We will now discuss the ethical implications of systems described in Sections 2 and 3. First, we acknowledge that from a philosophical perspective AI systems can be considered neither as moral, trustworthy, nor as intelligent agents. For a detailed discussion on that, we refer to respective publications. In a nutshell, the main arguments are as follows:

As argued, for example in [42], algorithms in general are not accountable, as they lack sentience. Therefore, they cannot be morally responsible. Furthermore, they do not have feelings, so they cannot inherit values. Because of this, we do not consider current machines to be moral or full ethical agents [27, 28], although this is a topic of interest in the field of machine ethics [17, 27].

Attempts to capture moral values are unlikely to deliver usable results that could be used as training data for machines. An example is the Trolley Problem-based Moral Machine project [11] and its critique that “provides reasons why ethical decision-making fundamentally excludes computational social choice methods” [20].

Secondly, the argument about trustworthiness goes along the line of trust and reliance, for which trust exceeds reliance [12]. Trust includes the possibility that a trusted person may fail by not acting normatively as expected. This may draw moral critique. However, since machines are not moral agents, they cannot be blamed. Heilinger concludes that “trustworthy AI seems to elevate AI systems to the status of moral agents, which is not only factually wrong, but also morally dubious, because it hides or reduces the moral responsibility of the appropriate potential targets of trust: those who develop and implement AI” [23]. Thus, AI is not trustworthy, even though it is frequently claimed [e.g. 1].

Thirdly, similar to trustworthiness, intelligence may be misleading in the context of AI, as it targets very narrow domains. This “has little to do with a general creative and cross-sectoral intelligence of humans”, as stated by Heilinger. He concludes that “the term AI itself can give rise to misleading intuitions, which may subsequently also influence AI ethics itself by creating an idea of the technology under consideration that does not match with its reality” [23].1

There are various ethical concerns discussed in literature regarding AI systems [36, 45, 47], including bias, privacy and security, transparency, abuse, authorship, reliability, and environmental impact, among others. As AI systems become more autonomous and automation increases, these ethical issues become more likely and severe. This is directly connected to the level of participation of humans and machines. A strong participation of human users—as moral, intelligent, and creative agents (which machines cannot be)—can lower the impact of ethical concerns.

The main requirement for this is to create an interface that would allow the AI components to support System 1 tasks while also providing means for users to perform System 2 tasks. As described in Section 3, hypertext components that allow users to freely connect information through links (as objects, rather than simply copying and pasting URLs) or through spatial hypertext (using AI or recommender features as described in [31]) would enhance human participation while still utilizing AI support. This is about the potential for expressing human thought in collaboration with powerful machines, which hypertext can facilitate in ways that are currently lacking.

“AI no longer constitutes a narrower subject area”, Heilinger writes, “in a similar way as electricity-supported strategies to solve problems do not constitute a worthwhile subfield of the ethics of electricity” [23]. Hypertext, too, is a cross-cutting topic. In this regard, we must consider the necessary levels of AI versus human involvement in problem-solving tasks for each specific application domain. If humans are to be considered as moral agents, they must be properly educated.

In regards to a currently experienced technological solutionism [29], Heilinger suggests re-evaluating “whether AI is indeed a suitable means to address a problem in the first place” [23]. We argue once again that the issue at hand is not a simple, binary yes-or-no question (as implied by the latter quote), but rather a matter of the gradient level of participation between machines and humans. This ultimately comes down to the relationship between AI and hypertext, with the latter serving as an interface that facilitates a symbiotic and effective cooperation between humans and machines.

Any of the mentioned ethical issues of AI could be evaluated from the perspective of hypertext in the described context; each one is likely worth its own publication. In the following section, we will discuss some of them in the following.

This section discusses the benefits of hypertext in the context of AI-induced ethical issues, focusing on three key Western normative theories. Please note that the depth of this discussion is limited by the paper's scope.

4.3.1 Deontology. Deontology refers to a normative ethical theory that suggests principles for action to be followed by any moral person, regardless of the consequences. A very prominent rule of deontological ethics is the Categorical Imperative, formulated by Immanuel Kant: “Act only according to that maxim whereby you can at the same time will that it should become a universal law”2 [cf. 30]. We want to raise three questions related to the Categorical Imperative:

Do we want AI-based machines to generally take over System 2 tasks, reducing the role of the human user to a mere recipient of information?

Is it generally acceptable to use machines that have been trained on past data?

Do we want AI to appear as humanized agents that evoke emotions in the human user?

The idea of AI taking over System 2 tasks is questionable for several reasons. Firstly, as argued above, it removes moral agents (i.e. human users) and leaves decision-making solely to a machine that lacks moral responsibility. Secondly, reducing humans to mere information recipients can lead over time to a “deskilling” [40] of important System 2 capabilities necessary for decision-making.

The fact that trained AI models are based on past data, introducing bias, makes it unlikely to develop new solutions to problems based on current ethical values. This is especially true if humans’ influence in decision making is reduced in favor of AI systems. Even worse, data that is trained on AI-generated previous data is doomed to become disconnected from human experiences, goals, and ethical understanding in solving current problems and shaping the future.

The scenario in Section 2 mentions that Otto was attracted by the AI's pleasant female voice. This is a common experience in today's AI applications: Personal digital assistants sound realistic, even mimicking emotions and using filler words like “um”, “uh”, “hmm”, or “er” as in our imperfectly spoken language. This creates emotions on the user's side as the tone of an almost perfectly synthesized voice, combined with the use of natural language, appears very human-like and “trustworthy”. Other more extreme examples in the context of creating emotions on the user's side include deathbots [26] or sex robots [14]: The first imitates the voice, tone, language, and knowledge of a deceased relative, allowing the bereaved to communicate with the bot as if it was the deceased; the second imitates a human in voice or physical appearance to which a human feels sexually attracted, allowing the user to engage in sexual activities.

The creation of emotions lies solely with the user. The machine simply imitates them. This makes humans emotionally dependent on the machine, which cannot be considered a moral agent. Instead, the dependency is on the owner or provider of the AI system, which is most likely a company whose goal is to maximize revenue. This is one example of a value conflict that becomes apparent in these examples. These topics are further discussed in their respective publications [e.g. 26].

We can conclude that none of the three questions above would be answered with a “yes”. In other words, none of the claims should become a “universal law”. It is not a matter of bashing AI, but rather acknowledging the level of involvement of both humans and machines in the decision-making process. Introducing hypertext as a human in the loop paradigm that increases the relevance of users, as moral agents, in decision-making processes while leaving AI for System 1 tasks would lower the ethical issues raised in the first two questions.

The third question would be addressed by AI components that appear abstract, rather than using human-like language or voice. One example of this could be recommender-based spatial hypertext systems, as described in Section 3, which allow humans to express ideas or emotions while generating suggestions for augmenting human thinking. As there is no natural language involved, the risk of one-sided emotions and related dependencies on AI providers is greatly reduced. For a description of such a system, refer to [5].

This shows that hypertext supports goals in deontological ethics by empowering human users and utilizing the computational features of AI components.

4.3.2 Consequentialism. Some other ethical aspects can be investigated by the theory of consequentialism, which includes the theory of utilitarianism. This theory of normative ethics considers the outcome of an action, making it a form of teleological ethics (derived from the Greek word télos meaning “aim” or “goal”). From a consequentialist's point of view, the good is to be maximized as the consequences of our actions [30].

The question arises of whether the decrease in human participation in System 2 tasks in favor of AI, as indicated by the scenario in Section 2, helps maximize the overall good. This boils down to the question of whether AI alone (i.e. pure automation) or AI as an augmentation tool for humans (i.e. the symbiosis of human and machine) would be more beneficial in the long run. For the first mentioned scenario (automation), we also have to consider the possibility of deskilling humans with respect to decision-making abilities.

Thousands of years of human culture have shown that humans possess valuable capabilities. Humans are moral agents, while machines are not. This suggests, although not discussed in detail, that the symbiosis of humans and machines is necessary and beneficial for System 2 tasks in order to maximize moral good.

Secondly, it is likely that many of the AI models will be provided by companies. One of their main goals is to maximize profit. As machines themselves are not trustworthy, companies are very likely to subtly manipulate customers towards reaching the respective company's goals. Human-like AI systems can capture human emotions, building one-way “trust”. Companies may misuse this to disguise their inherent monetary values, mimicking a “trustworthy (artificial) person” that captures the customers’ attention and “trust” that would otherwise not be given to companies themselves.

The question we must answer is whether a System 2 task should be performed by an AI (automation) or left to the user while the AI only provides System 1 support (augmentation). Using hypertext systems would allow people to organize their information. Systems that support recommender-based spatial hypertext would enable users to focus on System 2 tasks, while “intelligent” machines augment human thinking and communication. This would reduce the emotional attachment and dependency that a human may develop towards the machine and decrease the influence of companies as they remain System 1 task providers.

Finally, recommender-based spatial hypertext systems require less resource-intensive knowledge bases [31, 44] compared to LLMs. This leads to lower energy consumption in creating knowledge networks [46] and significantly reduces the use of underpaid human labor for training LLMs.

Therefore, from a consequentialist perspective, hypertext supports the value of a healthy environment, reduces energy costs, and avoids modern slavery based on exploiting cheap labor in second or third world countries, promoting equality.

4.3.3 Virtue Ethics. Finally, virtue ethics targets living a good life, focusing on people's virtues and character. Considering that machines do not count as moral agents, as discussed above, humans would contribute the moral perspective. This is most important for System 2 tasks, less important for System 1, of which the latter can be conveniently automated.

In order to allow virtuous, ethically acting humans to become part of the system, we need to provide appropriate user interfaces. Chatbot-like systems, as mentioned in the scenario in Section 2, generate an output that is cognitively challenging for humans to rewrite or restructure as it is represented in natural language. On the other hand, spatial hypertext is a medium that allows for easy organization of visually represented, emerging structures. Other hypertext paradigms, such as node-link structures, allow users to add associations where there were none before. In other words, hypertext allows people to contribute their associations based on their creativity and ethical beliefs—something a machine cannot provide. This allows humans to contribute their moral values.