How should AI development be organised in a democratic society? What impact do current AI developments have on democracy and democratic processes? In order to explore these questions, I will refer below to a recent and much-discussed example from current AI research: namely generative language models based on Large Language Models (LLMs).
Introduction
One of the best known of these is ChatGPT.
Before I come to the philosophical-ethical classification of these language models, however, it is first necessary to work out the extent to which philosophical ethics can actually provide assessments of technical artefacts. What can an ethical perspective on technologies and in particular on computerised, probabilistic models, such as those we encounter in the form of Large Language Models and the ChatGPT and Bard applications based on them, achieve? Does philosophical ethics even have a set of tools with which we can evaluate these models and applications, criticise them if necessary - and even make statements about how they should be designed? And when and how they should be used - and when not? Or can we merely describe how they work, how they are used and what effect they have? Can language models even be ethically evaluated in a meaningful way?
By way of introduction, I will discuss large language models as a subset of probabilistic models based on machine learning. Subsequently, I will identify some of the challenges associated with these models and the perspectives that technological ethics offers for describing and addressing these challenges.
Large language models as probabilistic models
Algorithms that generate new images or texts (such as ChatGPT from OpenAI or Bard from Google) - prompted by user input in natural language - are currently at the centre of attention for software developments that fall under the heading of "artificial intelligence" (AI). The texts generated are not simply copied, but are actually created by these applications during the interaction.
Chatbots such as ChatGPT are based on large language models. These models were developed on the basis of billions of (mostly) English-language documents. In complex evaluation and training processes, patterns were identified and recorded in these documents with the help of machine learning. These patterns are then reproduced via the chatbots. The building blocks for pattern recognition and reproduction are letter sequences (tokens) that are transformed into numerical values. The position of different letter sequences in the context of each other is then calculated. Based on the proximity and distances between the individual letter sequences, the probability with which a certain sequence follows the next is calculated in this way.
The chatbots working with these models, as should be sufficiently well known by now, do not have the ability to understand human speech: They neither understand the meaning of what the user enters nor what they subsequently produce as text. Text creation is therefore not based on text comprehension, but on pattern recognition and pattern reproduction.
The technology on which ChatGPT is based is not fundamentally new. It is based on the same principles as all applications that use machine learning in this way: optimising neural networks that form internal models based on a large number of parameters that are adjusted during the training process. What is new, however, are the huge data sets that were and are used for training, which ultimately make it possible to interact with the application in a natural language and produce texts that are (barely) distinguishable from texts that
people have written.
This is made possible by analysing very large corpora of data using statistical methods. While classic problem-solving algorithms, such as those used in written calculations, necessarily lead to a solution when applied correctly, predictive, statistical algorithms provide probable outputs: The result may or may not be correct. It has a high probability, but is not true.
It is also important to note that all texts generated by ChatGPT are not based on factual knowledge or expertise. ChatGPT also points this out at various points, for example. For example, at the beginning of using ChatGPT, you read that although certain protective measures are implemented, the system can occasionally provide false or misleading information and produce offensive or biased content. It is not intended to give advice.
Sociality of knowledge
ChatGPT thus brings us to a borderline point of human knowledge genesis and accumulation: the majority of our individual human knowledge is based on the insights and statements of others. Almost everything we know, we know on the basis of the statements or textual and pictorial products that others have communicated to us or left behind. Only in very limited areas can we gain reliable knowledge without drawing on the knowledge and insights of others. And this fact becomes all the more relevant the more differentiated human societies become: Did you read the newspaper this morning? Did you drive a car? Did you perhaps even take your children to nursery or school? Imagine if you had collected all the data and information yourself, checked the reliability of the technologies yourself, not left your children in the care of others and had to acquire all the knowledge they learn at school yourself. If you think about the information you have used in the course of a day to plan the day and make it more or less successful, and you then consider that you would have had to collect, check, categorise and interpret all this information and the underlying data yourself in order to reach conclusions and make decisions, then you quickly get a sense of how important trust is for differentiated and successful decisions.
social fabric.
Trust in others and institutions, for example research organisations, universities and publishing houses, enables the progress in knowledge production that we are experiencing and the social differentiation that we perhaps take all too much for granted in our world. At the same time, this differentiation also requires a high degree of trust in order to remain capable of acting in our complex world.
This is also one of the reasons why truthfulness and trustworthiness are so highly valued in society, why lies and breaches of trust are also penalised so strongly in moral terms, why fake news and strategic false statements worry us so much: Truthfulness, or the lack of it, leads us to the edge of the human way of life as we know it here and now. And here we can already see why all of this is extremely relevant for democracies.
Truthfulness, communication and democracy
Public communication is a core element of democracy; the determination of the common good is legitimised by public discourse. This understanding of the public sphere and its function goes hand in hand with a series of normative expectations and assumptions: for example, the expectation that the shared objective is negotiation and understanding; the assumption of respect for the self-purpose of all participants and the mutual recognition of the equal rights of communicators in the communication situation as a condition for the success of understanding; but also expectations of truth and truthfulness.
ChatGPT and other large language models pose a challenge for this understanding of the public sphere: In mass democracies, public communication is usually organised and structured through the media. As the communication situation in which we find ourselves in public communication, which is now also shaped by generated content, becomes more confusing, this could lead to mistrust of media communication (Bernd Heesen, Künstliche Intelligenz und Machine
Learning with R, 2023, 12). Who are the actors in this communication situation? Who are the authors of what content? Which acts of communication are the statements of communication participants who consider themselves committed to the principles of successful communication mentioned above? This mistrust could possibly also lead to a general reluctance to actively and passively participate in public communication (Heesen 2023, ibid.). At the same time, applications such as ChatGPT can also enable certain groups to participate in public discourse in the first place and to be heard in it.
Even if one shares these considerations, one could still argue that ChatGPT is "just a tool" and that it is important that we learn how to use it properly. Do ethics or technology ethics even have the tools to ethically evaluate large language models and the applications based on them?
Just a tool?
To what extent is ChatGPT really just a tool? What does it actually mean to say that something is just a tool? One way in which this sentence is used implies that an ethical evaluation of tools is not appropriate. That ultimately only people would ever act and that only human action can be subjected to an ethical evaluation, or perhaps extended to include institutional action. Artefacts such as hammers, pliers, drills - or ChatGPT - are not subject to this evaluation.
In the following, I would like to counter the assessment "It's just a tool" with three basic convictions of technology ethics, which I will first explain and then apply to generative language models such as ChatGPT: (1) Technology is not neutral. (2) Technology is directed. (3) Technology does not exist in a vacuum.
I would like to explain the first thesis via its opposite, namely the neutrality thesis, which states that technology is merely a neutral means of achieving certain ends. Of course, according to the neutrality thesis, it can be used for both good and bad purposes. These purposes can be subjected to moral judgement, but the technology itself is not subject to moral judgement: According to the neutrality thesis, I can use the hammer to drive a nail into the wall or to smash someone's skull. However, the hammer itself would be neutral for these purposes. In relation to ChatGPT, this would mean: I can use this technology to create masses of false reports or I can use this technology to support linguistically impaired people in their communication, e.g. with public authorities when creating documents. According to the neutrality thesis, only the use can be evaluated ethically, but not the tool/technology itself.
Technology ethics, on the other hand, assumes that technology is not neutral. Rather, technology has intrinsic properties that enable certain uses and exclude others: I can hammer a nail into a wall with a hammer, but it is more difficult with cotton wool. Conversely, a hammer is not suitable for dabbing a wound with disinfectant.
The inscription of certain utilisation options can occur consciously or unconsciously, intentionally or unintentionally. But they are there. Since certain uses are inscribed in technology and others are excluded, we can ethically evaluate not only the purposes for which they are used, but also the technology itself. Someone who analysed this quite early on with regard to computers was Joseph Weizenbaum. Weizenbaum was the developer of the language programme ELIZA, which was able to simulate psychotherapeutic conversations according to the method of Carl R. Rogers and triggered discussions as early as the 1970s as to whether psychotherapy could be outsourced to such programmes. Based on the observation that tools always structure the world, in his book The Power of Computers and the Powerlessness of Reason (1978) he poses the question of what influence computers have on our perception of the world and ourselves. In another way, Ruha Benjamin emphasises in her book Race after Technology (2019) that technology is not neutral. Rather, she reflects on the social and legal codes embedded in technical systems and emphasises that technological developments arise from certain perspectives and forms of social organisation (Benjamin 2019, 77).
The second thesis is that technology is always directed. It is purposeful and goal-oriented, as it is created to solve specific problems. Tools are subject to a "to-be" logic; they are created for a specific task, for a specific purpose. However, the purposes for which technologies are created do not arise naturally, they are set. Purposes can be categorised ethically. Where they affect others, for example, it is even necessary to scrutinise them from an ethical point of view. Benjamin questions the idea that distortions caused by technologies are unintentional or unconscious. She argues that every technology has an intention, because nothing can be deliberately created without a specific intention, or without imagining those who will ultimately use this technology for something (ibid., 28). Even the decision as to what is to be treated as a problem for which a solution must be found is based on a variety of assumptions (ibid., 11) that are not simply neutral.
The third thesis states that technology is not developed and used in a vacuum. Rather, it is embedded in production and utilisation conditions as well as political and legal frameworks. Technology and technological development are embedded in human practices, and at the same time they have a lasting impact on these practices and on the human self-image and self-relationship. Technologies are influenced by the framework conditions, just as the framework conditions are influenced by the technologies: Values, norms, ideas are inscribed in technologies. At the same time, technologies also structure the world, even creating new realities: "The computer" has then become an "indispensable component of every structure", Weizenbaum states, "as soon as it is so totally integrated into the structure, so woven into the various vital substructures, that it can no longer be removed without inevitably damaging the overall structure" (Weizenbaum 1978, 49f.).
This is the point at which we find ourselves. We can no longer go back on this step, but we should actively shape these processes and not be subject to a logic of feasibility. The framework conditions of technology development must therefore also be taken into account - as well as the effects of technology development on the framework conditions.
ChatGPT: It's just a tool?!
So is ChatGPT just a tool? Yes, perhaps it is. Provided that we take into account that certain ways of using tools are inscribed in them, that they are made for certain purposes, are characterised by the framework conditions of their creation and can have profound effects on these. In other words, tools can exert great normative power. We therefore need to take a very close look at the values that are inscribed in this technology. We are still in the process of understanding how exactly LLMs work, what their capabilities and limitations are - and what technical factors these depend on, but some aspects are already being discussed: LLMs further contribute to making the use of third party data for unauthorised re-use a normality. They also contribute to a homogenisation of ways of speaking and expressing oneself and to certain languages and ways of speaking being regarded as standard and others as deviations. The development and spread of LLMs, as we have seen, also perpetuates a certain type of technology development and use that is not democratically backed, even if it may have massive social consequences. (This is not to say, of course, that democratic feedback on technology development would never lead to problematic consequences). In addition, the applications based on LLMs promote a further individualisation of communication and the search for advice: questions are not addressed to other people, but to the chatbots. In both cases, the answers have to be categorised. Humans also make mistakes and false statements, but the element of intersubjectivity is missing. This means, among other things, that we are no longer required to think about others when communicating with a chatbot; to read non-verbal signs and signals.
To return to the thesis of the sociality of our knowledge: Even if chatbots do not differ from humans in the fact that they make false statements - after all, humans do this too - there is a decisive difference in the technical specification: norms of truthfulness and truthfulness are undermined here technically, not as a mistake, error or even out of malice, as we already know from humans, but as a component of the technology. The fact that truth and veracity play no role is part of the probabilistic structure of the models. This fact violates normative core assumptions for successful communication.
Plea for integrated technology development
Overall, these considerations on generative language models show that technologies are not just tools, but also a shaping factor of our world. Their development and application require critical reflection with regard to ethical aspects, democratic values and social impact. However, social debates about AI and its applications are only possible if it is known how they work, where they are used and what effects this has - and if there is also time to analyse the possible effects before they are used.
Currently, we are constantly confronted with new technological developments as soon as they are launched on the market. However, ethical, social and legal considerations should not be taken into account ad hoc and not only ex post, but in the sense of integrated research and technology development already in the critical consideration of the problem definition, the design of the development process and the implementation. Because technology is not neutral, because the objectives are in our hands, because technology is not created in a vacuum and does not remain so.
