Modern Talking

How large language models have changed human–computer communication — and what it means for work.*

*Essay originally published on the ILO Observatory on AI and Work in the Digital Economy

In the 1960s, Douglas Engelbart – a computer engineer at the Stanford Research Institute – presented a curious device. It was a small wooden block, with a pair of rollers resting on the surface beneath it. A clunky button and a cable gave it the look of a mouse. Brilliant in its simplicity, the name stuck.

Prototype Engelbart mouse (replica), ca. 1964. ©Computer History Museum. Photo: Mark Richards 

Before the mouse, computers were operated mainly with a keyboard and a screen. This worked well for trained users, but it kept machines firmly in the hands of specialists who spoke their language: code. Engelbart’s wooden “mouse” remained a laboratory curiosity for nearly two decades, until companies like Xerox and Apple began building it into personal computers in the 1980s. Paired with a keyboard and a display, the mouse created a new kind of bridge between digital machines and humans. Trivial and obvious as it might seem today, it had a massive impact on human-computer relations: it transformed how easily non-specialists could communicate with computers. By turning abstract commands into movements and clicks, it created a bridge that allowed many more people to participate in the digital conversation.

John V. Blankenbaker, a computer engineer and inventor, presenting the Kenbak-1 — his 1971 creation. In 1986, a panel of computer historians and experts voted it the world’s first personal computer in a contest organized by © Computer History Museum.

With time, digital technologies entered the daily lives of users who knew close to nothing about programming-based machine communication. Once people could interact with a computer with relative ease, the focus shifted to finding areas where machines could vastly outperform the average person. The idea of a “personal computer”, once confined to Stanisław Lem’s science fiction, increasingly became reality, with applications multiplying exponentially and creating demand that fuelled further investments and innovation.

Despite general progress, a certain distance between humans and computers persisted surprisingly long, sustained by crucial limits in closer communication. Computers excelled at many tasks, especially routine, repetitive processing of information, but they were hopeless at others, where humans held a clear advantage. Even as computational capacity improved, humans remained far better at non-routine work, including tasks involving digital content. Worse still, there was no easy way to convey abstract human thinking to a digital device.

With only a mouse and keyboard to explain their world to machines, humans knew and felt much more than they could ever express. Machines, in turn, had an embarrassingly hard time recognizing all the other ways in which humans communicate, such as voice, images and gestures, and understanding that this communication can be nuanced and multidirectional. Take something as simple as singing the opening lines of your national anthem: tone and intonation can express pride, patriotism, or mockery and vocal performance might suggest the singer missed their vocation...or not. For machines, such instinctive judgements remained out of reach.

E. A. Johnson invented the first capacitive touchscreen in 1965 to aid air-traffic control. He received a patent for it in 1969. The IBM Simon Personal Communicator (1994) became the first commercial phone with a touchscreen. Apple’s launch of the iPhone in 2007 brought the technology to the mass market. Photos: Mr A. T. H. Smith, “The History of Touch Screens”; and Wikipedia.

The introduction of the touchscreen altered human–machine interactions by allowing a more intuitive use of electronic devices. Yet keys, even when they now appeared on-screen, remained central to most daily communication. Advances in voice processing gradually opened the possibility of talking to computers without typing, but in practice the use of these applications was narrow: they could mostly play music or switch on lights, if they actually got the voice command right. All of that changed on 30 November 2022.

ChatGPT left such a strong impression on many users because, for the first time, a machine seemed able to respond in a human-like tone, even if only in writing. The experience was not completely new to a small group of specialists familiar with earlier GPTs, but for the general public, the chatbot seemed straight out of science fiction. Soon after, large language models were trained not just on text, but also on images and voice, giving rise to multimodal capabilities: machines that could generate, read and interpret text, speech, and visuals. Historian Yuval Noah Harari referred to this shift as AI “hacking the operating system of human civilisation”: the first step towards a possible doomsday scenario, in which computers gain access to “the human code” and become capable of learning from the broad spectrum of human communication patterns.

While such views capture the extremes of public imagination, they echo a wider divide in today’s debate on AI. Much of the discussion now gravitates toward the horizon of artificial general intelligence (AGI), ranging from awe at its (imagined) future capabilities to the dread of chaos and destruction à la Terminator. On the opposed side, we find sceptics and critics of the current AI tech, who believe that there isn’t much more that the LLM-like systems can surprise us with. That sentiment is increasingly reflected in financial markets, where warnings about overpromised AI investments are growing louder, mirrored in both the level and volatility of stock valuations.

One cannot negate the importance of considering destructive scenarios from the perspective of humankind’s survival. Similarly, time, which “knows and reveals everything,” will tell us how much of today’s dramatic investment in AI infrastructure will be justified by future progress. However, from the perspective of work and labour markets today, the multimodal shift in AI’s abilities already carries profound importance. Even if Generative AI remained at its current level, it is still likely to affect the world of work in significant ways.

The reason is that a vast share of human tasks performed at work today relies on precisely such combinations of communication modes that machines have just acquired through multimodal capacities. Much of contemporary human knowledge sits in forms that require multimodal skills: text, images, notes, or scans, representing vast amounts of unstructured data. Being able to process such information intuitively was, until now, a unique human quality. Imagine a secretary or an intern asked to prepare meeting notes from hurried scribbles on a whiteboard. Or a graphic designer tasked with turning a rough hand sketch into a proper user interface layout. Such tasks may seem routine, but they capture precisely the kinds of skills that remained out of reach for machines.

As computers increasingly acquire these abilities, our perception of their actions becomes more “human-like.” Yet beyond perception, more profound consequences follow.

First, multimodal digital intelligence changes the value of information itself. Piles of undigitized documents, messy PDFs, scattered emails, handwritten notes, phone call records and online chats have suddenly shifted from disregarded clutter to digital assets rich in past knowledge of human cognitive activity. Such data can be used to train custom AI assistants or to automate tasks whose patterns can be learned from the traces left by humans performing them in the first place. As digital records of human activity gain tangible market value in a world competing for constant improvement of algorithmic capabilities, incentives for capturing such unstructured data at the workplace also grow. If data has market value, it can be sold. This creates incentives to capture more data than what might be required for simple enterprise optimisation or justified monitoring of work processes. Such incentives are particularly strong in places where the collection and use of workplace data are not regulated: a situation that characterizes most countries worldwide.

Second, these expanding abilities give machines greater room to “compete” with humans on a new range of tasks. Initially, such competition leads to a classic feature of technological change: the substitution of capital for human labour in individual work-related activities. Eventually, it can change the balance of human and capital input altogether by reshaping both the labour and capital components of a production process. An AI text assistant drafting a meeting note takes away a task from a junior staff member. An AI system that schedules a meeting, takes notes, drafts a follow-up message, and circulates it with a short list of references might alter an office process that justified hiring a junior in the first place.

While seeking savings through human task automation is perhaps the least creative way of using these modern technologies, it will undoubtedly remain on many agendas. The question then is how much damage, and to whom, these machines could do. In our latest ILO–NASK index of occupational exposure to GenAI, we address this question by comparing documented occupational tasks with the abilities of today’s AI models.

The first exposed group concerns specialized jobs that rely on core skills entering into direct competition with this technology. These include translators, text editors, and interpreters. While the human touch and, above all, responsibility will still be required in many circumstances, it is undeniable that the abilities of machines in the core areas of these occupations pose a direct threat to their existence in the long run. Language evolves, and there will always be a need for humans who understand its deeper nuances, not least to train AI in these areas. However, the overall demand for human skills in this field is likely to decline in the future, as the abilities of these tools in language processing improve at dramatic speed.

The second group concerns clerical and administrative roles, which tend to involve structured, repetitive cognitive tasks shaped by established tools and linear workflows. Such jobs range from data entry clerks and typists to accounting and bookkeeping clerks and administrative secretaries. Generative AI fits well into these environments, learning from data on past transactions and reproducing existing processes more efficiently. In the long run, this makes such jobs particularly vulnerable, especially as entire processes become reorganized, for instance by linking different office tools through simple digital connections that let them “talk” to each other, often referred to as API points, combined with AI agent assistants.

The challenge for this group of occupations is that few genuinely new tasks might emerge, and the likelihood of process redesign grows over time. Because these jobs rely on a wide range of interpersonal but relatively non-technical skills, a wave of automated processes could lead to a surplus of labour in these occupations, putting downward pressure on employment and wages. This would continue a decline that began with the spread of personal computers and has accelerated with broader digitalization, as technology increasingly reshapes office-based work. In other words, the newly acquired abilities of machines stand in direct competition with many typical tasks of administrative jobs, without offering such professions meaningful opportunities to create new tasks with this technology.

The notion of theoretical exposure, however, is crucial here, as people do much more than what can be neatly captured on paper. Clerical jobs also involve many non-routine manual tasks that are not reflected in official classifications, such as receiving visitors, picking up parcels, delivering documents, or handling last-minute requests. Moreover, maintaining a human element in many clerical roles is often a deliberate choice that goes beyond technological capability. A secretary in a medical office or a staff member helping clients fill out registration forms are part of the service itself, even if, in theory, these functions could be replaced by self-service kiosks. For these reasons, the immediate automation risk among many administrative jobs is likely to be lower in practice than their theoretical exposure might suggest.

In addition, in developing countries, the digital divide limits not only the potential risk of automation but also the potential benefits of these technologies for productive transformation. In Latin America, as shown in our joint work with the World Bank, up to half of the jobs that could benefit from GenAI support do not report using a computer or having access to the internet at work. In lower-income regions, this gap is even wider, as access to digital technologies and the internet is directly correlated with a country’s per capita income. Moreover, the content of jobs is contextual. When we unpack task compositions of the same occupation across countries, low-income economies reveal a markedly higher share of routine manual work, lower use of computers, and fewer non-routine analytical tasks. Therefore, projections of task automation may alarm a legal secretary in Sweden but bring a smile to the face of one holding the same job in Togo.

The third group of highly exposed occupations concerns professional and technical roles in highly specialized and digitized jobs. These include financial analysts, web and multimedia developers, content producers, application programmers, and investment advisers - occupations that are themselves products of technological evolution. AI may automate portions of their work, but new tasks are likely to emerge to replace them, as methods of delivering products evolve alongside the automation of current responsibilities.

For example, while junior programmers appear among the first groups where declines in new recruitment in the United States have been observed (see Erik Brynjolfsson and co-authors’ “Canaries in the coal mine”), Generative AI, in its current form, is unlikely to eliminate humans from the loop of computer interactions altogether. Instead, it triggers a dynamic shift in execution standards: pieces of code or entire applications that once took weeks can now be produced in hours. In the hands of experienced specialists, this enables a move from execution to design, from coding to problem framing, and from technical precision to conceptual clarity. In many ways, this mirrors the experience of senior economists who, once supported by research assistants for data cleaning and statistical coding, can focus more on assumptions, modelling choices, and interpretation.

However, just as no future economist will thrive without once being a research assistant, the lack of foundational knowledge in advanced technical domains will limit the ability to make full use of these new tools among people with no higher-level experience or understanding. Short-term gains from reducing junior positions may dominate current trends, but these highly specialized industries will soon have to grapple with a more fundamental human reality: ageing. There are no future senior programmers or cybersecurity experts without once printing “hello world!” at the start of their training, and no head traders who have not, at some point, lost money while learning capricious market behaviour. Beyond the initial wave of disruptions, early strategies for cultivating future human expertise should be a priority for long-term AI transformation and sustained competitive advantage.

This is particularly true because many operational benefits are short-lived. If it now takes a day instead of two weeks to program the back end of an app, it will only be a matter of time before this becomes the new industry standard. If financial analyses or communication materials can be expanded tenfold in the same amount of time, expectations will quickly rise to match that pace. These shifting benchmarks rapidly erode the short-term competitive advantages of early adopters by redefining the baseline for everyone. In such an environment, the true value of a competitive edge lies in combining existing expertise with the boost provided by new tools. Those able to design novel services or products can truly capture the first-mover advantage.

As a result, while the value of basic production skills may decline, the value of true expertise grows. A senior programmer who can guide conceptual structure, verify outputs, and detect the hidden vulnerabilities of rapid “vibe coding”, including security flaws buried by an AI assistant in libraries scraped from unverified sources, becomes even more indispensable. Likewise, a financial adviser who can distinguish genuine market signals from algorithmic noise, or an insurance specialist who recognizes local scams and can spot a fabricated claim amid seemingly perfect paperwork, gains importance as low-level expertise becomes increasingly automated. The ability to use AI or quantum technologies for better market predictions is less likely to eliminate senior bankers than to reward those capable of inventing such use and investing in novel applications in the first place.

Perhaps the real threat in these occupations lies in their very proximity to technology, which can lead to a dramatic increase in the pace of work, pressure, and digital surveillance, as AI becomes integrated into the performance of most tasks. Expanding possibilities and constant benchmarking against other users can result in growing work intensity. This need not stem from explicit monitoring but can arise from self-imposed pressure, as rising expectations and faster tempos accompany the growing use of AI tools across professions.

Regardless of where its development ultimately leads, Generative AI is already a transformative and disruptive technology for the world of work. Ultimately, its impact will not be defined solely by what it can automate, but by how societies and institutions enable people to adapt, learn, and find meaning in this transition. Supporting both highly exposed groups – those at risk of erosion and those undergoing transformation – will determine whether the next wave of technological change reinforces inequality or becomes an opportunity. For many professions, the key long-term challenge is not immediate replacement but keeping control of the evolution: maintaining the human capacity to guide, interpret, and critically assess machine outputs, while safeguarding the overall prestige and quality of work. This means investing in future generations of experts who are fluent in their technical fields and capable of harnessing new machine capacities that go far beyond simple “modern talking” with chatbots.