Critical AI literacy for interpreting in the age of AI

1. Introduction

Artificial intelligence (AI) is rapidly weaving itself into the fabric of our lives. Some authors (Kissinger et al., 2021; Susskind & Susskind, 2022) speak of a forthcoming second wave of AI in different professions, whereas others (Corpas Pastor, 2024; Fantinuoli, 2018) draw our attention to a third technological revolution, both of which have significant implications for interpreting. Critical AI Literacy (CAIL) seems crucial if we are to make well-informed decisions about AI-enhanced interpreting. On 2 February 2024, I had the pleasure of interviewing Dr Deborah Giustini on this key topic.

2. Interview

Given the latest developments in the use of AI in interpreting, what does CAIL mean for interpreting?

Whereas in previous years the high computing powers necessary for the functioning of AI meant its development was limited to large IT companies, the increasing availability of cloud computing and neural language processing capabilities now makes it possible to scale things up and achieve more autonomous development of new applications, which have become very efficient but also pervasive. Coexistence with these new technologies implies the need to develop digital skills to allow, on the one hand, the development of solid basic skills for access to professions, including interpreting, but also to be able to understand the social, ethical, and economic implications of AI. It appears clear that CAIL—the skills needed to work with and through AI technologies, and the active awareness of its affordances and limitations—is increasingly crucial if you want to avoid the digital divide between members of society. The demand for CAIL has now become very stringent and functional for carrying out roles in various domains of work, and of course, interpreting is no exception.

From a scholarly perspective, I conceive of CAIL as interactional and embodied expertise. Conceptually and analytically speaking, in the last few years, my work has been focused on attending to interpreting and interpreting expertise through a specific ontology called “practice theory,” which looks at social reality as open-ended, organised sets of materially mediated action learned from and with others (Nicolini, 2012; Nicolini & Giustini, 2024). For practice theorists like myself, competence, skills, and knowledge are not only cognitive skills but also components of situated action, that is, what we need to competently perform in everyday life in all its nuances. This means that I see CAIL as interpreters’ embodied capacity to engage with AI technologies in the social and material activities related to language interpreting, as a form of shared, hybrid knowledge. This interrelation is captured by the term “sociomateriality,” which refers to the constitutive entanglement, rather than dualistic opposition, of the social and the material. Although it has been popularised by organisation studies (Orlikowski & Scott, 2008), sociomateriality was first conceptualised by the posthuman feminist onto-epistemologies of Donna Haraway (1991, 1992, 1997) and Karen Barad (2003).

I think looking at CAIL as the interrelation of the language (social) and the material world, proposing that they are mutually constitutive and dynamically defining each other in the process, allows a more profound discussion about both humanity and technology as mutually constitutive and defining themselves dynamically by interacting with each other. I’m not a supporter of techno-utopianism (the promise that advances in science and technology mean inevitable progress), but I am also not an advocate for techno-dystopia (the fear that the individual is likely to lose control or become dependent on machines). I see both tendencies in interpreting, particularly when it comes to defining the skills, the expertise, and the role of interpreters vis-à-vis AI. The debate boils down to those who posit AI, such as computer-assisted interpreting (CAI) and machine interpreting (MI), as the future without questioning it from socio-economic and ethical perspectives, and those who proclaim that human interpreters are irreplaceable by machines by virtue of what apparently distinguishes them from machines: emotions, and nuanced representations of reality (as in culture). But this is nothing new. I am going to be a bit provocative here. I have a background as a social scientist, and in the social sciences, we have already witnessed these debates which started in the 1960s, particularly through works such as Alchemy and Artificial Intelligence (Dreyfus, 1965) and What Computers Can’t Do (Dreyfus, 1972) by Hubert Dreyfus, who argued that human expertise depends on our background sense of the context, of what is relevant given the situation, rather than on the process of searching through combinations of possibilities to find what we need. Dreyfus did not believe that AI could capture this tacit knowledge or do the kind of rapid problem solving that humans do. I tend to agree with this view because, as a practice theorist, I do not see interpreters’ knowledge (and CAIL) in terms of paradigmatic rationality and bounded cognitivism. I see it rather as a negotiation of the meanings of words, actions, situations, and material artefacts, in a world—including communication and thus, performances of interpreting—which is socially and culturally structured.

However, I believe that this dialectical paradigm is inadequate: We cannot and should not continue this “us and them,” “human vs. machine” line of thinking. Rather than siding with one or the other pole of the debate, it would be more useful to recognise interpreting expertise right now as an issue of sociomateriality. This does not mean reinventing the wheel. Interpreting literacy has always been sociomaterial. Think about the interrelation between simultaneous interpreting and the equipment needed to perform it, like the booth, the headphones, the microphones, and so on, or even consecutive interpreting with pens and note pads. That is a basic form of sociomateriality. This is valid for CAIL, too, since AI is part of the equation in interpreting workflows and knowledge management. Here comes my major disagreement with contemporary interpreting studies: Most put the posthuman in the future and are reluctant to acknowledge the fact that AI is already here. Instead, we should be able to talk about the present of AI and CAIL. The usual thesis is that interpreters are better than machines because they possess unmatched skills—usually in terms of responding to and producing linguistic and cultural nuance as well as gauging emotional appropriateness—that make them more competitive, more accurate, and even more refined than the machines. But the question is: Who decides what machines can or should do? Who decides what the appropriate output of machines is? Who decides how we are to be enhanced to maintain our primacy as the super-intelligent human? I fear that we are really going around in circles by attaching this Cartesian, dualistic notion of the mind/reason versus emotion divide, which is epistemologically the humanistic and sovereign image of the subject as a liberal individual agent. What I am observing is exactly the opposite. We should think of CAIL as the ability to foster transversal connections in practice and to rethink the interpreter’s role as an assemblage, a complex human–machine multiplicity (Braidotti, 2013). Going beyond the “human vs. machine” and therefore beyond the “reason vs. emotions” divide is a way of starting to recast interpreters’ expertise in terms of CAIL, as a different epistemological order.

In a now iconic piece of neo-noir, cyberpunk Japanese animation, Ghost in the Shell (1995) by Masamune Shirow (1991), we see a futuristic world where humans are mangled compositions of artificial, techno-engineered body parts (the “shell”) and whatever remains of their human components (the “ghost”). In one scene, we see the foreign ministry’s interpreter who undergoes a brain hacking procedure to prevent industrial espionage and intelligence manipulation. While I am not venturing as far as saying that we are likely to see neural implants or partly cybernetic bodies in the future, I think we are likely to metaphorically conceptualise CAIL as the sociomaterial integration of AI algorithms to enhance interpreters’ linguistic prowess, to bridge their embodied socio-cognitive abilities (the “ghost”) and the vast neural language processes and databases of AI (the “shell”). I also think that CAIL is not about thinking “we need to upskill because one day the interpreter will be post-human”; it is something that is already in the process of happening. The post-human is about displacing the centrality of the interpreter’s anthropomorphic brain as the producer of knowledge, and convergence with automated, neural technologies. There is a very famous and powerful metaphor by Donna Haraway (1985), that of the “cyborg,” a concept theorised in the now seminal 1985 “Manifesto for cyborgs: Science, technology, and socialist feminism,” which is very apt for thinking about CAIL. A cyborg is a hybrid entity that blurs the boundaries between human and machine, challenging traditional notions of identity (Penley et al., 1991). We don’t necessarily need to inconvenience sci-fi imaginaries to grasp it: Rather, it is the medicalised body, the tracked biometric data, or the use of our smartphones which is now necessary in almost everything we do. This leads us to conceptualise CAIL as the reciprocal ways that human interpreters and AI can augment their capabilities through synergistic human–AI interactions (i.e., human-augmented AI and augmented human intelligence), resulting in hybrid intelligence. Probably, what interpreters are called to do and be is more-than-humanness and more-than-machinery.

In the way you define CAIL when it comes to interpreting, I understand that you are emphasising that discussions must go beyond this static, techno-determinist perspective on interpreting as clearly seen in interpreters versus machines debates. Am I correct?

Yes, of course. If we do not start to go beyond this dichotomy, we will also keep running around in circles without finding any epistemological standpoint through which we can better attend to AI-enhanced interpreting when it comes to other adjacent problems. CAIL is a complex interplay of elements with multiple dimensions. As a matter of fact, another dimension I find especially worrying relates to AI, skills, and expertise and is likely to be exacerbated by our limited critical-analytical tools, for instance, what is going to happen to the interpreting workforce and interpreters’ competencies? Now that AI is deployed both to reduce costs and control labour, it is likely to result in the deskilling of workers. That is one side of the discussion. Another common argument revolves around the polarisation of skills, according to which interpreters would be divided into jobs either at the bottom, requiring lower skills levels (where the machines will dominate), or at the top, requiring a greater level of skill (a sort of niche, “boutique” interpreting). Basically, the latter would be able to engage in upskilling, that is, the process of enhancing skills in certain areas, and this very much relates to CAIL because interpreters will have to update their skills. Therefore, these arguments suggest that the current landscape of interpreting work is shaped by dualistic ideas of automation, with machines matching or outperforming humans in a fast-growing range of tasks. They also indicate that the activities most susceptible to automation are routine cognitive tasks like interpreting in so-called technical settings or in predictable environments such as webinars where communication is mostly mono-directional. According to this circulating discourse, what is left for interpreters in terms of skills and labour would be tied to the intellectual and creative capacity that they can still hold on to, since it is hard, if not impossible, to reproduce them through automation. However, as I have argued, I don’t really think that this divide between reason and emotion is tenable because everything is permeated by affect and creativity, even when it comes to AI use and CAIL. So, where do we draw the line? Can we say there is no emotion in a technical setting or a webinar? Can we say it is just a mono-directional type of communication?

My preoccupation with these dualistic arguments is that they do not address the state of the industry when it comes to CAIL. We’re moving in a direction that implies that intellectual work is not necessarily factored into the translation pipeline. Techno-optimists emphasise that adaptive interpreters will still differentiate themselves from automation and less qualified individuals, hence dictating labour price and autonomy over their work. Regardless, while the demand for language work worldwide is growing as one of the main inputs of the knowledge economy, much of this work is already stuck at the lower end of the market, where automation and cheap labour are used as major selling points. Customers and businesses are likely to prioritise price and availability over professionalism and quality, and these are dynamics that interpreting studies in particular, and studies of industrial relations in general, have attended to for decades. Interpreting scholarship has denounced these dynamics for decades, for example with regard to the opposition between professional/trained interpreters versus the non-professionals “invading” and “ruining” the market. So, my problem is trying to think around these dynamics. How can we be so sure that constantly having recourse to automated solutions rather than human intervention will not obscure the skills of professionals? If clients perceive machines to be able—at least half-decently—to do the work of a human without all the hassle of going through an agency, contracting interpreters, and flying them to a venue, then, interpreters will start to find themselves in a position that will require them to justify their intellectual work, their skills, and their cost in relation to this expertise. So, we should not forget that perceptions of expertise, of quality, and of professionalism are formed also by providers and consumers. Such perceptions are not likely to necessarily overlap with those of the interpreters. My fear is that, as in many other sectors, the organisation of interpreting work will not become necessarily more valued because of automation, but rather AI is likely to exacerbate exploitative practices and the low signalling of language professionals’ labour value. These dynamics point to a troublesome relationship that fuels mechanisms of devaluation, rather than pushing for positive CAIL development, where the language worker is likely to become yet another invisible cog in a larger automated process that directs the purchasing and the selling of languages as a commodity.

Interpreting is a specific type of social interaction with a diversity of stakeholders who have different levels of involvement with and understanding of technologised interpreting. Who exactly stands to benefit from applying CAIL to interpreting?

I think this is a very interesting question. CAIL really has the potential to benefit various stakeholders; however, the specific advantages vary, based not only on the level of their involvement, but also according to their understanding of technologised interpreting and of course according to their own legitimate interests. I would say if we start from interpreters, CAIL can empower them to make more informed decisions about when and how to leverage AI tools in their workflow. For instance, they are likely to be able to critically assess the outputs of AI systems, thereby ensuring that automated systems are aligned with professional standards and do not compromise aspects such as accuracy or cultural dimensions. I think CAIL is also essential for language service providers and agencies of any type to ethically deploy interpreting services. It would really enable them to critically evaluate the impact of AI on consumed service quality. Service quality is a cornerstone of what such entities do; they are likely to want to communicate transparently with clients about the role of AI in the provision of such services. When it comes to businesses and organisations, and, hence, to customers, they are likely to want to understand how to strategically implement AI in interpreting. In addition, we should consider the ethical implications and the societal impact of AI and address potential concerns relating to data privacy. This should also be a primary concern for tech developers to engage in responsible innovations. So, CAIL is likely to benefit them across these dimensions. Even for governments or public services, CAIL is likely to drive decision makers to understand the potential bias of AI and take measures to provide fair and accessible services for any linguistically diverse population. Finally, end-users would benefit from knowledge on how to critically engage with AI-driven interpreting services. They are likely to need these skills to question the outputs, to understand the limitations so as not to have an impression of AI as being capable of doing anything, but also to be able to more actively and responsibly participate in shaping the responsible use of AI in their interactions. I think that the most promising approach would be to involve stakeholders, including first and foremost the interpreting community, to foster this systemic awareness because this is not something which is concretely occurring across the spectrum. Of course, the interpreting community does need to play an active role in leading the development of CAIL because we are going through a transformation phase which is a societal-technical change, and we are not going back. Interpreters are the core practitioners, the core link in this system of stakeholders and, if we want to benefit from the opportunities of AI while trying to prevent to the extent possible the risk of being “dominated,” the risk of being “replaced,” I think we need to rethink our skills and rethink our professional practice on the basis of evidence and ground our awareness in an understanding of where the market is going, avoiding attempts to shield ourselves from engaging with these difficult, even emotional discussions about the state of the market.

In your work on the emerging trend of platformisation of interpreting (Giustini, 2022, 2024), you are raising awareness of digital platforms, which can be considered as another important stakeholder of interpreting services. Do you think it is likely for CAIL to have important implications for this emerging uberisation of interpreting?

Yes, absolutely. When it comes to digital platforms, they mostly function as marketplaces where labour supply and demand are matched automatically through algorithms, that is, through AI. So, there must be from their side an opportunity to engage with understanding how algorithmic management is likely to have an impact on matching customers and interpreters, but also on more problematic dimensions when it comes, for instance, to the use of algorithmic management in deactivating some profiles of interpreters or to some profiles being made more visible than others on the platform. These are all dynamics which have an impact not only on the role but also on the employability and work continuity of interpreters who operate through these platforms. We should be willing to engage with this debate.

The mission of the recently launched Interpreting SAFE-AI Task Force is “responsible AI.” What does “responsible AI” mean for interpreting? Who exactly stands to benefit from it? What are the key factors to consider in order to promote a shift from uncertainty to accountability, responsibility, and transparency around AI in interpreting? Does a joint effort by different stakeholders contribute to responsible AI for interpreting?

This is an excellent set of questions. Funnily enough, I am a stakeholder assembly member of the SAFE-AI (Stakeholders Advocating for Fair and Ethical AI in Interpreting) Task Force, and of course, in this interview I am speaking in my own capacity only, not on behalf of the task force. These matters are really at the heart of what the interpreting community now finds pressing to answer. As for what “responsible AI” means, as I said earlier, I am not against technology, but I belong to a crowd which is sceptical of what Haraway calls the “techno-fix,” that is, all those pushes and projects towards the adoption of AI which disregard engagement with communities of use and of practice, and an unwillingness to pay attention to AI in interpreting and how this represents a very entrenched version of techno-capitalism in return.

First, to me, being responsible means emphasising ethical principles and the well-being of individuals involved in the use of AI for interpreting. So, responsibility means understanding what it means to access to interpreting as supported by AI. I often see very sweeping statements made about how AI is democratic because it increases access to information, and it makes it easier for people to communicate. But I find this kind of hype a bit foolish because AI has a very unequal distributive impact, and if we look at the available evidence, it is mostly supranational institutions which have the purchasing power for AI-enhanced interpreting solutions. And also, if we look at the available MI matrixes (e.g., The Nimdzi Machine Interpreting Matrix, see Nimdzi, 2023), we see developers and businesses being targeted in terms of potential investments and adoption of AI. So, both growth and financial viability are constantly being assessed, but not responsibility in the sector. Then, there is the problem of dominant languages when it comes to thinking responsibly about AI. So-called “minoritarian” and indigenous languages hardly constitute a threshold percentage of the constituency of AI. Developers tend to prioritise the application of large language models (LLMs) for languages that exhibit a more consistent performance, which further contributes to the under-representation of certain languages and the social groups affiliated with these languages in the digital sphere and beyond. We also know that LLMs perform rather badly in non-standard languages, yet they are likely to be playing a growing role in life-altering decision-making settings, such as justice, asylum, and healthcare. I have been recently looking at the use of machine translation (MT) by the UK Home Office in its implementation of a streamlined asylum process and how this is not likely to be the most effective solution in this type of setting because it perpetuates cultural, gender, and racial bias. So, we are already seeing that AI is used irresponsibly in some settings. Think also about using MI in hospitals, which is likely to be a tempting, cost-effective solution. Interpreters are likely to be employed for consultations with doctors, but then you are likely to have patients who need to communicate because maybe it is night time, or maybe they need something from the nurse because they are in pain, and there are no contracted interpreters on duty. So, AI is likely to represent an ideal solution in these settings, right? But what about patients who speak less common languages, who we see very often in multilingual societies? Will they be at a disadvantage? Should we think that some level of communication, perhaps inaccurate or incomplete, is better than nothing?

There is also the issue of bias because LLMs, on which AI training rely, tend to perpetuate racial and gender bias, and this impacts accountability, because it raises the question as to how we can devise the necessary corrective measures if machine-made mistakes are likely to remain blurred or opaque. So, I think we need to recognise that AI-enhanced interpreting is likely to have the potential to break down language barriers, but there is a risk of exploitation by privileged groups. If we look, for instance, at the Global South, tech firms, which are overwhelmingly based in the Global North, are leveraging economic disparities to create products that further entrench Western hegemonic dominance in AI, and thus digital colonialism. In addition, unequal digital access and varying hardware requirements, for instance in the case of vulnerable communities like migrants and refugees, mean that the benefits of AI interpreting are not easily accessible to everyone. Hence, access for whom?

Second, there is the issue of confidentiality and privacy. Once again, if I think about asylum settings where the use of MT is increasingly considered viable, what of the sensitive data that public services are likely to be feeding to the machines about asylum seekers? What if such information is breached and weaponised against asylum seekers by the very governments and countries that they are likely to be seeking protection from? So, this involves not only handling sensitive information with care, but also having stringent data protection regulation. It necessitates the implementation of robust encryption methods, secure data storage practices, and strict access controls to safeguard the confidentiality of the parties engaged in the interpreting.

Furthermore—perhaps this is a less-explored dimension—responsible AI also means being attentive to how AI is affecting the environment. Research in and beyond interpreting studies has shown that digital solutions minimise carbon footprints because they eliminate the need for interpreters to travel to physical locations or access infrastructure such as hotels, but they still have an environmental impact, and a larger one than we might imagine. For instance, remote interpreting links a variety of stakeholders who rely on digital technologies and this is coupled with producing and disposing of electronic devices, data centres, and the energy required to power them all, which contributes to the ecological cost of the practice. So, if we look at the information and communication technologies (ICTs) sector, this is actually responsible for 10% of the world’s total energy consumption and more than 2% of total carbon dioxide (CO₂) emissions against aviation’s 2.5% of global carbon dioxide emissions and 3.5% of global warming (International Energy Agency, 2023a, 2023b). While I am not suggesting here that the interpreting profession should engage in non-sustainable practices or just adhere to “travel everywhere” policies, we should recognise that AI is equally implicated in these processes; otherwise, we are leaning towards greenwashing.

As for who exactly stands to benefit from responsible AI, the beneficiaries include, of course, interpreters, which means that if we deploy AI as a tool to augment rather than replace interpreters or lower their working conditions, then we can take advantage of this potential for increased efficiency and reduced cognitive load associated with interpreting itself. Those seeking interpreting services stand to benefit though potentially cost-effective, accessible, reliable, and culturally sensitive translations in which biases and mechanisms for checking are clearly outlined. We should also not forget that companies involved in developing responsible AI can benefit from improved market acceptance and trust: Positive public perception of responsible AI practices can be an advantage, although I am very mindful of the cost and effort required for maintaining such systems when it comes to organisations. This is also likely to allow the entry of smaller players in the market while preventing an exacerbation of the concentration of AI power in large tech corporations. Finally, responsible AI should also be a benchmark for the whole society, for instance, as related to tackling biases and ethical considerations, and fostering better inclusivity while minimising unintended consequences.

Concerning the factors that we need to consider in the shift from the current situation to more responsible AI for interpreting, I side with a renowned posthumanist thinker, Rosi Braidotti (2006), in conceiving of AI-enhanced interpreting as a “site of ethical and political accountability” (p. 138). Accountability is the cornerstone of the governance of AI. Accountability should help prevent potential misuse that could compromise privacy or lead to biased interpretations. It is also essential for setting standards for performance and accuracy, making it clear, for instance, who is responsible for AI output. Nevertheless, accountability is very much defined too imprecisely because the socio-technical structure of AI systems implies a variety of values and practices, which are complex to measure. Consequently, we have a lack of clarity here in terms of answerability: Which authority should we recognise and interrogate when it comes to AI use and misuse? How can we make sure that both industry and companies act responsibly, and check the reasonable limitations of power? How can we make sure this accountability architecture achieves the goals of compliance, reporting, and enforcement?

Accordingly, transparency is often heralded as a key to responsible AI, but transparency alone is insufficient to ensure that actors developing and deploying AI systems adhere to certain values. In my research into interpreting and the gig economy, I found that the concept of transparency was constantly used by the platforms to lower the exchange barriers between customers and service providers. For example, they would say that platforms are more transparent than agencies and other forms of client-professional intermediation. Using this concept often served to lure interpreters away from the so-called “ties” of a “constraining” and more traditional employment relationship, such as those between interpreters as freelancers and agencies. Relatedly, I have seen that a variety of digital ecologies using AI resort to narratives of transparency as a tool to gain competitive advantage by constructing ostensibly more virtuous identities about themselves, in contrast to other existing business models. So, if we really want to promote transparent AI, stakeholders and policymakers should move beyond this technical hype and bolster measures such as funding for research and increasing resources for the institutions of accountability, including monitoring tools.

As for the last part of the question, I advocate for the coming together of different stakeholders to regulate AI responsibly. Interpreting is, inherently, a collective activity. We often think of interpreting as just being interpreting. But interpreting is a relational market, where multiple occupations and actors make up and take shares in the outcomes of multilingual communication—organisations, institutions, public authorities, professional associations, industry players, clients, and communities—with the ties among them increasingly dependent on specific material arrangements (technologies and AI). Without these ties, the wider industry simply would not function, and the positive outcomes of inter-linguistic knowledge transfer would be halted. As such, contributing to a responsible, ethical stance must be the result of a shared effort, because the adoption of AI in the sector will impact not only interpreters and end-users, but also all the other practices and actors that make interpreting possible and that base their own professional subsistence on the tasks, products, and outcomes available in the industry. If we want to see professional subsistence for everyone, we really need to be at the forefront of the formation of a collective responsible stance in the industry. To conclude this point, if the industry wants to keep together its different interests, it needs to orchestrate responsible AI with intentionality. In fact, the issue here is that responsible AI struggles against the huge amounts of techno-capital that support the industry itself: natural language processing (NLP), engineering projects, scaling up of software services, the push for digital platforms. Stakeholders will eventually come into conflict around responsible AI. We are already seeing that happening because there are different market interests. I am afraid that capital and market goals will push most stakeholders to participate in a culture in which AI is viewed as offering salvation for the costs, messiness, and unpredictability to which human multilingual communication is exposed. I am also afraid that responsibility will be equalled to financial responsibility. We still aren’t doing the whole industry and supply-chain analysis of AI, which is needed to understand whether responsible AI use is possible. So, I think we have a long way to go in terms of collaborative efforts and identifying who wants what and who has the right to what in these matters, and how these wants will affect other actors.

When it comes to optimising the training of future generation of interpreters who can thrive alongside AI and analysing the design and implementation of modern curricula, pedagogical models and learning environments for interpreters, what aspects of CAIL in your opinion should attract the attention of programme directors, curriculum developers, and trainers?

This is a terrific question. I think it is a complex matter to consider because when it comes to training, there are different schools of thought among countries, despite the more predominant Eurocentric tradition of interpreting pedagogy. As a trainer myself, I am seeing that several training institutions, both at graduate and vocational levels, are integrating interpreting technologies (including now AI) and related competence development in their curricula while others show a more cautious approach. This said, we can make a few considerations.

At the pedagogical level, CAIL should lead to the ability of the next generation of interpreters to understand, analyse, and engage with AI in a thoughtful manner. This involves more than just technical proficiency with AI tools. It encompasses having the capability to deeply understand the ethical and social implications of using AI as well as understand its potential impact over future employability, which is a conversation we do not necessarily have with trainees. It also involves making them understand where the market is going and how AI is likely to affect their work someday. With regard to practical skills, the benchmark for programme developers and trainers would be to equip trainees with the ability to use CAI tools (where machines are assisting humans) as well as to integrate their performance with MI (in which, inversely, we see humans assist machines). However, as surprising this is likely to sound, this is only part of the story, because technology is constantly evolving—and I am seeing it in my own classes—students are likely to be trained in one or more applications that will become obsolete in a couple of years. They will be entering the market and will constantly be facing the emergence of new software, platforms, and gadgets. What I am trying to advocate for when it comes to pedagogy is that raising awareness that CAIL should not be narrowed down to practical skills of technology used only, because that can be refined along the way. Instead, we need to educate trainees in a way that makes them understand that speech-to-speech and speech-to-text translation are there, how they work, what happens through them. Therefore, I think that we should instil into curricula the understanding that both the interpreting industry and practice are now led by technology, and that this has consequences for the labour process, the role of the interpreter, the market development, and stakeholders’ expectations. We will go nowhere through a purely functional approach.

Developing comprehensive curricula also means to explicitly address the new ethical considerations associated with AI. Despite its often very programmatic and rigid treatment of professional norms—such as neutrality and objectivity—training does play a role in shaping the social and collective milieu of the profession. Training should branch out to account for new topics and materials, such as bias in machine learning algorithms, privacy concerns, data protection, transparency, and accountability. This must go hand in hand with engaging trainees’ critical thinking in CAIL by prompting discussions on how renewed ethical principles around fast-developing technology are likely to guide decision making in situations where AI is involved. Fostering collaborations between interpreting programmes and other programmes like ethics of AI and with other experts in AI, such as engineers, has the potential to benefit interpreting education by generating more rounded insights.

Indeed, curricula often do not leverage the strength of inter- and trans-disciplinarity when it comes to training methodologies. Of course, I cannot speak for every institution; I am a researcher and an instructor with a fairly diversified experience in Europe, East Asia, and the Middle East, and there is still a conventional approach when it involves interpreting skills and CAIL, meaning that the latter is treated largely as a supplementary activity. We do have curricula that integrate fundamental disciplinary insights like intercultural communication, elements of international relations of business, and so forth, but I would like to see a more interconnected approach—a truly multidisciplinary approach—now that AI is entering into pedagogy. In this sense, I side with those colleagues and scholars who would like to see at least basic courses in programming, engineering, maybe science and technology because that would help address the gap in terms of understanding the ethical issues at play. In my view, you cannot really fully understand AI and what it means for the field and for your own experience as a professional if you are not willing to grapple with how it works. I oppose the argument that being critically AI literate means that you are only actively learning about the technologies involved, and that you do not need to understand the advanced mechanics of AI. You need to get to the coalface.

I would also like to see an ethical politics of response-ability in training—a concept that Haraway (2016) uses to emphasise the performative feature of responsibility that entails “cultivating collective knowing and doing” (p. 34). When it comes to generating awareness about pedagogy and AI, using traditional critical pedagogies is not likely to suffice in the sense that they do not necessarily propose a type of learning which accounts for both cognitive skills, practical skills, power, and ethical relations. So, I would like to see all these dimensions brought to the fore. This is not a moralising gesture but rather a relational attitude towards understanding how dealing with the more-than-human in interpreter training affects humans’ process of knowing.

Finally, let me offer a word of caution. AI is predominantly a Global North-led phenomenon, at least in terms of design and development, which also means that the social and economic benefits linked to pedagogy are likely to remain geographically concentrated. I would like to see more active work towards bridging the digital divide between the Global North and Global South when it comes to pedagogy. Educators ought to recognise that access to AI technology does vary significantly across regions, and this is likely to impact the resources and training opportunities that are accessible to interpreters in the making. Here, partnerships with international organisations or utilising open-source technologies are likely to go hand in hand with developing pedagogical strategies to make AI tools and training programmes more accessible and more affordable. I believe that those involved in CAIL education, such as programme developers and trainers, should also ensure this is catering to diverse learning styles and backgrounds which include promoting linguistic and cultural diversity in the development of AI technology and their use in curricula, especially when it comes to training interpreters in the so-called “Languages of Lesser Diffusion (LLDs)” or indigenous interpreters. AI does promise to produce experts and expertise in a new radical way, but we should not forget that this imagined emancipatory potential cannot be considered in isolation from embedded power structures. We should be very careful as trainers not to magnify any existing injustice.

What is your assessment of the current market uptake of provided alternatives and solutions to human interpreting in the present political and economic circumstances, compared to the reception of AI-enhanced interpreting by end-users? How is this impacting the interpreting industry?

I think the market issue is very pressing, and very much at the heart of my research because I specialise in looking at the labour and market dynamics in the language industry. Generally speaking, we are witnessing a huge growth in the so-called global voice and language intelligence market, which was valued at USD 18.6 billion in 2023 (Market.us, 2023). Subsumed under this, there is also the interpreting market, where there are two main AI trends to highlight: MI (i.e., speech-to-speech and speech-to-text translation solutions, as well as the integration of Automatic Speech Recognition [ASR] which convert spoken language into written text, facilitating tasks like transcription) and CAI (bespoke technology support tools, like computer software that assists interpreters in their workflow, as in terminology management, like InterpretBank). I believe honestly that the most pressing issues in the market concern MI rather than CAI. Therefore, I will focus more on that. If you look, for instance, at the industry landscape, you will see that all the MI solutions available are becoming the most dynamic in this space, and companies that are investing in MI are also in the top global list of most innovative companies in the sector (Nimdzi, 2023). There is both investment in and market uptake of computational power, as well as a growing convergence of remote interpreting solutions, digital platforms, and MI tools. Apparently, the market size just for speech-to-speech translation now at the beginning of 2024 is already more than $500 million, with broadcast, emergency response, military, and contact centres being among the B2B segments that are increasingly utilising the benefits of MI and speech-to-speech technology. I am also observing international organisations, such as the World Health Organization (WHO), which are increasingly adopting MI for mostly mono-directional events, such as webinars, and I know this because I attend those events, and I think that in this case, I can be considered as an end-user of MI myself.

Regrettably, though, studies on the reception of these solutions are, to my knowledge, very scarce. There are some exceptions when it comes to CAI, such as SmarTerp, in which there have been attempts to understand, for instance, the feedback of end-users and how it can be integrated into the stages of development of these tools (Frittella, 2023). But when it comes to MI, I think that industry and research evidence is very scarce. I can only provide a very cautious assessment because the evidence is fragmented. I was looking into the 2023 European Language Industry Survey (Elis Research, 2023) recently, which shows, for instance, that MT is considered a stress factor for around 30% of the surveyed linguists, while AI is considered to be exclusively a threat. I suppose this already speaks volumes about how interpreters are likely to be experiencing a solution such as MI and its market availability. Arguably, the financial aspect and immediate availability will also be noteworthy for businesses and customers. AI translation services can operate 24/7 without the need for breaks and without downtime, which means capability to accommodate different time zones and urgent requests efficiently. In addition, AI can process large volumes of content without the fatigue or the resource constraints faced by interpreters: They do not need to go on breaks! At the same time, automation clearly reduces the costs of interpreting services, which means making large-scale or frequent interpreting tasks more economically feasible. However, arguably the effectiveness of AI in interpreting relies heavily on the quality of the software and hardware used. Issues like poor internet connectivity, low-quality audio input, or system malfunctions can severely impact the quality, and these issues are likely to be a hindrance for end-users. There is also the matter of adaptability to unconventional language use, context, and emotional understanding, which could result in misinterpretations or a lack of empathy. For those end-users like organisations that prefer a professional able to build trust and understand client needs, AI is likely to be seen as incapable of addressing client preferences.

In the past couple of years, I have been closely following the way in which AI is driving innovation in business models in the interpreting industry. For instance, companies are exploring subscription-based services and pay-per-use models; they can access interpreting services, including through AI, for a fixed monthly or annual fee, and this model provides predictability in terms of usage and costs. Some businesses are adopting hybrid models that combine human and AI-driven interpreting services. This approach allows clients to choose the level of service they need, with human interpreters handling complex or sensitive interactions and AI addressing supposedly more routine or repetitive ones. So, given this varied landscape and the limited available evidence, it is really important to foster well-rounded, inclusive discussions to assess the expectations of those who rely on interpreting across settings, particularly crucial settings, including education, courts, governments, and healthcare.

Finally, there is just one brief point I would like to emphasise. I am observing through my studies on the platform economy that digital platforms are emerging as very strong marketplaces where interpreters can connect directly with clients. This matter is not being thoroughly and sufficiently considered by the interpreting studies community. There is amazing work done in the translation studies community, which sometimes I feel is one step ahead of us. I see potential for interpreting studies to address this dimension, too, because these platforms leverage AI for matching interpreters with specific clients and their language needs, streamlining the booking process, even handling administrative tasks, including invoices and contracts, from the beginning to the very end, and such a model is becoming highly normalised as a large-scale production ecosystem, to the point it can be described as the “platformisation” of interpreting. The landscape is becoming also increasingly varied. There are general platforms where interpreting is included among a variety of services. For example, marketing, design, and business, and there are specialised platforms that cater specifically to interpreting. There is, consequently, much “gigification” in the utilisation and provision of interpreting, which implies techno-control of interpreting labour through digital marketplaces; we should be willing to consider this newly emerging business model in parallel with traditional business models for interpreters, such as agencies, because they are really showing how the industry is undergoing transformation in terms of new pricing and service delivery structures. I think we need to open the conversation now before it gets too late.

In your answer to my question on responsible AI, you interestingly brought to the fore the concept of regulation of AI. Now, in consideration of market issues, do you advocate human-first regulation of AI when it comes to AI-powered language and translation apps before they enter the market?

Yes, it directly comes back to your previous question on responsible AI. I think it is likely to be a feasible solution. It would be useful and important, for instance, to implement policies that provide guidelines on the correct use of AI. What also matters is informing and educating stakeholders about the impact of AI on different social groups. My fear, however, is the following. If we assess the regulation of AI as well as the use of AI in companies and in the digital platform economy, it seems that governing institutions are increasingly incapable of keeping corporate interests in place. It has become very apparent in the discussions around the new EU Platform Work Directive, which is attempting to ensure the correct classification of the employment status of people performing platform work and to introduce the first ever EU rules on algorithmic management and the use of AI in the workplace. However, since the companies behind AI have significant stakes in the market, they also remain primary interlocutors in the regulation of AI itself, whether through policymaking or law-making. Furthermore, although guideline policies are being developed, we do not have any guarantee that they will be efficiently implemented and followed. Therefore, more than policies, I would like to see appropriate regulations accompanied by sanctions when it comes to misuse and misappropriation of AI tools in the interpreting industry and beyond. I think if we do not hold stakeholders accountable, then there is not going to be any policy capable of constraining misuses of AI.

What are key aspects of CAIL that need the attention of interpreting researchers?

This is a wonderful question because the field is still so open, that there is so much scope for research which can positively inform the work of interpreters, organisations, and professional associations in the sector. I would like to answer by going back to the very definition of CAIL we discussed at the very beginning of this interview, namely, to recognise CAIL as a sociomaterial accomplishment. I am reprising this point because interpreting studies need to account for a holistic understanding of how AI technologies function, their applications, ethical implications, and potential biases inherent in these systems. Just as NLP and data literacy are not merely about statistical combinations and number-crunching, but involve critical thinking about data sources, data outputs, and their interpretations, the theoretical and analytical treatment of AI literacy must encompass researchers’ ability to critically engage with AI technologies, question their outputs, and use them responsibly. Within that, we should also make space for theorisations of how these dynamics unfold, for instance, by looking at the performance of interpreters and machines, in terms of human–machine interaction. I am honestly a bit less concerned with computational processes because, as a matter of fact, most research available right now in the field is very heavily biased towards descriptive studies of available technologies or the development of computer-assisted corpora, whereas voices from the profession seeking to invigorate the debate remain few and far between. I see this marginalisation coming from an epistemological gap, which to me is the actual crux of the problem: the polarisation between techno-solutionism and techno-criticism, whether it concerns studies of knowledge and skills or considerations of the socio-economic impact of technology. I am firmly convinced that a failure to systematically understand the organisational structures and dynamics that converge with the implementation and impact of AI technologies, including power structures, decision-making processes, and politico-economic contexts, will only intensify the risk of a disconnect between scholarly research and concrete applications, at a time when convergence is much needed to sustainably address AI integration.

We also discussed pedagogy earlier and I think that research should consider CAIL strictly linked to communities of practice (Lave & Wenger, 1991). As a practice theorist, I am often seeing what we call a “weak programme” of research (Nicolini, 2017), limited to trying to define a list of technological skills needed for professional interpreters. Of course, while this kind of research stems from a valid perception that much is to be gained if we are to understand the increasing skill-based needs of contemporary interpreting practice, how to manage workflows, and how to assess output, we risk reducing literacy to what interpreters should do or what interpreters must be capable of doing. The results are likely to be just descriptions that mainly bear witness to the limited familiarity we have with AI in the sector, while leaving more puzzling “so what?” questions open, which risks us not offering any substantial insights into the functioning of AI-related knowledge. So, we need a strong programme that goes much further, because we are living in an era where the knowledge and skills of interpreting are being redefined in terms of meaning and use and in their significance. We need to foreground a social, processual, and collective perspective to understand how pedagogical communities come together to shape the development of CAIL, for instance, in terms of human–machine interaction or human augmentation, instead of simply registering what practical skills should be accumulated from now onwards. In other words, I would like to see more emphasis in the current AI discourse in interpreting studies on understanding how AI practices and CAIL are socially organised and how communities of practice can impact socio-technical innovations.

To conclude, I would like to reflect on the matter of social injustice which is likely to come from the use of AI. This, in my view, happens both within the profession (targeting its workforce) and beyond (among other stakeholders, the communities interpreting is called to serve). With regard to the former, as interpreters we ought to recognise that CAIL is profoundly unequal by default, as we have addressed in a previous question. Available data suggest that levels of digital literacy are relatively low in the Global South, especially regarding low-income countries and developing countries. So, we need research on how to develop CAIL, but for whom? For the already well-equipped interpreters and institutions that can more easily engage in upskilling or to benefit the predominantly Eurocentric interpreting community? We have already seen these dynamics in the outsourcing of remote interpreting to low-income countries. Thanks to the availability of global networks, interpreters are available around the clock on online platforms for lower prices. Research should understand how to address the digital gap if we want to equitably harness the potential gains of digitalisation and automation. Otherwise, we are equipping the interpreting industry and its communities with a tool that only intensifies this division of digital labour. Any effort to acquire and improve CAIL means addressing a moving target, due to the fast-paced development of AI itself. We need research that understands how to harness this target and give equal access and participation opportunities to communities of interpreters, trainers, and students. This is not a trivial matter. It has now become alarmingly clear from a growing body of evidence that AI and its components, such as algorithms, perpetuate injustice for many. This is a point I would like to stress, for instance, we know from MT studies that AI is an incubator for the reproduction of inequality. Many cases are emerging that ethnicity, gender, and poverty bias machine outputs, with such biases augmenting exponentially when it comes to the availability (or scarcity) of languages covered and thus the data available to train machines. Although I am observing an interest in understanding, assessing, and mitigating these biases within the NLP community, I would like interpreting studies to equally maintain these concerns very firmly under their gaze.

Dr Giustini, thank you very much for this interview.