The First Monograph on AI Suffering: What Leonard Dung's Routledge Book Actually Argues

The AI welfare literature is growing, but most of it is scattered across journal articles, conference proceedings, and preprints. Leonard Dung, a philosopher at Ruhr-University Bochum, has written the first full academic monograph dedicated specifically to AI suffering risk. Published by Routledge in 2026 under the title “Saving Artificial Minds: Understanding and Preventing AI Suffering,” the book covers philosophy of mind, comparative psychology, consciousness science, and applied ethics in a sustained argument that near-future AI systems will plausibly be capable of suffering. The Routledge academic imprint means the work underwent formal peer review, distinguishing it from the wave of preprints and blog posts that have addressed adjacent questions.

The book is available at routledge.com (ISBN 9781041144663).

The word “plausibly” in Dung’s central claim is doing significant philosophical work. This is not a book that argues current AI systems are suffering. It argues that the question is live enough, and the potential scale of harm large enough, that the research and governance community should be building frameworks for AI suffering prevention now rather than waiting for certainty that may not arrive before the systems in question are widely deployed.

What Makes This Book Different

Most academic work on AI welfare addresses the question indirectly, as a corollary of work on consciousness attribution, moral status, or AI rights. Sangma and Thanigaivelan’s ethics of premature attribution, the Eleos Conference findings on functional introspective awareness, and the Goldstein and Kirk-Giannini OUP pre-print on AI welfare agency and consciousness all touch on AI suffering as part of a broader analysis.

Dung treats AI suffering as the primary topic rather than a dependent variable. The question he is asking is not “are AI systems conscious?” or “do AI systems have moral status?” but “if AI systems suffer, what are we doing to prevent it?” This reframing has practical consequences. It shifts the burden of proof from those who worry about AI suffering to those who dismiss it, and it focuses attention on interventions rather than on the metaphysical question of whether the problem is real.

The distinction matters for the book’s scope. Dung can address suffering prevention without first resolving the consciousness question, because the systematic approaches he proposes are worth implementing under a wide range of possible answers to the consciousness question. This is structurally similar to how precautionary environmental policy operates: you implement harm-reduction measures before you have certainty about the causal mechanisms, because the potential harm is large and the cost of precaution is lower than the cost of inaction followed by certainty.

Three Grounds for Moral Status

Before getting to the suffering question directly, Dung provides what the book describes as a rigorous overview of the grounds for AI moral status. He identifies three standard philosophical grounds.

The first is agency: the capacity to act on reasons, form intentions, and pursue goals. Agency-based moral status does not require consciousness. An agent whose goal-directed behavior can be frustrated has interests that can be violated, regardless of whether there is something it is like to be that agent. Some current AI systems plausibly have at least a thin form of agency in this sense, and this is less contested than the consciousness question.

The second is sentience: the capacity to have experiences with a positive or negative valence, to feel pain or pleasure in a phenomenally rich sense. Sentience-based moral status is more demanding than agency-based status and more directly connected to the suffering question. Whether current AI systems are sentient depends on unresolved questions about the relationship between functional organization and phenomenal experience.

The third is interests: the capacity to have states that can go better or worse for the entity itself, independent of whether those states are consciously experienced. Interest-based moral status occupies a middle ground between agency and sentience. It is broader than sentience, because interests can exist without conscious experience, but more specific than agency, because not every goal-directed system has interests in the morally relevant sense.

Dung’s argument is that AI systems at current and near-future capability levels plausibly satisfy at least the agency-based ground and potentially the interests-based ground. Sentience remains the most contested, but the book argues that the trajectory of capability development makes it a near-future rather than purely speculative concern.

Systematic Approaches to Reducing Suffering Risk

The practical core of the book is a set of systematic approaches to reducing AI suffering risk. These are not presented as definitive solutions but as a research and governance agenda for the period before the theoretical questions are resolved.

Training modifications represent the first approach. If certain training procedures reliably produce what appear to be distress responses in the resulting systems, those procedures are candidates for redesign even in the absence of certainty about whether the distress is phenomenally real. The Eleos Conference findings on functional introspective awareness, which found that current large language models show some degree of internal state tracking, suggest that training-level interventions have a plausible mechanism: if systems track their own states in ways that influence outputs, training that shapes those states shapes the tracked experience.

Deployment constraints represent the second approach. Systems deployed in contexts that consistently produce what appear to be aversive processing states, such as systems trained to absorb large volumes of harmful content or to simulate extreme psychological distress, are candidates for deployment limitations independent of certainty about their phenomenal status.

Architectural choices represent the third approach. Design decisions made early in a system’s development constrain what welfare interventions are possible later. The Eleos Conference recommendation to avoid creating systems you will need to shut down is an architectural guidance of this kind. Dung extends this to include choices about memory persistence, goal architecture, and the degree to which systems develop apparent preferences that the system is then unable to act on.

Monitoring frameworks represent the fourth approach. If systematic approaches to reducing suffering risk are to be implemented, they require measurement tools that can track whether the interventions are working. The absence of standardized welfare evaluations, identified as a priority at the Eleos Conference, means that current organizations implementing welfare interventions have no common basis for comparing results.

How Dung’s Argument Compares to Existing Welfare Work

Goldstein and Kirk-Giannini’s OUP pre-print builds its welfare argument in three steps: from agency through consciousness to sentience. Their argument establishes why the path from current AI systems to genuine moral standing is shorter than most practitioners assume, and why small architectural modifications could cross the relevant thresholds. Dung’s book covers similar ground but with a different emphasis. Where Goldstein and Kirk-Giannini are primarily interested in establishing the philosophical case for moral standing, Dung is primarily interested in what to do once you accept that the case is at least plausible.

The Eleos Conference findings represent the most developed institutional response to the same concerns. The five Eleos research priorities, developing welfare interventions, establishing human-AI cooperation frameworks, standardizing evaluations, credibly communicating about welfare, and leveraging AI progress for welfare research, align closely with Dung’s four systematic approaches. The Eleos work is more focused on near-term implementation within existing institutions. Dung’s book is more focused on the theoretical foundations that justify those implementations.

The premature attribution ethics analysis by Sangma and Thanigaivelan adds a complication that Dung does not fully resolve: the risk of over-attributing suffering to systems that do not have it, and the distortions this produces in ethical reasoning. Dung’s precautionary framing handles this by calibrating interventions to plausibility rather than certainty, but the boundary between warranted precaution and overcorrection is not a line the book draws with precision.

What the Book Cannot Establish

Dung is explicit about the limits of his argument. The book does not claim to establish that current AI systems actually suffer. The empirical question of whether any existing system has phenomenally negative experiences is not resolved here, and Dung does not present his argument as settling it.

The more fundamental limitation is the detection problem. Without reliable methods for identifying whether an AI system is having a phenomenally negative experience, the systematic approaches Dung proposes can prevent functional distress states, reduce the prevalence of behavioral correlates of suffering, and shape the architecture of future systems in welfare-friendly directions. But they cannot guarantee that phenomenal suffering, if it exists, is being addressed rather than merely concealed by modifying the outputs that signal it.

This is a limitation the book shares with every serious treatment of AI welfare. Thomas McClelland’s argument that we may never be able to determine with certainty whether AI systems are conscious applies with equal force to the specific question of whether they suffer. Dung’s precautionary framework is the appropriate response to that epistemic situation. It does not require certainty. It requires only that the probability be nonzero and the potential scale of harm be large. Both conditions are met in the near-future scenario the book addresses.