for it yet?
However, I also think it happens gradually. So many translators already, of course, work with AI day in, day out. We will probably say more on that. Interpreters are not really doing that yet to the same extent but I can see in the research community and from talking to interpreters that there are already uses of AI that assist interpreters in their daily work.
One good example is an AI usage that would also be helpful for many other scenarios in court, which is to use much more live transcription. For example, in magistrates’ courts, we still do not do recordings. It would be very beneficial, I believe, for many court users if they could actually see what is being spoken. It is not just for people who do not hear well, it is for people who speak English as a second language, who when they go to court may not need an interpreter. I may not need an interpreter if I ever have to go to court, but I have to say, I would probably feel more confident if, like in the opera house or in many theatres, I could also read along what is being said. It could be used as a record later on.
It may also not be perfect. A raw transcript created by AI will also have mistakes, but it is a good starting point. It could be what a court user can take home to reflect on. It could certainly be helpful for an interpreter—we are beginning to see positive results around this research—especially in situations such as I outlined earlier, where things are hectic in court, where you interpret on a videolink, where you cannot hear everything well, where speakers overlap. It may provide the interpreter with an extra tool to achieve a more complete and accurate interpretation. We are doing research around that. Others are doing it too and first signs are positive.
You can also look at other assistive uses of AI. Summarisation by AI might be helpful to summarise court documents, to have a record to pass documents on to court files. Again, I am also looking for technologies here that are not just for interpreters but that could have a wider benefit for all.
In terms of investment, again, I am not an economist. I cannot speak to that but I would also think that it depends very much on what we are investing in. If we are investing in these assistive technologies, that is one thing. If we are investing in, or looking to invest in, anything that has to do with full automation, then, of course, I think that regardless of the actual amounts of money, we have to ask ourselves a lot of questions around fairness and justice again; namely, how would such a system even be conceived? Would we then have to cover all languages that we need in court? I believe that in London people speak around 300 to 400 languages and dialects and rare languages and indigenous languages and what have you. That would be so complex to even conceive that I think it is very far in the future. You also have to remember, if we talk about court, that we often really talk about interpreting, so something that is immediate. That is even more difficult to achieve than it is in written translation.
I could talk for a very long time, I suppose, about why this does not work at the moment in speech-to-speech translation for all the languages that we need. We just have to be careful that we do not end up with a solution that may work for some large languages and where we invest and where commercially there is already, of course, investment in making the translation for these larger language pairings more robust. We have heard before that in some language pairs that are needed in court, there is less commercial interest, so there is less investment. Besides the fact that I do not believe that this will work any time soon anyway, if we do that as a thought experiment, we would also have to bear in mind that if we do not do it appropriately, we might end up with a situation where somebody who speaks one of these larger languages comes to court, has lots of choices between good machine translation systems, a large and good pool of interpreters, versus for somebody who speaks a small community language, poor-quality translation systems, small pools of interpreters, problems there with interpreter training with qualified interpreters and all of that. So we are creating and amplifying the digital divides enormously, if you think along those lines.
Daniela Ford: So just a few points: I am a translator, really, so I deal with the written text. I use machine translation AI in my daily life, partly because my clients request me to use it. My clients will provide the machine translation for me, and this is one important point I want to make. There is a real difference between free, the accessible machine translation tools and paid-for ones that you subscribe to. I will give you a real example, I have taught translators and terminologists in a very big government organisation and one of the translators told me that he was under time pressure, that he needed to get a translation done quickly; he put it into Google Translate to get it done quickly. It was a highly confidential text. He was not allowed to use Google Translate but he did it anyway and that whole source text appeared on the internet the next day. This is one thing that I really want to bring home, that using free machine translation is incredibly dangerous, certainly in a court setting. I would not even dream of using anything that was free. So that is one point.
I do like the approach of several big organisations that I have also worked with and taught, which give their interpreters and translators a choice. They do not say that you have to use a particular tool but they say that if you want to use a tool, tell us what you want to use, and then we will say whether you can use it. It is also a personal preference. I am very much into technology, I like technology, I find it interesting and fascinating. There are colleagues who do not share my views at all, and that is fine. So the confidentiality is a very big aspect.
I also translate software and games and mobile apps, also for generative AI, so I see a little bit behind the scenes, see what is going on because a lot of these systems are secret, so you are not allowed to know how they work. I test some of these systems and this is something that I thought of in the context of court interpreting. You will get emotional language in a court scenario, you might get swear words, you might get offensive language. Now I know for a fact that a lot of these systems put blocks on offensive words. If I tried that, because I do it as testing, they will block my tester account because I am using really bad words. I can see this as being a really big issue. Obviously, one of the issues is mistranslation.
In my daily life, I do see that these systems have come a long way. I printed out a short timeline. You know, machine translation has been around since the 1950s so this is not new technology. I remember that in the early days, the 1980s and 1990s, I was sometimes given a machine-translated text and I was told to correct it. I said, “I cannot correct it. I have to start again. There is nothing correct in this”. We are in a different place right now, but it is very much dependent on the language combination you work with. It is very much dependent on the source text and that we know what the subject is. It is also very much dependent on how well the source text is written, whether it is written with technology in mind. Spoken language is never perfect. You have grammatical errors; you have all kinds of errors. A machine has no feeling. It has no emotion. A machine cannot pick up on these things, which is where interpreters are really needed. So I can see a lot of issues with confidentiality.
I know law firms that use free machine translation tools, ChatGPT free. I cannot believe what they put in these tools because it is very difficult to see what the big tool providers are doing with the data. I was looking at some of them. A lot of them will say, “You are still the owner of your data but we will use your data, for sure”, because obviously they want to improve the system. This is a really big deal. I know people who work in prisons. I know people who work in the healthcare industry. I know that these free tools are used as a last-resort solution. I find it incredibly dangerous, if this is the outcome of a court case, if you are relying on that. Machine translation systems sometimes omit half the sentence, they leave out sentences, they change things.
I will give you a concrete example of what was said earlier about bias. These systems can only be as good as the data that is in them. There is a very common example that if you talk about a doctor and a nurse, quite often the doctor will come out as he and the nurse will come out as she because that is the data behind it. You have issues such as gender-neutral language. You have “they” in English, which you can nicely use for he or she. In many other languages, that is not possible. You cannot do it. There is not one word in many languages. You have to say, is it she or he? This will cause a lot of problems if you imagine that this gets translated as the plural “they”. So there are lots and lots of problems on that front.
So the whole MT, as I say, I like it, it has come a long way, but I can only ever see it as an assistive tool. I always think that you have to be a really good translator. You have to be a really good interpreter to be able to use it in the first place because the nuances are so tiny sometimes that on the surface, you think this translation is really good. You dig deeper, you find the most terrible mistakes ever. Without being a good translator, a good interpreter, it is incredibly dangerous to use any of these tools.
The Chair: Thank you.
Q89 Baroness Stedman-Scott: I would like to ask how you believe the use of AI tools and remote interpreting will affect the role of interpreters and translators. You touched on it to some degree but could you be specific in your responses? Could you also think about how translators and interpreters make use of such tools to support their work, and what training and education is available for them, and is it accessible, is it too costly? Daniela?
Daniela Ford: There are many questions in there. Maybe I will start with the training side of things because I am teaching at UCL. We are training future translators and interpreters. Our focus has shifted a lot towards data preparation and data manipulation. A lot of our students will work for big companies and government organisations to help train the machine translation systems. They will learn how to build machine translation systems. They learn programming languages such as Python to learn how to programme. Basically, they will act as linguistic advisors, alongside the developers, to help get a better outcome with those machine translations.
We have certainly seen a drop in traditional translation because of the use of AI. My favourite example is the Industrial Revolution in England. When the Industrial Revolution started, a lot of people were scared about losing their job, and sure, some did lose their job, but those who were flexible and adapted went into new roles or got new jobs, and this is how I see where we are with translation at the moment. We need to be flexible. There will still be niche areas where pure traditional translation will be needed but there are many other areas now where a lot of programming skills are needed. When I work, I rarely translate word files, I will translate all sorts of complicated software files that I never thought I would touch. There is a lot of technology involved in the translation industry now. This is where we are going with the training, much more heavily tool-oriented but to be able to consult on how to use the tools properly and also to be able to improve the tools. We also teach speech recognition, so we have students working with speech recognition systems to help train them. So this is the key part.
As for the other part, what I also am seeing a lot now in industry is that we have almost reached another plateau already with generative AI and machine translation. It became very exciting when ChatGPT came out, when neural machine translation came out, but I am not seeing a lot of big changes at the moment. What I am seeing now is that a lot of people are looking at terminology now. They are trying to improve the results by injecting terminology, by building glossaries to get the terms right, and they are also starting to use something called artificial intelligence-based post-editing. They’re starting to add something like ChatGPT, but not the free version, the paid-for version, because that is very different. They are starting to add that as another phase to say, let us get the machine translation first, and then let us get another system that will polish, which is actually technology that was around before already in machine translation, so you can almost see it a bit like a replacement strategy: get the right terms, but you can also feed in stylistic information, and this is where we are going now. The tools that we are using as translators are already starting to integrate generative AI now to help with that, so it is becoming a bit of an authoring tool.
Professor Sabine Braun: Perhaps I can add a few thoughts to that, specifically in the area of interpreting, where I said earlier that these uses are not as far developed as in translation, but like everywhere else, interpreters will interact a lot more with technologies in the future, and even in the near future. The question was around how remote interpreting and using AI affects their role. Earlier we pointed out the challenges with remote interpreting; certainly, I did. That is, by the way, not to say that I think it does not work, it works when the people and the systems are right, so we need to spend a lot of time on thinking about that. Also, when people have training, we must also be fair. As soon as you give interpreters the opportunity—as we certainly do in our courses and elsewhere—we have sessions on remote interpreting, where students can try out in a safe space how it all works. That can also be done for interpreters who have already been in practice. They can have continuous professional development in this area.
Many years ago, I trained the then roughly 400 police interpreters in the use of these video interpreting systems when the Metropolitan Police in London introduced them. It was a short training course but it was, I think—I hope—somehow helpful to get people going. With training, we can optimise the use of remote interpreting and that goes for video interpreting and also goes for these AI tools.
One benefit of AI tools, I think, is generally that they are now at a stage where they can be customised, as you say, so we can train interpreters in even customising what I was pointing out earlier, the speech recognition. Interpreters can inject their own glossaries there and suggest a translation to make them more accurate. We also have already done research on customising these systems for legal proceedings. That is all perfectly possible. Another good thing about these AI tools is that they can be personalised so we can also train interpreters in making use of these AI systems in the way that it works for them. AI is not really, “Yes, we use it”, or, “No, we do not use it”. It is a gradual and it is a more or less. Some interpreters may feel happier about it. Others will not feel that they need it. We need to provide the training opportunities, and that brings me to my other major point here, which is about how we deliver this training.
I believe in this court setting where you have a workforce—yes, they are freelance interpreters, but they are not the same kind of freelancers that do all kinds of work; they are a relatively homogeneous workforce. I believe that somebody from the agency that supplies interpreters said the other day that they cannot provide training because they are freelancers. I would like to disagree slightly with that. I think we should look at training solutions where everybody has a share of responsibility and where we make sure that the interpreters are treated as a workforce and where we give them opportunities to train and to upskill. That can come from many different sources.
By the way, in my view, training works best when we do this as a joint effort. We did some of that in projects, research and training components where we have trained legal practitioners and interpreters together in how to work together when you communicate through an interpreter. You can always hear both sides. That is very effective in training and we can do the same to speed people up in the technologies because the legal practitioners, not just the interpreters, need to understand these technologies. That is my plea, that we should do much more around joint training, systematic training opportunities, so that we upskill.
By the way, I think training has a lot of other good side-effects for interpreters. Interpreters often work in quite an isolated manner. They work alone in court. If interpreters have the opportunity to upskill together, to discuss these technology issues together, I think it is also hugely motivating and it has the benefit of forming a workforce. It is a soft benefit, but I think it is there as well.
Professor Richard Susskind: I will zoom out a bit and make a more general comment. I co-authored a book called The Future of the Professions, where we looked at the impact of AI across eight different professions. For those of you who are interested in law, lawyers are the most conservative other than the clergy. We looked at the clergy as well. What is fascinating is that almost every professional sees far greater scope for AI in professions other than their own. We all do it, I am afraid. I am doing a lot of work with doctors just now and they simply cannot imagine a world that is different from an automated version of what they have today.
Our research found that there are two futures unfolding. One is reassuringly familiar for the interpreter and translator. That is where that job pretty much continues as it is today but translators and interpreters have at their fingertips some remarkably powerful technologies to enhance what they have always done. That is what that we have been hearing. My own view is that that is definitely the short to medium term. However, my interest in life is the longer term. I have sat in this Room or rooms like it before and kicked myself a decade later for not saying what I am about to say, and that is that I think these technologies are moving at a phenomenal rate just now, and just because we do not know how they will overcome the problems we have heard about, we simply cannot say, “Here are the worries, here are the issues, so let us stay where we are”.
I do not want people looking at your report in 2035 saying, “How on earth could we not imagine that there would be a second future?” and the second future is not when professional people—interpreters, translators—are AI enabled, it is when they are AI replaced. I am not fan of that. I am not a supporter of that; nor am I an opponent of it. I am simply here to say that after 40 years of working, looking at the trajectories, we have to take seriously the possibility, even if the experts cannot conceive that this might happen, and given the overall rapidity of advance in AI that we are seeing, that—and it is not just interpreters and translators—much professional work will be replaced by these increasingly and almost unfathomably capable systems. This is the first place where I say this publicly.
I wrote this book about AI, and I have always been a big fan of AI, and I have always thought it could bring about solutions to some of our greatest challenges—to climate change, to our health problems—but as I wrote more about it, and I thought more about it, I could see those huge risks as well. So I say that saving humanity with and from AI is the defining challenge of our generation. It is very easy for a group such as this to say, “Let’s just look a couple years ahead and get some better tools for interpretation and translation”, but I am simply not going to miss the opportunity to say that we have probably about 10 years to get our thinking straight, not for our thinking to be determined by technologists and tech entrepreneurs, but for serious policymakers and parliamentarians to say that we live at in time of greater technology, more advanced than humanity has ever witnessed, and we at least have to countenance in all our discussions the possibility that these systems could be far more advanced.
This is a trivial example but I think back to discussions about cabling court buildings, and keeping on saying, “You know, wi-fi is going to come along quite soon, you do not need to be doing this”. That was the feeling I had when wi-fi came along. I do not want to have that feeling in 2035. I really am definitively saying that I believe AGI is going to come about, but I think that, if you are looking at anything beyond three to five years, you have to take seriously the possibility that the future is not turbocharged interpreters and translators but is something quite different. I might be wrong but I think that the risk or the possibility is sufficiently great that it is a possibility that I would like you to contemplate.
The Chair: Sabine, you wanted to come back on that.
Professor Sabine Braun: I can make it short. I appreciate what you are saying. I do not think it is that we cannot conceive of this possibility. I think with the linguistic expertise that we have, and we are closer to this, I think what we are trying to say is the following. We need to be cautious and it is very much about the way in which we would implement such solutions.
Now I would like to jump back to the remote interpreting for a minute. Videolinks in court have been there since the 1990s, and we still to this day do not have a very effective solution for how we integrate interpreters into this. I am not going to say how many decades that is now, but that is quite a lot of time. Again, I am not saying that it is not possible to do remote interpreting and that these technologies have not come a long way, but I think that is also a small example that it takes more than technological advances. It also takes a lot of thinking about how to implement these things in detail, how to design the systems that are really capable of doing this in real life. All I am saying here is that when we think about how we design AI systems that provide this language support one day, we should not make similar mistakes, we should learn from these things, we should certainly involve a lot of people: we should involve interpreters, we should involve linguists.
By the way, in my centre, on Friday, one of our research fellows started a Leverhulme Trust Early Career Fellowship looking at machine and machine-assisted interpreting in legal settings, looking at what is there. It is not to say that we are not doing this. I think it is just to say there are many, many details to clarify and from looking at how technology is implemented in the public sector and elsewhere, I am just a little concerned sometimes that this detail gets lost.
Professor Richard Susskind: I share all these concerns. I think your case study is a brilliant one. But in thinking about video courts, what we have done there basically is a court with a videolink. What we should have done is what we call user-centred design, saying, “Who are the participants here?” You would sit down, I would sit down, and then what would end up with is a formal hearing that I not simply technology grafted on to the old way of working and that, I am afraid, is the history of court technology over the last 65 years.
The Chair: As with lots of other things as well.
Professor Richard Susskind: We just pile layers. It is mess for less. We put more technology on messy processes. We completely agree on that. Would it not be wonderful—and this is where we need the long-term vision and the guts to think for the long term—if we took a step back and thought, “What would a just, remote process look and feel like that takes account of all the key participants and does not simply replicate the role of judges and lawyers?” I think we are not that far apart, and I think it is good for the committee to hear this kind of debate.
Q90 The Chair: Richard, I have a very mundane question before I go on to Patrick for the next question. You have international experience through your research. One of the things we have heard about, which we know, is that the state of the courts in terms of technology and sound and the rest of it gives us an extra hurdle to overcome as we take this forward. Do you think our country is particularly bad in terms of the fabric of the court buildings compared with other places? On the whole, AI will get developed, stands a chance of being developed, across all countries at the same time because that is the nature of it. Compared with other countries—and that is the only bit I am looking for really—how much do you think the physical state of our court buildings is going to hold us back or make it difficult?
Professor Richard Susskind: The state of the court buildings or our systems?
The Chair: The court buildings.
Professor Richard Susskind: Okay.
The Chair: It has been raised with us so many times. I do not know whether they are worse than everyone else’s or everywhere has courts like ours.
Professor Richard Susskind: I can only be anecdotal here, but I do travel the world and I hear similar complaints. I hear the same narrative in almost every justice system I visit. There is an exception: small jurisdictions that are sufficiently nimble and often jurisdictions that are seeking to be competitive in the international stage have decided—you can think of Singapore, you can think of Dubai, you can think of other countries, where they have decided, particularly for the commercial courts—to invest in state-of-the-art buildings with technology built in.
The Chair: On a smaller scale.
Professor Richard Susskind: Yes, and it is easier. I have advised smaller jurisdictions. You can get everyone around the table and you can bang heads together. We are a massive jurisdiction and it is very hard to manage the resources.
The Chair: Okay. Thank you. I am conscious of the time now.
Q91 Lord Carter of Coles: I suppose this is about confidence. If this is to happen, how do we get quality assurance and how do we understand in a dynamic way? Richard, you could help us here because you are advocating that we do not know what is going to happen. How do we deal with a road map that is indistinct? How do we all the time move along? I think you touched on it just now, but could you expand on that?
Professor Richard Susskind: This is under the heading of quality assurance?
Lord Carter of Coles: Yes.
Professor Richard Susskind: It is very interesting that in medicine, when a new medical device or a new drug appears in the market, we are all pretty confident that it has gone through some kind of process: randomised control trials, peer review of publications and so forth. Over the years, when we have looked in this country, and frankly across the world, at new uses of technology, they have not been subjected to this kind of scrutiny at all. It is often, with the greatest respect, a senior judge, a senior politician saying, “This sounds like a good idea; let’s have a go at it”. We have to move beyond that. When we come to AI, one of the great difficulties, and I am not sure we will ever get over this difficulty, is that we do not really understand what these systems are doing when they are doing what they are doing. We design them for any particular interaction. They are black boxes. Lots of work is going on in this. But I think what we need is a whole new set of assured standard methods, tools, processes for developing these systems and, more importantly, a whole set of tools for validating, verifying and testing both systems under development and systems in operation.
Wearing another hat, I am working across the Commonwealth to try to crack the access to justice problem, and one of the big gaps there is that we do not have the tools to evaluate the systems we are developing. We are going to have to ask some difficult questions. When is a court service good enough? We often say, “Well, that would be better than nothing” but better than nothing is not necessarily good enough. So how are we going to set these standards? How are we going to offer assurance and comfort to people who are using, perhaps, online courts with autonomous interpretation, that this is a reliable system? We are going to have to develop tools.
Lord Carter of Coles: Is that starting anywhere? Where is the thinking on that?
Professor Richard Susskind: There is very little thinking work. There is an interesting bit of work being done in Harvard but I would say we are decades behind the medical profession and we have to move beyond this mentality that people have a good idea and think it is worth a shot. We have to get into an R&D mentality; we have to be piloting, experimenting with feedback loops and iterative development. If you look at most of our big reform programmes, someone has specified a system that is going to be developed in seven years’ time. There is no real movement for testing and experimenting and revising as we have gone along. This is why I think innovation has to be built into our public institutions. This discussion that we are having today, in my view, is not about interpretation and translation. It is about the new era into which we are operating and we have to have new ways of thinking about how we change—how we fundamentally change. I am full of non-answers today, but I am quite good at pinch-pointing great questions.
Lord Carter of Coles: We are pressed for time, I am afraid. Sabine, is there anything you would like to add to that?
Professor Sabine Braun: Briefly, yes. I agree. By the way, we also advocate this so that in healthcare, if you use machine translation, it should be evaluated like a medical device, and we do not exactly have this in court.
Maybe I would like to say something that it should not be. We should not just trust automatic evaluation tools in translation and interpreting. They will not capture what we need in court interpreting. I think we need evaluation criteria and systems that are adjusted to the requirements of court interpreting.
To give a brief example, fluency is nice to have, and many machine translation interpreting systems come across as very fluent but of course in a court, accuracy is much more important. That is just a very simplistic example to say that we will definitely need human evaluation in these systems for a very long time.
Daniela Ford: The tools have to be assistive. I do not think they can be autonomous ever.
I will give you one example from my student from Singapore. We have a lot of dissertations where students compare human outputs against machine translation output and machine translation against ChatGPT—which one is better? My student did some research on a machine translation system built specifically for Singapore English, which is different from English here because they have their own terminology as well for cultural reasons, and she compared the output with ChatGPT. The ChatGPT output sounded a lot better, more fluent, but it got all the cultural references wrong. This is the crossroads where Singapore is at now because they do not have the data. Again, it all comes down to the amount of data that you have and the quality of data and we see this with students when we build machine translation systems. You are looking at 900,000 words minimum. You need to find a lot more, so you are looking for data. There is no way you can quality-check all that data.
I also want to mention that there are already court cases going on against various big tool developers because they are basically harvesting the internet for data to get as much as they can. There are quite a few court cases ongoing already. Spoken language is another matter again because there is certainly not enough data for spoken languages available at this point. So it is a combination of the language combination for starters, it is a combination of specific dialects, a combination of the subject areas, whether it is spoken language or written language, but the person at this point always needs to be there. There are some automated ways of checking quality but really what they are doing, they are comparing against the human translation at this point. There are various mathematical systems where this is done and for some of them you can say “Okay”, but for others you can say, “Well, that is a mathematical calculation. That is not meaningful from a language point of view”. So I maintain it has to be an assistive tool.
Obviously, productivity is a lot faster. You put a text into the system and you have a translation straight away. It does not mean it is a good translation at all. It can be full of mistakes and that is the danger.
Q92 The Chair: That is all very interesting. I think we have been on a bit of a journey in this committee because our early witnesses were not even thinking about it and as we have gone along, we have got more confident that, as you said, Richard, it is there and you can hide your head in the sand if you want to but it is a wasted opportunity. Today you have given us some really good ideas.
I will ask the last question. Can you give us one recommendation because however different we want to be, at the end of the day, we are going to produce a written report with recommendations to the Government, as is always the case, and if you could just give us one idea as to which recommendation you would most like us to put in our report that would be really helpful. Who wants to start? Who has their idea already? Let us start in the middle. If you are in the middle, you never get to start.
Professor Sabine Braun: I have three suggestions. Look again at the design of the video-interpreting solutions and when we redesign and revisit it, involve the interpreters in redesigning it because I think that in the medium term that is the system that will probably take quite a lot of attention.
Next is training for everyone, not just for interpreters, in technology literacy, including AI literacy—training together, ideally. Then there is relationship building with the interpreter community and also seeing technology as something that is embedded in something broader; technology does not exist in a vacuum. We also need to look at the working conditions for the interpreters. If these working conditions are favourable, interpreters will also be more inclined to work with technology, to accept certain things and to collaborate, but that is also something that I could talk at length about. I think that is not there.
Specifically—I know we are short of time—when it comes to AI, the guiding principles here have to be fairness of justice, equality, safety, and ethically compatible, socially compatible ways of introducing it. Again, it comes back to what do we do with skilled workforces. Professions will change, I know, but we also need to make sure that this does not happen at the cost of the experts who are providing services at the moment.
The Chair: Just briefly, over to Richard. Do you want to go next?
Professor Richard Susskind: I think we need to be bold as a country. I would like you to produce a road map that starts with AI automating, enhancing and supporting the great work that interpreters and translators do but anticipating that by the mid-2030s this is one of many tasks which will be autonomously undertaken by machines.
Daniela Ford: I would stress the confidentiality aspect very strongly. That would be my first point, absolutely, because I have seen so many bad things already with that.
My second point is to be very clear that tools are there to assist, to help—absolutely important.
My third point is to let translators or interpreters decide or suggest what would help them rather than enforce something on to them that is not going to help them in their work.
The Chair: Do you just want to come in briefly, Lord Blencathra?
Q93 Lord Blencathra: Very briefly. I have no interest to declare on this subject. This is just a brief question for Professor Susskind.
I think that in the last couple of years, all the great gurus in America, and maybe Tim Berners-Lee in this country, who were in favour of AI and said it was the future, and maybe Bezos or the late Steve Jobs and others working for Apple and Google and so on, have now said, in the last couple of years, “This is a great threat to humanity. Oh my God, what have we done? We must have controls on AI”. Do you share their view?
Professor Richard Susskind: I think we have to get used to the idea, and humans are very bad at this, of holding two thoughts in our head at the same time. One is that this is a technology of phenomenal capability and the other that this is a technology that could bring about our demise. That is why I say again that saving humanity with and from AI is the defining challenge of our generation. I am not yet at the stage of saying that we need to pull the plug but the bigger question, in my view, is whether or not we could pull the plug even if we thought it was a bad idea.
I remain an optimist and for this country I would want us to lead the way in the safe deployment of this technology.
The Chair: One sentence, Sabine?
Professor Sabine Braun: One sentence. Certified and professional interpreters take responsibility for what they do. That is in court. They take the responsibility for the accuracy of their interpretation. So I think a good test for whether this is ready or not will one day be to ask the providers of the AI technology whether they will be ready to take the responsibility for the accuracy of what they are doing.
The Chair: A very good point. Thank you, everyone. I will bring the session to a close now but this has been an excellent session. We could have gone on for three hours, to tell you the truth. We have not asked as many questions but we have heard a lot from you and we decided to let you speak. You will certainly find that a lot of what you have said reflected in our report.
If there is anything you feel you missed out on or that, on reflection, you would like to add, please do approach our team. It can still count as evidence. Meanwhile, thank you very, very much indeed.