Scott Robbins

Across academia, industry, and government, “AI” has become central in research and development, regulatory debates, and promises of ever faster and more capable decision-making and action. In numerous domains, especially safety-critical ones, there are significant concerns over how “AI” may affect decision-making, responsibility, or the likelihood of mistakes (to name only a few categories of critique). However, for most critiques, the target is generally “AI”, a broad term admitting many (types of) systems used for a variety of tasks and each coming with its own set of limitations, challenges, and potential use cases. In this article, we focus on the military domain as a case study and present both a loose enumerative taxonomy of systems captured under the umbrella term “military AI”, as well as discussion of the challenges of each. In doing so, we highlight that critiques of one (type of) system will not always transfer to other (types of) systems. Building on this, we argue that in order for debates to move forward fruitfully, it is imperative that the discussions be made more precise and that “AI” be excised from debates to the extent possible, especially as discussions move more toward discrete systems and use cases. Researchers, developers, and policy-makers should make clear exactly what systems they have in mind and what possible benefits and risks attend the deployment of those particular systems. While we focus on AI in the military as an exemplar for the overall trends in discussions of “AI”, the argument’s conclusions are broad and have import for discussions of AI across a host of domains.

The rapid acceptance and use of large‑language models (LLMs)—including ChatGPT, Gemini, Claude, and related systems—has sparked a contentious debate over their suitability for children. Proponents argue that early exposure equips children with essential competencies for a future dominated by generative AI. They cite potential gains in cognition, personalized tutoring, and social interaction. Critics, however, warn that children’s access to LLMs may jeopardize children’s physical, cognitive, social, and mental health. They echo concerns that later emerged after two decades of unregulated social‑media use. Drawing parallels with the delayed regulatory response to social‑media platforms, this paper advocates a precautionary ban on children’s use of LLMs. First, we delineate the precautionary principle and its four requisite conditions—epistemic uncertainty, credible scientific indication of serious harm, irreversibility of damage, and proportionality of response. We then contrast the rigorous, evidence‑based admission processes governing other products such as medicine with the comparatively lax market entry of digital technologies, illustrating how the latter’s laissez‑faire stance facilitated widespread harm. Next, we show early empirical findings and high‑profile incidents (including deaths) suggesting that LLMs can contribute to emotional dependency, impair problem‑solving skills, and exacerbate mental‑health vulnerabilities among children. Although some studies report modest benefits, the current evidence base remains preliminary and inconclusive. Given the plausibility of severe, potentially irreversible harms and the absence of robust safeguards, a temporary, proportionate ban on LLM use by children emerges as the most prudent policy. This prohibition should persist until developers can demonstrably prove that the net benefits outweigh the risks, thereby shifting the burden of proof onto proponents of allowing children access to LLMs.

Opponents of digital sovereignty characterize the methods to achieve it as authoritarian, protectionist, and anti-innovation – ultimately leading to digital fragmentation. Digital fragmentation, roughly, is the idea that “the internet is in some danger of splintering into loosely coupled islands of connectivity.” The internet is built on, so the argument goes, the foundation of open accessibility, free movement of data and interoperability. The exercise of digital sovereignty, it is claimed, chips away at this foundation – and threatens to fragment the internet as we know it. The argument against digital sovereignty is largely premised on an assumption that is in no way obviously true: the way the internet is benefits individuals. The current functioning of internet services requires consumers to give up large amounts of sensitive personal information that are used to generate targeted advertisements. These targeted advertisements have been associated with election tampering as well as a driving force in the dissemination of conspiracy theories leading to genocide, an armed insurrection, and making it much more difficult to get the world vaccinated. Opponents of digital sovereignty may be correct in their analysis that, in the current context, methods to achieve it will cause the internet to somewhat fragment. However, too often these opponents seem to defend a status quo that is fundamentally broken and dominated by a few companies which have time and again abused their power. Until the world agrees upon rules which protect privacy and ensure the functioning of democracy, the states must accept the trade of a fragmented internet for the protection of their citizens.

In this paper I assess the ethical and epistemic utility of explainable AI algorithms. I first distinguish between different types of outputs that AI can have. The first class of outputs are verifiable (either through a third-party or in virtue of contributing to a win in a game scenario) – that is, there is a way to independently verify the outputs of the model. The second class of outputs are non-verifiable and include outputs like ideals (finding the best of something) and generative AI. While there is some epistemic value gained by explaining the outputs of verifiable AI, there is no intersection between explanations offered by xAI and ethically useful explanations. Therefore, if there is an ethical problem with the use of opaque AI systems, explainable AI won’t be able to help solve it.

This paper examines the implications of generative AI (GenAI) emulating human expression, i.e. human communication and human creative expression. While GenAI seems to offer benefits such as increased efficiency and productivity, its use raises significant practical and conceptual concerns: GenAI comes with the increased efforts of prompting, verification and editing, and causes the deskilling of its users. It also comes at a monetary cost and causes various ethical issues e.g., a lack of authenticity. We further show that GenAI’s fundamental issue is that it is by design not able to output human expression but only human-like expression. Using AI for tasks that are fundamentally about communicating is replacing communication with something that is not communication. Finally, we show that the consequences of the use of GenAI cannot be avoided on an individual level by those individuals avoiding the use of GenAI and will necessarily lead to an erosion of human expression in general. This is because GenAI: will lead to a distrust in human expression when the authenticity of authorship over time becomes unclear; will cause a devaluation of human expression when human expression can be mimicked with less effort by GenAI; and, will discourage human expression altogether when GenAI has set the bar too high.

Meaningful human control (MHC) is increasingly becoming an important topic in AI ethics beyond the domain of autonomous weapons systems. MHC has been conceptualized, analyzed, and applied. However, in this article, I show how all the current attempts at realizing MHC have fallen short because we have not taken the important first step of deciding what machines should and should not be doing in the first place. We must first ensure that the output we have delegated to the machine is appropriate – only then do we have to do the work required to ensure that MHC is realized. Here, I argue that machines should not be evaluating – that is, we should not be delegating evaluative outputs to machines. This is practically important because machines that have evaluative outputs cannot be evaluated for efficacy. We can’t say how effective they are. This is ethically important because machines don’t have the moral agency required to make evaluations. Furthermore, machines should not be in a position to change our values – which they would be if they were evaluating. Finally, evaluations are judged based on the considerations used to reach a particular judgment. Contemporary AI cannot provide these justifying considerations – so we have no way of evaluating their evaluative outputs.

When a person recommends a restaurant, movie or book, he or she is usually thanked for this recommendation. The person receiving the information will then evaluate, based on his or her knowledge about the situation, whether to follow the recommendation. With the rise of AI-powered recommender systems, however, restaurants, movies, books, and other items relevant for many aspects of life are generally recommended by an algorithm rather than a person. This volume aims to shed light on the implications of this transnational development from both legal and ethical perspectives and to spark further interdisciplinary thinking about algorithmic recommender systems.

Recommender systems are increasingly being used for many purposes. This is creating a deeply problematic situation. Recommender systems are likely to be wrong when used for these purposes because there are distorting forces working against them. RS’s are based on past evaluative standards which will often not align with current evaluative standards. RS’s algorithms must reduce everything to computable information – which will often, in these cases, be incorrect and will leave out information that we normally consider to be important for such evaluations. The algorithms powering these RSs also must use proxies for the evaluative ‘good’. These proxies are not equal to the ‘good’ and therefore will often go off track. Finally, these algorithms are opaque. We do not have access to the considerations that lead to a particular recommendation. Without these considerations we are taking the machine’s output on faith. These algorithms also have the potential to modify how we evaluate. YouTube has modified its algorithm explicitly to ‘expand our tastes’. This is an extraordinary amount of power – and one that if my first argument goes through, is likely to take us away from the good. This influences our behavior which feeds back into the algorithms that make recommendations. It is important that we establish some meaningful human control over this process before we lose control over the evaluative.

The concept of Meaningful Human Control (MHC) has gained prominence in the field of Artificial Intelligence ethics. MHC is discussed in relation to lethal autonomous weapons, autonomous cars, and more recently, AI systems in general. Little, however, has been done to analyze the concept. Those using MHC tend to look at it narrowly and intuitively—as if it is clear what it means. They fail to see the many issues concerning human control over machines. In this article, I break the concept into its three constitutive words (‘meaningful’, ‘human’, and, ‘control’) to outline the many meanings of MHC. While the intention is not to come to the real meaning of MHC, this analysis brings up the many issues that should be considered if meaningful human control is to be realized. These include: which humans count as meaningful in the application context, whether the control those humans are given must be meaningful, whether humans must retain control over the things that are meaningful in life, whether the style of control is human-like, whether each actor (designer, operator, subject, government) has the control they need, and what it is exactly that a human is controlling (e.g., the training data, the inputs, the outputs, etc. of the AI system).

Artificial intelligence (AI) is becoming increasingly important for the infrastructures that support many of society’s functions. Transportation, security, energy, education, the workplace, the government have all incorporated AI into their infrastructures for enhancement and/or protection. In this paper, we argue that not only is AI seen as a tool for augmenting existing infrastructures, but AI itself is becoming an infrastructure that many services of today and tomorrow will depend upon. Considering the vast environmental consequences associated with the development and use of AI, of which the world is only starting to learn, the necessity of addressing AI alongside the concept of infrastructure points toward the phenomenon of carbon lock-in. Carbon lock-in refers to society’s constrained ability to reduce carbon emissions technologically, economically, politically, and socially. These constraints are due to the inherent inertia created by entrenched technological, institutional, and behavioral norms. That is, the drive for AI adoption in virtually every sector of society will create dependencies and interdependencies from which it will be hard to escape. The crux of this paper boils down to this: in conceptualizing AI as infrastructure we can recognize the risk of lock-in, not just carbon lock-in but lock-in as it relates to all the physical needs to achieve the infrastructure of AI. This does not exclude the possibility of solutions arising with the rise of these technologies; however, given these points, it is of the utmost importance that we ask inconvenient questions regarding these environmental costs before becoming locked into this new AI infrastructure.