The oldest technical question is whether the categorical imperative can become a decision procedure. Powers (2005, 2006) poses the problem most clearly: universalization looks formal enough for computation, but only if a system can formulate the right maxim in the first place. Wallach & Allen (2009) therefore treat Kant as the classic top-down option, while Lindner & Bentzen (2018), Kim, Hooker & Donaldson (2021), Singh (2022), Sebti & Ben Hamed (2025), and Olson (2026) translate maxims, duties, and universal law into logical form. Geng and colleagues' UPAR framework (2023) takes the idea into LLM prompting. The gain is precision; the cost is that Kant's richer account of motivation and judgment has to be handled somewhere else.

The harder question begins where rule-following ends. A Kantian agent does not simply behave in accordance with duty; it acts from duty and can regard itself as bound by a law it gives to itself. Schönecker (2022) argues that this requires respect and moral feeling in a sense that cannot be artificially reproduced. Chakraborty & Bhuyan (2024), Çilingir (2024), Manna & Nath (2021), and Brożek & Janik (2019) reach related conclusions from different accounts of freedom and rational agency. Tonkens's 2009 challenge and White's 2021 reply frame the issue well. Sanwoolu's 2025 distinction is helpful because it prevents a false choice: an AI may fail as a Kantian agent while still being constrained by Kantian design principles.

After ChatGPT, Kant becomes useful for alignment talk, but not as a shortcut. D'Alessandro (2024) is especially important here because he warns that deontological constraints can produce paralysis or unsafe behaviour if they are applied naively. Chaly (2024) gives a different emphasis, treating alignment as a fallibilist process closer to Kantian enlightenment than to a finished rulebook. Mougan & Brand (2023) use Kantian ideas to defend procedurally just fairness metrics, while Rathje (2024) and Wright (2023) mark the gap between simulated rule-following and genuine moral reasoning. The common lesson is modest: Kantian rules can discipline AI design, but they cannot replace the work of interpretation.

The responsibility literature asks who is answerable when an autonomous system causes harm. Kantian approaches tend to resist the idea that responsibility simply disappears into a gap. Kiener (2025) argues instead for an abundance of responsibility that needs allocation; Demirtas (2025) deflates the gap; Vallor & Vierkant (2024) redescribe it as a vulnerability gap; and Hindriks & Veluwenkamp (2023) understand it as a control gap. Munch, Mainz & Bjerring (2023) even ask whether some gaps may be desirable. What holds these positions together is less a single doctrine than a Kantian intuition: responsibility has to be organized through roles, duties, and institutions before harm occurs, not merely assigned afterward.

The Formula of Humanity gives the literature its strongest language for human dignity: persons must never be treated merely as means. In AI contexts, that problem is not limited to spectacular cases of manipulation or domination. It also appears in ordinary relations with persuasive systems, chatbots, and platforms that compete for attention. Van der Rijt, Coelho Mollo & Vaassen (2026) argue that treating chatbots as persons can violate a duty of self-respect; Aylsworth & Castro (2024) defend a duty to protect one's attention; and Hanna & Kazim (2021), Knell (2022), Dierksmeier (2022), and Shimizu (2025) develop dignitarian foundations and design duties. Shevlin's work on social AI shows why this matters: the moral risk may lie not in the machine having dignity, but in what our relations to it do to ours.

Kant's political philosophy becomes central once AI leaves the lab and enters administration, policing, welfare, hiring, and public decision-making. The issue is not merely whether a model is transparent in a technical sense. It is whether the reasons that govern people can be made public in a form they can contest. Wright's "Rightful Machines" (2022) asks what lawful AI would require inside a just order. Beckman and colleagues (2024) and Frost (2024) use Kant's publicity principle to criticise machine-learning administration, while Loi, Ferrario & Viganò (2021) recast transparency as "design publicity." Maclure & Morin-Martel (2025) push the point institutionally: voluntary ethics codes are not enough when public power is at stake.

Kant's theoretical philosophy supplies a final set of questions about cognition rather than conduct. The issue is whether machine learning can be understood through synthesis, spontaneity, and the unity of apperception — or whether those ideas mark precisely what current systems lack. Schlicht (2022) reads deep learning through Kant's account of synthesis; Evans (2021, 2022) and Soeteman & van Lambalgen (2024) build apperception into working architectures; and Baiasu (2022) and Stuart & Dobbyn (2002) probe self-consciousness more directly. Shetty (2025) and Seo (2024) use Kant to name the opacity and limits of AI judgment, Berger (2022) extends the discussion to aesthetic judgment, and Hui (2026) reads the whole field through Kant's critical architecture. Here Kant is not simply a source of ethics. He is a way of asking what it would mean for a machine to have a world at all.