1 Introduction

It is probably safe to say that “artificial intelligence” (in short, AI) has become not only a buzzword, but also one of the most transformative technologies of our time. Together with the current trend of the digitalization of societies, it is bound to have a lasting impact on people’s lives as well as the cultures around them. Interesting models in AI are found in the domain of machine learning that are based on artificial neural networks to perform deep learning trainings and to create fresh predictions upon what the computer has “learned” from a wealth of data [77]. This makes the models potentially more powerful because they can be used for tasks that they have not been explicitly programmed for, although they have been trained to solve a specific task and broad applicability is still in its infancy. Newer models try to incorporate a form of multimodality, meaning that many different input sources are used to generate stronger potency of the algorithms for a broad applicability [2]. However, artificial neural models work with hidden layers of nodes that are in a sense opaque so that not even programmers know exactly what calculations have been performed to obtain the respective results. The main issue with this manifests itself in the problem of “AI alignment” as a concern of “AI safety engineering” [12, 17, 44, 54]. More specifically, it is known as the problem of “instrumental convergence” where a model sets itself a new instrumental goal so that it can better fulfill its original instrumental goal [14, 51]. The problem is that on the one hand, the model becomes nontransparent, and on the other hand, it may have unforeseen consequences. As such, it could eventually harm human agents and thus become misaligned with its human creators’ intentions. An example may help: an AI tasked to eliminate cancer in humans may draw the conclusion that eliminating people would be the most effective way to do so. Hence, eliminating humans would be the new goal to aid in the achievement of the original goal, which was to eliminate cancer in humans. Although such examples may sound a little too much like science fiction, the conceptual problems behind them are real and serious based on the notion that AI-agents become increasingly more self-governed and lack of explainability in their decision formation.

There are many teams working on AI safety and the large institutions in the field appear to implement their considerations responsibly. However, the speed of AI development has increased considerably not only in the past years but also in the last months and weeks. With more than just one key player on the market, there is a competitive landscape for creating ever newer and better models, which poses unique risks and benefits for society at large. Given the tremendous impact of potential machine learning issues, these risks ought to be taken seriously and it should be discussed how they could be potentially mitigated (cf. [4]).

The next chapter consists of a brief overview in the history of machine learning, predominantly focusing on illustrating how fast the recent developments have become. Based on this “economy of AI development”, some of the associated societal risks and benefits are being discussed, followed by suggestions of how these risks could be addressed.

2 The increasing speed of digital innovation in AI models

Interestingly, the idea of an “artificial intelligence” stemming from machines is not particularly new, but for a long time the approach was very different from today. Also, the advent in machine learning based on artificial neural networks was already seen in the twentieth century, although due to a lack of data and computing power, it had not grown popular until the second decade of the twenty-first century. The following discussion highlights that there was an increasing speed of digital innovations in AI models.

1940s

A necessary but not sufficient condition was the invention of the digital computer, which first saw the light of day in the 1940s in the form of the Zuze Z3 and the Atanasoff-Berry computers

1950s

This constituted the emergence of a new discipline now known as computer science, and about a decade later, the famous mathematician Alan Turing [119] came up with what we today call the Turing-Test. The test basically tries to answer the question of when we might want to call a computer “intelligent”. Turing’s answer was that if we could not distinguish a computer’s answers from a human one, then surely, we could call this machinery intelligent. A few years later, in the mid-1950s, John McCarthy was the first person to name this an “artificial intelligence”. Needless to say, the term has all the right connotations and the potential to ignite our imagination for humans to still be intrigued by it [74, 93]

Frank Rosenblatt may arguably be called the grandfather of modern AI research due to his work on the Perceptron in 1957, which appeared to be the first attempt to create an artificial neural network loosely inspired by the adaptive learning capabilities of the human brain. Although his endeavors created a considerable social hype, he did not get very far with his project because the technical limitations of his time did not allow him to move beyond one layer of hidden nodes, or, “neurons”. For this reason, the idea of artificial neural networks was abandoned rather quickly. In 1959, Arthur Samuel coined the term “machine learning” referring to a system that could play chess better than its human inventors

1960s

In the 60s, new electronic systems were used to solve mathematical problems in the real world and applied to computers in commercial factories. A decisive moment was when Joseph Weizenbaum programmed ELIZA in 1965, which was basically the first chatbot. Now for the first time, the public was able to “speak” with a computer. Although Weizenbaum wanted to demonstrate that such interactions were extremely superficial, he later found that many people started to anthropomorphize ELIZA. Hence, already in these days, people started to theorize that one day computers might be able to think for themselves [53]. Three years later, a movie entitled 2001: A space Odyssey created a story with a machine that became sentient and eventually killed the space crew

1970s/1980s

In the 70s and 80s, a lot of effort was made to translate the electronic capabilities onto robotics. Geoffrey Hinton—sometimes referred to as the father or the godfather of deep learning—has taken up Rosenblatt’s idea of the neural network once again and introduced more layers to it, effectively creating a multilayer perceptron (cf. [56]). This was now possible due to better computers and a revolutionary method known as backpropagation [104]. The result was that the program was able to make accurate predictions about data it has never seen before

1980s/1990s

In the late 80s, initiatives incorporated neural networks for working on self-driving cars and in the 90s, Yann Lecun (who is also often counted amongst the AI pioneers and was working alongside Hinton) created a program that recognized hand-written digits. However, for at least another decade the field was stifled by the lack of computing power and the lack of huge data sets. At this time, skeptical voices claiming that AI was not as promising as once thought led to a decline in funding, eliciting what is sometimes referred to as an a “AI winter” [53, 134]

2012

The birth of the modern AI movement can be traced back to September 2012. This is when Hinton and his team were now operating in a world with computers that were fast enough and had enough data due to the internet so that they were able to create AlexNet, an image recognition system that could solve the ImageNet challenge (a large collection of images that had to be labelled independently by the computer) better than any previous technology. The impact this finding had can hardly be overstated. The results were highly promising, which led to a large-scale global adoption of artificial neural networks and subsequently many innovations in the domain emerged [64]. Today, neural network models have become commonplace: Tesla uses it for self-driving cars, Google uses it to learn about our search parameters, Netflix and YouTube use it to estimate which movies we like, Amazon employs it to find what we are likely to buy, Uber has it to allocate the driver as well as to calculate the pricing and the roads, and every year, new start-ups come on to the scene employing new AI innovations. Many apps like TikTok, Instagram or Tinder are completely AI-driven, and the same is becoming more mainstream in the physical space

2015/2016

In 2015, DeepMind’s Alpha Go has won against the European champion Fan Hiu and [47] it defeated Lee Sedol, the world champion, in a very complex game called “Go”. The program learned the game all by itself with no supervision

2017/2018

These exponential AI improvements have become palpably visible especially in the past 6 years. In August 2017, Google introduced the so-called transformer model, which has been foundational for LLMs (large language models) to emerge [124]. Based thereupon, in June 2018 the generative pre-trained model GPT-1 was released by OpenAI [97]

2019

Not even a year after this, OpenAI leveled up their LLM in February 2019 and released the largest model at the time with 1.5 billion parameters (Bp) called GPT-2 [98]. Then, in October 2019, Google released BERT, which significantly improved search queries [87]

2020

Google AI made a huge innovation public in January 2020 with Meena, a chatbot that is conversant in most everyday topics [1]. Only three months later, Facebook’s Meta AI published BlenderBot 1.0, an even stronger chatbot model, which was made available completely open-source [102]. A month after this, OpenAI published GPT-3, today one of the most powerful language models, with a staggering 175 Bp as well as a few- and zero-shot learning approach [16]. In September 2020, The Guardian in collaboration with OpenAI published an article that caught the public’s attention because it was fully written by the AI ([48], p. 3)Footnote 1

2021

In the following two years, the exponential developments have increased even more, due to increasing computing power, big data, and model innovations [38, 134]. In January 2021, EleutherAI issued The Pile v.1, an 800 GB dataset of diverse text for language modelling, aiding future research in the domain [45]. In March, BAAI published Wudao 1.0, which is a “ super large-scale Chinese corpora for pre-training language models” [136]. In June, EleutherAI introduced a new generative pre-trained transformer called GPT-J-6B, which was developed in Japan and had 6 Bp [130]. A few days later, Google used the improvements in LLMs to publish a new chatbot model for dialog applications known as LaMDA (for a revised version, see [117]. Around this time, the Chinese company BAAI updated their product with Wudao 2.0. [141]. Briefly after, the Chinese institution Alibaba Dharma Academy published the M6, an innovative model due to its usage of multimodal databases including texts and images [72]. Just a month later, Facebook’s Meta introduced their improved chatbot model, the BlenderBot 2.0 consisting of 9.4 Bp [65]. Likewise, DeepMind released an open-source license for AlphaFold, an AI that can predict protein structures better than humans can, which is immensely valuable in the fields of biology and medicine. In August, AI21 published the largest LLM with 178 Bp called Jurassic-1 [70], Baidu introduced Plato-XL with 11 Bp with the clear intent of anthropic alignment [7], and Allen AI/AI2 released an 11 Bp QandA-model [111]. In October, NVIDIA and Microsoft hit the new record with Megatron-Turing NLG using 530 Bp (cf. [108], followed by Yuan 1.0 with 245 Bp by Inspur AI from China [135]. In November, Alibaba Dharma Academy revised their M6 model [72] and Coteries in France issued Cedille, a 6 Bp model based on GPT-J-6B [80]. In December 2021, a new player known as Anthropic published their 52 Bp model (cf. [7], Google Inc. and Google AI broke the record with GLaM, a model with 1.1 terra parameters [25] as well as with Gopher, which was noteworthy because it incorporated 280 Bp with novel multimodal dataset [99]. The year went to an end with a Christmas gift from Meta AI using a 13 Bp model named Fairseq-13B [6]

2022

In 2022, OpenAI implemented new data resulting in a new variant of GPT-3 known as text-davinci-002 [91], followed by GPT-NeoX-20B from EleutherAI in February [11, 68]. In March 2022, Google’s DeepMind released Chinchilla with 70 Bp [57], BigScience published tr11-176-ml (only available on the github repository). April brought along PaLM with 540 Bp by Google [83], CodeGen using 16 Bp by Salesforce, VLM-4 by Salesforce [88], the 200 Bp Luminous by Aleph Alpha [52], the 13 Bp mGPT from Sber [107], and the 10 Bp Noor from TII [114]. The most groundbreaking model in April was called Flamingo produced by Google’s DeepMind because it combined a 70 Bp language model with a 10 Mb image model [3]. In May, Meta AI released OPT-175B with an open-sourced code ([137, 138]. And just in the past few days at the time of this writing, Google AI published LaMDA 2 [117], DeepMind released an intriguing model called Gato with 1.18 Bp [100], and Google Research published UL20B [113]

At the time of this publication, the last few weeks showed groundbreaking innovations in multimodality by combining natural language processing (NLP) with computer vision, the most prominent models are Dall-E 2 by OpenAI [105], Flamingo by DeepMind [3], and a few days ago, Imagen was introduced by Google [105]. All these models are incredibly powerful in using text inputs to generate images or taking images and converting them to natural text

2022/2023

Around the switch from the year 2022 to 2023, an interesting business move was made that came along with a considerable social impact. So far, most of these developments had occurred behind an ‘expert veil’ since the general public neither was up to date on the most prominent models (except from anecdotes and scattered media reports) nor was there large-scale interest in them since access was mostly reserved for professionals working in the field as well as some selected beta-testers. In November 2022, [92] decided to go public with a version of GPT-3.5 in a chatbot graphical user interface (GUI) and provide free access to the world. Within days and weeks, global news agencies reported on the introduction of the application dubbed ChatGPT whereas people from all industries started to adopt the technology for their business ideas, writing papers and articles, generating slogans, and helping out with tasks requiring natural language applications [42]. Within just a short period of time, the global community was faced with the questions of how to best deal with such AI models and how to regulate them [41, 50, 96, 140]. At the beginning of 2023, Microsoft [78] as a prominent investor in OpenAI has added to the momentum by announcing that they planned to incorporate GPT into their search engine Bing. Fathoming their business model threatened, at the same time Google publicly declared that they intended to stay ahead of the game and thus to also implement their NLP model called Bard for their search engine [94]. This provoked some popular commentators on the web to proclaim that the “AI wars” have begun [18]

This is supposed to show just how accelerating the speed of AI development has become (an up to date timeline of AI and language model developments can be found here: [116].

3 The rapidly changing economy of AI development

On his YouTube channel, Alan Thompson, a leading researcher in the field, stated at the end of May 2022:

I documented just how quickly this stuff is changing. In March, April and May 2022, we were generating something like a new model every three to four days coming out from places like DeepMind, Google Research, OpenAI, EleutherAI, and the list goes on. It’s almost a full-time job to keep up with it [115].

With this rapid acceleration in the development of the AI economy, there are some benefits and risks emerging. The potential risks are of notable relevance because they might have a considerable impact on human societies, which is why it is important that they are being discussed. In the current paper, the quintessence of the AI economy is conceived in the fact that there is a competitive market of developers that have new ideas for models, data usage and different access to computing power to create new AI systems. Eventually, there appears to be a uniting goal, which is to be the first to arrive at an Artificial General Intelligence (AGI). This is, metaphorically speaking, a race to the moon.

3.1 Benefits of AI competition

In digital business models, the idea of “the winner takes it all” appears to have a lot of merit [66]. The reason for this is that inventors of digital platforms and technologies control not only the access to their know-how and the new possibilities, but they also gain a competitive advantage where customers are willing to pay more for a service than will be the case later on, and as soon as they are connected to the innovator’s infrastructure, it becomes increasingly more difficult to switch partners [125]. As such, the key developers have an inherent incentive to invest heavily in these new technologies and to be faster than their competitors [106, 133].

The benefits in these dynamics largely lie in the associated market dynamics. For a long time, the popularity of machine learning research was confined to a small population of experts. The limitations in data availability as well as computing power did not seem to make the field very attractive and thus, there were not as many players on the market as today [39]. This had three consequences [82]: (i) funding was capped, (ii) human resources were limited, and (iii) there was less potential to share knowledge with other researchers and hence make faster progress together. Economically, this last point has a two-sided effect. On the one hand, a larger market may motivate the agents to collaborate and to share their knowledge to better keep up with the fast developments. On the other hand, each protagonist has the business incentive to be faster than the others so that one can reap the benefits from the major AI breakthroughs. This leads to an ever-increasing acceleration of AI development as long as the players on the market believe that they are close to a profitable turnaround [34, 43].

The main benefit may be the innovations in new machine learning models. One could argue that a fast improvement in these systems eventually profits the consumers and therewith human societies at large since AI solutions can be applied in virtually every setting we dare to imagine [32]. Hence, if there were no risks to them, it would be admirable to get ever-increasingly fast developments until a singularity is achieved (which is the idea that AI systems would surpass human intelligence).

3.2 Risks of AI competition

Of course, market and business principles do not explain all the developments in the field. There are major ethical concerns governing some of the decisions of the leading players. OpenAI, one of the major institutions in the field and the authors of GPT-3, chatGPT and Dall-E 2, is deliberately set up as a hybrid organization between a business company and an NPO with philanthropic goals. They portray a sense of urgency, which becomes apparent in their mission statement where they claim that the top priority of AI development must lie in safety. The team at OpenAI vows that if any of their rivals comes close to achieving a true AGI, they would cease their activities and concentrate on collaborating with this competitor to see the project come to fruition [90]—although it is a philosophical question when (if ever) one would be inclined to grant a system the label of an AGI.

There are three central concerns that should be acknowledged in developing potent AI systems, all of them being manifestations of a discipline known as “AI safety engineering” (for a glimpse into this discipline, see [12, 17, 44, 54]. The first deals with the problem of AI misalignment, the second with the problem of human abuse, and the third with the problem of information control (cf. [127]).

AI alignment is the study of finding ways to make sure that an AI’s instrumental goals do not conflict with our human terminal goals. The previous example of eradicating cancer by getting rid of humans altogether would be a manifestation of clear misalignment where the human goals are not adequately reflected in the AI’s objectives. Since in machine learning, the system can “learn” by itself, it also has the possibility to implicitly set itself new goals to fulfill its programming effectively. Since it is complicated to peek into a system’s hidden layers, it is also not obvious whether an AI has established such a latent goal that might be at odds with human objectives. These things need to be weighed very carefully if a general AI is to be released “into the wild” [19, 139].

Even if there is perfect alignment with the human intentions, the intentions of individuals may not always be for the greater good. This might occur either through ignorance or through human maliciousness. A person could task an AI to do harmful things (either willingly or unwillingly) and hence create damage in the real world harming real people. This is not a problem if researchers experiment with machine learning algorithms in their labs. But as soon as people could access these technologies more broadly, they also would have the potential to abuse them if they wanted to. And just with anything else, there will surely be some people who will want to use these technologies with harmful intentions [28, 75, 127].

The third problem has to do with explainability and information control, meaning that the AI has the capacity to retrieve data but also to create brand new information, either in text but also in picture or video form. To a certain degree, this is already possible with OpenAI’s Dall-E 2, DeepMind’s Flamingo and Google’s Imagine. Soon, it will be impossible to discern between fact and fiction in the digital sphere because everybody would have access to powerful data generators based on AI. This is at least part of the reason why many players have refused to give the public access to these applications at the beginning (cf. [21, 84]).

These topics in AI safety engineering are highly relevant, but chances are high that in an increasingly accelerating AI economy, the speed might lead to an underrepresentation of these risks. This would lead to AI being used by the public or by controversial protagonists prematurely, before either the technology or society would be ready [5, 9, 46, 62, 63, 126]. There are two ways how this could come about: first, the incentives of “going live” might start to overrule the AI safety concerns of the current players, or second, the speed increases the prospects for newcomers so that new agents enter the market that might not share the same concerns because they could be driven by short-term incentives.

3.3 Managing the risks

These rapid AI developments harbor their own economic dynamics, which raises the question of how the associated risks can be managed and mitigated. The risks deal with an underrepresentation of AI safety concerns [127]. The main question, then, is how to ensure that the focus on trustworthiness remains a pivotal concern, both in the books of large incumbents as well as when it comes to newcomers entering the field.

Via EU funds from the Horizon 2020 program, the European Union has attempted to foster developments in the direction of socially acceptable machine learning tools. There are four European networks aspiring research excellence in AI, which are called ELISE, HumanE-AI, AI4Media, and TAILOR. The latter was deliberately launched to work on the topic of trustworthy AI and to establish a network of research centers that can act as expert hubs in the domain. In an interview with Inria, Marc Schoenauer (project-team leader of TAU, an initiative involved in TAILOR), declared: “If systems lack transparency, are biased, are easily affected by noise or attacks, or if their results cannot be explained, then society will never accept AI.”([60], para. 5).

There are ethics guidelines presented by the European Commission [36] stating that trustworthy AI should be:

  1. (1)

    Lawful—respecting all applicable laws and regulations.

  2. (2)

    Ethical—respecting ethical principles and values.

  3. (3)

    Robust—both from a technical perspective while taking its social environment into account.

These points were further specified by seven key requirements for an AI system to be deemed trustworthy. They are the following:

  • Human agency and oversight:

    AI systems are supposed to help humans to make informed decisions and at any time, proper oversight mechanisms must be ensured, such as human-in-the-loop, human-on-the-loop, or human-in-command approaches.

  • Technical robustness and safety:

    AI systems should be secure, accurate, reliable, and reproducible. Additionally, there must be a fallback plan in case something goes wrong.

  • Privacy and data governance:

    The quality and integrity of data should be respected, making sure that access is restricted to the respective agents involved.

  • Explainability:

    The decisions of AI systems should be explainable in a manner adapted to the concerned stakeholders. People always have to be aware that they are interacting with an AI and they should be informed about the model’s capabilities and limitations.

  • Diversity and fairness:

    Unfair biases and discrimination must be avoided. AI systems should be accessible to all, regardless of nationality or other demographic variables.

  • Societal and environmental wellbeing:

    AI systems should come to the advantage of all people, both in the present and the future. Environmental and social impacts should be carefully considered.

  • Accountability:

    There should be mechanisms to guarantee responsibility and accountability of AI systems and outcomes. There ought to be ways to audit the systems, enabling assessment of the algorithms, the data, and processes in place.

Based on these principles, the TAILOR network was formed, which now includes 55 partners across Europe. Its vision is to achieve human-centered trustworthy AI and make Europe the global role-model for responsible AI. Although the aspiration is noble, considering that a lot of the innovation in the field does not only originate from Europe (cf. chapter 2), it may be unlikely that the continent itself will be “the” role-model in the field, but might still be a contributing force. The network is built around six core objectives [112] (Table 1).

Table 1 The six objectives that form the basis of the EU initiative known as TAILOR (2022)
Full size table

The EU guidelines have been criticized as being somewhat myopic, vague, and having a lack of long-term risk considerations. Furthermore, “red lines” like autonomous lethal weapons and social scoring systems had not been given enough priority [76]. They were criticized as having no hierarchy of principles [101], being non-binding, providing little incentives to adhere to the ethical principles, and having a lack of regulatory oversight to monitor the enforcement of EU ethical guidelines [73]. Although useful for providing a starting point concerning trustworthy AI, companies and practitioners worry that the EU guidelines are not specific enough for the harmonization with company law and governance principles, as well as leaving many concrete questions unanswered [55]. A review on the timeline in the European legislation of AI concluded that it is going to be a long road for the specific formulation and implementation since there are many open questions that must be dealt with in a nuanced manner [40]. Another reason seems to be that it may be difficult for regulators to keep up with the pace of the new inventions of modern AI models that keep changing (see chapter 2).

When perceived in light of the rapid advancements in AI development and the associated economic risks in safety engineering as highlighted above, not all TAILOR objectives appear to be equally promising. The first objective is geared towards a network of research centers, which is helpful for advancing knowledge in the field, but does not necessarily mitigate the risk of the developmental speed leading to a loss of vision on safety engineering for certain companies. The reason is that these research centers and the individual AI organizations may not necessarily overlap or collaborate. Similar thoughts can be applied to the second objective. It sets out to define and maintain a strategic roadmap, which is highly relevant for the excellence centers in the TAILOR network, but it is not binding for all external institutions. The third objective wants to create the capacity and critical mass to develop the scientific foundations for trustworthy AI. There are two problems with this: (i) the most promising resources and technologies for generating large statistical models with more than one billion parameters are mostly in the hands of established companies in the industry whereby academic institutions neither have the funds to currently calculate the models nor to keep up with the speed; (ii) and even if TAILOR could generate a trustworthy AI, this would do nothing to impede other industry players that do not abide by the same ethical guidelines. The fifth objective runs into the same problem of that progress in the state-of-the-art foundations of trustworthy AI is only relevant for the players that already subscribe to these ethical tenets. Much more promising are the fourth and sixth objectives. Objective four intends to build strong collaborations between academic, industrial, governmental, and community stakeholders. These collaborations can act as platforms or gatekeepers providing access to the necessary resources, know-how and market touchpoints in the field. If the networks become global and exclusive in the sense that it is almost impossible to succeed without being a part of them, then the networks’ ethical guidelines become a mandatory part of the AI developments, regardless of the speed they take on. Comparable to this, objective six wants to increase knowledge and awareness in the foundations of trustworthy AI, which can influence stakeholders to make sure that the companies abide by established ethical guidelines.

Although the EU has invested a lot of resources striving for laws in the regulation towards trustworthy AI, it is by far not the only regulator to do so. As (Smuha [109], p. 16) stated:

Europe has not been alone in making strides towards addressing the ethical concerns raised by AI systems – and that is a good thing. […] Discussions with other countries and regions, taking place at global fora such as the G7 and the G20, and within other organisations such as the OECD, the Council of Europe and UNESCO, are essential to work towards a much-needed global governance framework for AI.

The initiatives started to become more pertinent [47] when China announced its three-year guidance for the internet and AI plan [20], whereas the US issued their national AI research and development strategy plan [86], and South Korea published the mid- to long-term master plan in preparation for the intelligent information society [47]. In 2017, Canada [22], Japan [110], Singapore [59], Finland [79], the United Arab Emirates [85], and China again [131] issued further regulatory strategic initiatives. In the consecutive years, a host of other countries from Europe, Northern America and Asia began to publish digital strategies and plans for regulations dealing with AI (for a timeline review, see [123]). International collaborations were fostered, among others, by the EU’s communiqué on AI in 2018 [35], the OECD principles on AI from 2019 [89], the EU’s white paper on AI in 2020 [37], and the UN’s strategic plan for AI and robotics for 2020–2025 [122].

A recent research paper on the principles and practices in trustworthy AI highlights that most of the developmental processes and operational advancements occur in the industry. The companies form the sector that eventually connects the users, academia, as well as the government with the market. In the industry, operation occurs along the interfaces of data management, algorithm design, development, and deployment [69]. As such, it is necessary that initiatives working on trustworthy AI form collaborative platforms with companies from the industry that are key players in the market.

3.4 Implications for the (digital) humanities

Almost all academic enterprises are affected by these developments. This is true even for disciplines where this might seem somewhat counterintuitive, like the humanities. Since the speed of innovation is permeating society unhalted, it may elevate critical thinking faculties and the (digital) humanities to a particularly valuable place in these ongoing dynamics. This last chapter poses an example of what this might mean for the humanitarian disciplines.

There are two directions where the current AI developments may become relevant for the humanities: first, the digital humanities can and should preserve a critical humanitarian glimpse upon the rapid machine learning advances that are likely to have a transformative impact on our societies. Second, the humanities by and large may need to evaluate how these evolving technologies could be implemented in their own fields of study in the most promising fashion. Both of these points shall be briefly discussed in the following.

As seen in chapter 2, the AI models are being developed at an accelerating speed. At the time of this writing, in the past few months almost every day a new machine learning model with novel features was published. Already today, custom-tailored digital catering via algorithms defines which books we read, which financial investments we make, which movies we watch, which doctors we consult, which news articles we consume, which food we eat, and even which dates we go on (i.e., through algorithmic matching via Tinder, Bumble and Co). Since machine learning models are getting larger, faster, better and more multimodal, the transformation they will have on our societies is likely to impact every aspect of human life and conversely our social interactions (cf. [128, 129]). There is the danger that technological engineers, software developers and businesspeople do not necessarily have the same goals and incentives as individual consumers, ordinary citizens, policy makers, or societies at large. As such, given the rapid advancements, there may be the risk that the needs of some stakeholders would be overheard [13, 15, 27, 31, 71, 95]. This is where digital humanists can assume a crucial role in these dynamics: with a focus not primarily on profit or technological concerns, they should research and discuss how such changes impact humans and societies, preserving a critical stance towards these developments and at the same time making constructive suggestions to help engineers who often have less time to dwell on ethical concerns, or have to meet deadlines and revenue goals. This has been described as a foundational task of the digital humanities, which would be a valuable contribution to the community [10, 61, 67]. In line with the above discussion, for this to work best it appears necessary that digital humanists stay in close connection with researchers and practitioners from the industry, so that (i) they remain in touch with the rapid changes in the field, and (ii) their critical reasoning influences the discussion as well as the decision-making exactly in the places where the AI systems are being constructed.

At the same time, due to the broad applicability and pervasiveness of AI models, they are bound to play a significant role in all academic disciplines. It is likely that no field will be omitted from these dynamics, which might be comparable to the universal introduction of computers in the 1940s. Individuals, regardless of whether this is in their private lives, in companies or in academia, have become accustomed to the fact that computers are often integrated in what humans do. In the past hundred years and more, computers have not taken away our responsibility to think, act and make decisions, but they have helped us in doing so. They are helping us with storing, processing, and retrieving information, as well as translating some of them to physical labor via robotics. However, due to the inherent limitations of classical computing systems, they still require humans to engage in a multitude of micro-decisions. If machine learning algorithms now are becoming increasingly better, they will push these boundaries further aback, allowing us to hand over more of the repetitive work to the machines. This does not imply that the work humans do becomes mindless or less credible. It only means that there needs to be a form of “AI literacy” when working with them. Neither the industry not science has become worse by enabling all workers to use computers. In fact, the opposite is the case: there are more options to solve problems, and tasks can be managed more effectively. What is necessary, though, is that people who use computers need to know what they can and cannot do with them. The argument here is that this would be no different with the introduction of AI systems. After all, modern machine learning algorithms are nothing else than statistical models based on a huge set of matrix multiplications, and there are publications highlighting the inherent limits of such systems. For example, it lies in the nature of their architecture that AI models will never be able to do things that are quite simple for humans, such as making logical inferences or entertain abstract reasoning—because all they do (and this is what they are really good at) is making extrapolations using statistical operations [29, 137, 138]. However, such claims may be debatable since they strongly depend on what we mean when making reference to terms like “abstract reasoning” or also just “reasoning” per se [26, 49].

In the humanities, deep neural networks have already been applied in a variety of tasks, i.e. Topic Modelling [8, 81, 118] or Authorship Verification [58, 132]. Three years ago, a paper called “The computational case against computational literary studies” by Nan Z. Da [23] has made some headlines by claiming that the literary studies have used neural models incorrectly. One of the points made was that this further exacerbates the replication crisis in the discipline. Kathryn Eccles ([33], p. 90) put it well when saying that “Da’s article had an immediate impact”. It provoked numerous responses with many scholars defending the approaches used in the past. A few days later, Da [24] reiterated the claim in her essay entitled The Digital Humanities Debacle, whereas Ted Underwood [121] pushed back in his defense called Dear Humanists: Fear Not the Digital Revolution. James E. Dobson [30] suggested a synthesis of the two perspectives by stating that going forward, both qualitative criticism as well as computational statistics will be valuable for further advancements in the field. Given the speed of AI development and its potential pervasiveness (see chapter 2), just like in any other discipline, it appears unrealistic that the humanities will not be working with AI models; almost like most of us today work with computers in our everyday jobs. Today, computers do not take away our responsibilities to think, act and make decisions, but they help us to be better informed and to work more efficiently. The same will arguably be the case with the general adoption of AI systems.

Humans and AI will likely work together hand-in-hand, akin to our symbiotic relationship to computers nowadays. Since machine learning algorithms are (i) only statistical models and (ii) remain opaque in their inner workings due to the multitude of interconnected nodes (or “neurons”), they can provide us with important clues but not with definite and secured knowledge. This means that humanists should not simply take an output at face value but as valuable information providing statistical evidence for certain thoughts. The same holds true in the natural sciences: engineers are training AI systems to recognize cancer cells on MRI images. This does not make the physician unnecessary, but instead shifts his or her job to becoming a controlling agent, making sure that the computer has found the right solutions (mostly checking for false positives), thereby making the whole process of diagnosing cancer much more efficient and accurate [120]. In other words, the AI provides valuable information whereas the physician checks the results and makes the necessary decisions afterwards. Another example is DeepMind’s AlphaFold, which is an AI that in many cases works better in analyzing protein folding structures than the way molecular biologists hitherto had to disentangle them. This leads to groundbreaking results in biology and medicine. However, it does not imply that the structures can be taken at face value—the results need to be carefully tested experimentally in the lab to confirm that the folding occurs the exact way that AlphaFold is suggesting. The emergent human-AI interactions produce findings that would have taken decades for biologists to discover, meaning that the whole process becomes a lot more effective [38, 103].

These ideas can be applied to the (digital) humanities: neural network models are not supposed to force a result upon a researcher. The results obtained by an artificial neural network are fuzzy, depending on the learning algorithm not always reproducible, and it is unclear what calculations were performed due to the nature of deep learning architectures. This problem is not unique to any given discipline. However, human experts can use the outputs as evidence and make sense of them by framing the results with their expertise. As such, humanity researchers can—and most likely inevitably will—work with AI systems, just as they today also work with computers.

4 Conclusion

The present paper intended to give a brief overview through the history of AI development, with a focus on demonstrating that there has been a rapid acceleration in these dynamics, especially in the past few months and weeks. Although there are certainly some benefits in these fast market developments, there are also some risks that belong to the domain of AI safety engineering. The main concerns are an improper AI alignment, the abuse of such technologies and problems with informational control (that is, discerning fact from fiction in the digital space). It is important to work on solutions to mitigate these risks on the levels of politics and the law as well as the society. Some ethics guidelines proposed by the EU commission were discussed and framed by TAILOR, an EU initiative working on trustworthy AI. The present paper discussed that some of the TAILOR objectives are more promising than others, given the risks previously discussed. The most promising objectives are the ones dealing with building strong collaborations, and the ones dealing with raising awareness in the domain of trustworthy AI. There are some implications that can be deduced for the (digital) humanities. First, it was claimed that digital humanists must preserve a critical glimpse upon these rapid changes occurring in the world of AI development, making sure that ethical issues remain prominently discussed. Second, it was held that AI systems will become commonplace in both the industry and academia, meaning that humanity experts will have to figure out how these new technologies can be adequately integrated. The suggestion was made that outputs from deep neural networks should not be taken at face value and cannot replace qualitative criticism, meaning that there should eventually be a nuanced expert-AI interaction.