Delegating Communication to AI Without Disclosure: Agency, Trust, and Moral Responsibility

The technology used to transform human communication has evolved far beyond basic grammatical correction. Artificial intelligence (AI) can now alter a message’s argument, tone, and intent, raising a question that goes beyond whether AI saves time or effort. What is the receiver of the communication entitled to believe about the thought and responsibility behind a message when someone receives it under a human name? The ethical problem is especially clear when the recipient does not know that something external to the author participated. A polished apology, a routine appointment confirmation, and an essay may contain equally extensive AI assistance while making different claims about personal attention, authority, and intellectual achievement. This essay presents three modes of communication delegation. In weak delegation, the human originates the communicative purpose, uses AI to refine or draft the message, and critically reviews the result before sending it. In moderate delegation, the human sets the purpose but permits the system to write and transmit the message without reviewing its final content. In strong delegation, the system initiates, composes, and transmits a message in response to a stimulus without the represented person’s awareness of that particular exchange. For example, someone receives and email and an automation responds. These are analytical categories proposed, rather than classifications attributed to the philosophers examined in this essay. The categories focus on the participation over a particular message.

I argue that undisclosed AI assistance can be permissible when the person represented as the message’s author endorses the message and remains accountable for it, and when the message does not mislead recipients about what it says or whose judgment it expresses. Unreviewed communication presented as a person’s considered judgment carries a strong presumption against permissibility, especially in intimate relationships, friendships, education, and consequential decisions. However, the degree of automation alone does not determine whether the delegation is wrong. Prior authorization, recipients’ reasonable expectations about the level of automation, the stakes of the message, and opportunities to correct misunderstandings also matter. This essay draws on several philosophers’ arguments: Kant clarifies duties of respect, Mill tests consequences for well-being and institutional trust, Arendt examines judgment, and Aristotle asks what habits support flourishing. Nietzschean and feminist ethics then challenge assumptions about moral authority and independent agency that could otherwise make my argument too rigid.

Before applying these frameworks, authorship must be distinguished from composition, endorsement, and accountability. A person can endorse words composed by someone else, as when approving a translated letter. Conversely, typing every word does not establish sincerity or responsible judgment. Therefore, defining authorship exclusively as mechanical or cognitive production would decide too much in advance. This definition would treat AI as replacing the person, even when it only helps them express what they already mean. For this essay, authorship concerns whose work a message is presented as. Composition means creating the wording, the composer is the person or system that creates it. Editing means revising wording that already exists, the editor is the person or system making those revisions. Editing may range from correcting phrasing to changing a message’s content or tone. A person may compose a draft that AI edits, or AI may draft a message that a person edits. These roles can overlap, and none alone establishes who endorses or is accountable for the message. Endorsement means accepting what the message communicates. Accountability concerns who must explain, correct, and answer for it.

Non-disclosure is not automatically deception. An omission becomes morally significant when it predictably leaves someone with a false belief that matters to a decision. Recipients need not be told about every tool used to write a message because what matters is whether AI involvement changes what they can reasonably infer about the person represented as its author: that person’s attention, judgment, or endorsement. For example, a recipient may reasonably assume that a personal apology reflects that person’s attention, or that a recommendation records the signatory’s assessment. They might accept AI-generated scheduling information while rejecting an AI-generated account of a friend’s feelings presented as the represented author’s own. This distinction avoids requiring disclosure of mundane spelling assistance while recognizing that silence can mislead when a message implies personal involvement the represented person did not provide.

A foundational question about authority appears in Plato’s distinction between what is pious and what is loved by the gods (Plato, Euthyphro, 10a–11b). Approval and the reasons that justify it are not automatically the same. Applied here by analogy, an employer’s permission or a university’s policy cannot alone explain why undisclosed delegation is morally acceptable. Rules set by employers and universities shape expectations and obligations, but their ethical justification must still rest on reasons such as fairness, respect, and learning. This raises a question similar to Plato’s: is disclosing AI involvement morally right because an employer or university requires it, or should these organizations require disclosure because it is morally right?

Kant locates moral worth in a good will acting from duty rather than in the success of its outcomes (Kant, 1785, section I). His universal-law formulation asks whether an agent’s maxim can be willed as a universal law (Kant, 1785, section II, Ak. 4:421). Therefore, the maxim needs careful formulation. For example, “I will use a writing tool” describes too little to reveal the moral problem. A more relevant maxim could be: “I will present an AI-generated message as my considered judgment without determining whether it reflects a judgment I have actually formed.” The ethical concern is the misleading impression of personal endorsement, rather than the presence of technology. The concealed absence of endorsement, rather than the presence of technology, is central.

Kant’s example of a false promise illustrates a contradiction between a deceptive practice and the trust that the practice presupposes (Kant, 1785, section II, Ak. 4:422). Similarly, a person who delegates an unreviewed AI-written apology may seek to have the recipient take it as sincere and carefully considered, even though that person has not reflected on or endorsed what it says. The connection between communicating an apology and personally acknowledging the wrong can then be broken. The words suggest an act of reflection that may not have occurred. This is an application of Kant’s reasoning, not his own verdict about AI. It is also more precise than predicting that all language would collapse if AI became common practice in communication. A convention of openly authorized automated notices could remain coherent. The difficulty arises when an automated message creates a misleading impression of personal endorsement and invites the recipient to trust that impression. This distinction is especially clear in strong delegation, where the recipient may receive an apology without the person named as its sender even knowing it was sent. The recipient may feel reassured or decide to forgive, believing that the person represented as the apology’s author has acknowledged the wrong and expressed remorse. However, the message provides no basis for attributing those feelings to the represented person. From a Kantian perspective, the disconnect is morally troubling because the recipient is led to respond to an apparent expression of remorse without being able to judge whether the represented person actually endorses it, undermining their ability to make an informed choice about forgiveness (Kant, 1785, section II, Ak. 4:429-430).

Kant states the Formula of Humanity as treating humanity “in every case as an end withal, never as means only” (Kant, 1785, section II, Ak. 4:429). Here, respecting the friend as an end means preserving their capacity to decide how to respond, rather than manipulating them to serve the person sending the apology. For example, consider someone who has hurt a friend and asks AI to produce a convincing apology solely to protect their reputation and preserve the benefits of the friendship. They send it without considering what they did wrong or what responsibility they should accept, while presenting its expressions of remorse as their own. Here, apologizing becomes a way to save appearances. The person seeks forgiveness without providing the acknowledgment that the message suggests. The friend is treated merely as a means to restore that person’s social standing, rather than as an equal entitled to decide how to respond. This parallels Kant’s objection to deceptive promising, where someone pursues their own purpose by preventing another person from understanding the basis of the exchange (Kant, 1785, section II, Ak. 4:429-430). The problem is not simply that AI wrote the words, nor that the person hopes to preserve the friendship. It is that the person pursues this benefit through a misleading appearance of sincerity. Even if the friend feels reassured and never discovers the deception, the person has withheld information that might have changed how the friend chose to respond.

Nevertheless, the Kantian judgment depends on the maxim behind the action and on what the message represents. Johnson and Cureton’s discussion of the universal-law test emphasizes the importance of formulating maxims and notes disagreement about how to identify contradictions in them (Johnson & Cureton, 2025, section 5). This debate cautions against treating every undisclosed use as a Kantian violation without first specifying the maxim and asking whether it can be universalized and whether it deceives or otherwise fails to respect the recipient as an end. Weak delegation can preserve human endorsement, but reviewed work can still misrepresent independent achievement when an assignment requires it. A narrowly authorized automatic confirmation can be permissible when recipients understand its automated nature and it does not imply a personal judgment made specifically for them. Therefore, applying Kant’s approach requires us to ask what the message leads the recipient to believe and whether the agent’s action respects them as rational beings. Mill evaluates actions through their tendency to promote happiness, understood as pleasure and the absence of pain, with the interests of all affected people counting (Mill, 1863, chapter II). Therefore, the benefits of AI communication extend beyond productivity gains for the person using it. AI can help someone communicate across a language barrier, reduce the effort of routine correspondence, or make participation possible when writing is difficult. These benefits should count when deciding whether using AI is ethically justified. A full prohibition would require justification for the opportunities and well-being it sacrifices, including benefits to recipients who receive clearer or more timely information.

Mill argues that an “unintentional deviation from truth” weakens the “trustworthiness of human assertion,” which he calls “the principal support of all present social well-being” (Mill, 1863, ch. II). Applied to AI, the concern includes more than embarrassment if undisclosed AI use is discovered. Unreviewed messages can make promises that no accountable person is prepared to fulfill, cause recipients to act on unsupported claims, and transfer the burden of checking information onto others. For example, a software engineer might save time by using an automated code-review tool. If its feedback is clearly labeled as automated and not presented as the engineer’s personal assessment, teammates can judge it as tool output. The time needed to verify the automated output remains a cost to weigh. But if the feedback appears under the engineer’s name as a review they personally performed, teammates may rely on an assessment that never happened, undermining trust and shifting verification work onto them. A morally relevant calculation must include these benefits and burdens. These are foreseeable possibilities in the examples, not empirical findings that every automated exchange produces harm.

A disagreement among scholars makes the ethical issue harder to assess with Brink explains that Mill has been read through both act-utilitarian and rule-utilitarian approaches. Urmson’s influential rule interpretation is contested (Brink, 2022, sections 2.7-2.8). An act-focused evaluation asks whether this particular automated response produces the best available outcome. A rule-focused evaluation asks what communication practice would promote welfare if generally accepted. This difference affects the case of an apparently harmless concealed message. Its immediate benefit may favor sending, while a practice that makes personal endorsement unreliable may impose wider costs. Therefore, consistent rules about disclosing AI involvement and reviewing messages can be defended on utilitarian grounds because they help preserve trust. This argument does not require resolving whether Mill’s theory is best understood as act or rule utilitarianism. The distinction matters because judging a single action can lead to a different conclusion than judging the general rule behind it.

The theory’s strength is its attention to actual effects and alternatives. Its limitation is the difficulty of comparing time saved, damaged trust, exclusion, and harm to relationships under uncertainty. These effects cannot simply be assigned numerical weights, nor can loss of trust be assumed to outweigh every benefit. Mill’s discussion also allows carefully bounded exceptions to truthfulness when competing considerations of utility justify them (Mill, 1863, chapter II). A routine, previously authorized notice might be justified without individual review. However, that authorization may depend on a particular audience, purpose, and set of circumstances. A message accepted by one group at one time may later reach people who do not share the same expectations, or continue operating after its original context has changed. Therefore, prior authorization provides a reason to expect acceptable consequences, rather than a guarantee that the assessment must consider whether those expectations still apply. An unreviewed medical explanation or disciplinary message would demand stronger safeguards because misunderstanding could cause greater harm. These considerations support a presumption against misleading delegation where significant interests are at stake, while leaving room for justified cases of moderate or strong delegation. Arendt asks whether thinking can interrupt habitual conduct and prepare independent judgment. In her discussion of people following rules without examining them, she says they “get used to never making up their minds” (Arendt, 1971, p. 436). That warning applies when a person treats an automated answer as sufficient reason to stop asking what their words do to another person. Thinking is not simply intelligence or the production of fluent sentences (Arendt, 1971, pp. 417-423, 445-446). Her discussion of the internal dialogue links conscience to the person one must live with after acting (Arendt, 1971, pp. 439-444). The concern is that the person lets the system decide for them without thinking critically about its response. The AI analogy with Arendt work has limits. Ordinary AI correspondence must not be equated with the crimes examined in Arendt’s historical work. Moreover, Tömmel and Passerin d’Entreves distinguish actor-centered and spectator-centered accounts of judgment and identify a tension between them (Tömmel & Passerin d’Entreves, 2024, section 6.4). This prevents treating Arendt as supplying a completed checklist for responsible AI use. Her contribution is to emphasize that people must examine their own actions rather than excuse them by saying, “This is how things are normally done.” The practical inference is that review must concern the message’s meaning and effects, not merely its grammar. A person can click “approve” without genuinely judging, just as carefully designed delegation can embody prior reflection.

Aristotle connects the human good with activity in accordance with virtue across a complete life (Aristotle, Nicomachean Ethics, I.7). This account of eudaimonia challenges the assumption that a faster exchange is necessarily a better one. Communication can exercise patience, honesty, courage, and sensitivity to another person’s circumstances. An apology is not only a text that produces forgiveness, it can be an occasion to recognize a wrong and undertake repair. If a person repeatedly substitutes convincing language for that activity, the ethical cost may impact the character they cultivate even when no single message causes visible damage.

Virtue develops through habituation, but genuinely virtuous action also involves knowledge, choice, and a stable disposition (Aristotle, Nicomachean Ethics, II.1, II.4). Practical wisdom concerns deliberation about living well and attention to particulars (Aristotle, Nicomachean Ethics, VI.5, VI.7-8). This suggests a distinction between delegating a linguistic task and delegating the judgment of what another person is owed. Someone may use AI to compare several ways of expressing a difficult admission, then decide which is truthful and appropriate. Such assistance could deepen reflection. Someone who asks a system to simulate concern and sends the result unread avoids the very encounter through which concern might be practiced. The strength of Aristotle’s perspective is its attention to moral development over time. Its limitation is that it offers no simple rule for resolving conflicts between goods, such as developing our capacities and meeting the claims of relationships; those cases still require practical judgment. Aristotle’s discussion of truthfulness specifically concerns truthful self-presentation between boastfulness and self-deprecation (Aristotle, Nicomachean Ethics, IV.7). It should not be inflated into a ready-made rule governing every disclosure. Virtue ethics does not establish that unaided composition is always superior. For an exhausted caregiver, assistance with routine messages may preserve the time and attention needed for care. The Aristotelian question is whether the practice supports a good life and sound judgment and not if the effort has been maximized.

On the contrary, Nietzsche’s genealogy investigates the conditions under which moral values arose and asks what those values themselves are worth (Nietzsche, 1887, preface, sections 3 and 6). His First Essay contrasts forms of valuation associated with noble self-affirmation and reactive ressentiment (Nietzsche, 1887, essay I, sections 10-11). These arguments challenge the claim that existing communication conventions possess unquestionable moral authority. Applied to the present problem, genealogy asks whose interests are served when polished writing is treated as evidence of virtue, competence, or sincere feeling. Some expectations may reward inherited educational advantages or exclude people who need assistance. However, the fact that a convention has a history does not show that any replacement is justified. Just because AI is popular and ideas about authorship have changed over time does not mean it is morally acceptable to hide that AI is speaking on someone’s behalf. A new convention could improve participation, or it could make vulnerable recipients easier to manipulate. This application does not attribute a verdict on contemporary technology to Nietzsche. It draws a methodological lesson that both the old norm and its proposed replacement need evaluation. If recipients openly understand that a class of messages is automated, one ground of deception may disappear. However, the responsibilities associated with those messages still require an account. Additionally, Bernstein questions whether MacIntyre’s defense of Aristotle gives enough reason to trust the moral standards of a tradition. A practice can have established rules and valued skills without necessarily serving a good purpose (Bernstein, 1984, pp. 6-29). This is Bernstein’s criticism of MacIntyre’s interpretation, rather than a direct argument between Nietzsche and Aristotle. Applied to communication, it suggests that older ways of writing are not automatically better simply because they are familiar. For example, a handwritten apology can be dishonest, while an AI-assisted apology can express genuine remorse if the person represented as its author carefully reviews and genuinely means it. Equally, an automated message is not morally good simply because it saves time. The important question is whether the way we communicate supports honesty, responsibility, and the recipient’s ability to understand the author’s intended meaning. Review and disclosure should be defended for these reasons, rather than because they preserve older habits. From the care ethic angle, Held questions moral theories that focus on people as independent decision-makers while giving little attention to emotions, relationships, and the support people need (Held, 1990, pp. 321-344). She explains that support from others can help people make their own choices, rather than take that ability away (Held, 1990, pp. 342-343). This suggests a limit to judging AI use by how much help someone receives. For example, a person who has difficulty typing might use AI to turn spoken thoughts into a written message. If they check that the message expresses what they mean and choose to send it, they still control what they communicate. Receiving extensive help does not necessarily mean giving up personal judgment or responsibility. The care perspective also makes the recipient’s need significant. Consider a grieving person who receives a polished condolence message. The ethical question includes whether the message reflects attention to that person’s particular loss, rather than merely whether its wording is accurate. By contrast, a person with limited energy may use assistance to maintain relationships that would otherwise become inaccessible. These are applications of a relational perspective, not examples supplied by Held. They show why a demand for visible personal labor in every message could become unfair.

Care ethics contributes contextual sensitivity, but it must also confront unfair demands within relationships. Recipients should not be permitted to demand unlimited emotional labor, and institutions should not shift the whole burden of responsible automation onto people with the least time or authority. Therefore, combining care with justice means explaining AI’s involvement without forcing someone to reveal a disability. Responsibility should also depend on who could decide how the AI was used and what message was sent. For example, a neutral statement that a message was prepared with assistance can explain its production without explaining why assistance was needed. The aim is reciprocal respect and supported participation, rather than a battle between independent speakers and assisted ones.

The contemporary AI-ethics debate connects these concerns to organizational practice. Hagendorff proposes justice, honesty, responsibility, and care as basic AI virtues, supported by prudence and fortitude, to complement principle-based guidance (Hagendorff, 2022, sections 3-4). His discussion treats bias, incentives, and workplace pressure as obstacles to ethical action (Hagendorff, 2022, sections 4-5). The proposal supports my view that stating a disclosure rule is insufficient when an organization rewards employees for speed. Its limitation is that a conceptual virtue framework does not demonstrate that a particular training program will work. From my point of view, the person or organization responsible for a message should ask what the recipient is reasonably relying on: accurate information, independent intellectual work, personal attention, or an authorized institutional decision. They should make sure the message is accurate and does not exaggerate anyone’s personal work, attention, or authority. Consequential claims require substantive review by a qualified person or personal expressions should reflect genuine endorsement. Disclosure should come before it could affect how recipients understand or act on a message. Thus, a disclaimer located elsewhere cannot repair a message that falsely implies a specific human being. For example, in education, this means distinguishing help with expression from the intellectual work an assessment is meant to evaluate. A student may understand and endorse an AI essay and still need to disclose that assistance if the reader is assessing the student’s reasoning and writing. In employment, an automated rejection or performance assessment should identify an accountable person and provide a way to question errors. In personal relationships, disclosure matters when a message implies attention or emotional engagement that the represented person did not provide.

In strong delegation, responsibility should track who authorizes the system. The person who delegates remains accountable for foreseeable consequences, even without approving every message. Lack of awareness may reduce blame for an unforeseeable action beyond the delegated scope. If the person acts on an organization’s behalf, the organization may also share responsibility. Disclosure alone is insufficient and labeling a response as automated does not make an unsafe decision fair or a promise fulfillable.
Ultimately, the ethics of delegating communication to AI turn on how delegation affects endorsement, recipient reliance, credibility, and responsibility. Weak delegation can support agency and flourishing, but it is not automatically permissible when it conceals assistance relevant to an assessment or relationship. Moderate and strong delegation face objections when recipients are led to mistake unreviewed output for a person’s considered judgment. Nevertheless, bounded prior authorization and clearly understood routine automation show why a categorical ban based only on technical autonomy would be too broad. Kant explains why respect cannot be reduced to convenience and Mill requires attention to everyone affected and to the practices that sustain trust. Arendt and Aristotle reveal why judgment and character remain significant even when an output appears successful. Nietzsche and Bernstein challenge the assumption that traditional practices are morally right simply because they are familiar. Held’s relational perspective prevents autonomy from being confused with unaided performance. These frameworks disagree about moral foundations, and this essay does not resolve those disagreements by combining them into a single theory. However, it uses their tensions to test and refine a defensible practical judgment. What the problem reveals about ethical reasoning more broadly is that familiar labels, human, artificial, authentic, efficient, cannot do the work of justification. We must identify the relevant accountability, examine whose agency is enabled or bypassed, and explain how responsibility can be exercised. Communication is not valuable only because information reaches a destination. It also establishes relationships in which people can ask for reasons, challenge representations, and expect an answer. AI assistance is ethically defensible insofar as it helps sustain those relationships. Still, honesty may not answer every concern. Knowing that AI helped shape a message can change how it feels, even when the assistance is fully disclosed. Perhaps the small hesitations and imperfections of human communication are sometimes signs of a particular person’s effort to respond. If so, the question is not only whether AI mediation misleads us, but whether it changes the kind of relationship a message can create.

References

• Arendt, H. (1971). Thinking and moral considerations: A lecture. Social Research, 38(3), 417-446. https://jonudell.net/h/arendt.pdf
• Aristotle. (n.d.). Nicomachean ethics (W. D. Ross, Trans.). Internet Classics Archive. Citations identify book and chapter. https://classics.mit.edu/Aristotle/nicomachaen.html
• Bernstein, R. J. (1984). Nietzsche or Aristotle?: Reflections on Alasdair MacIntyre’s After Virtue. Soundings: An Interdisciplinary Journal, 67(1), 6-29. https://www.jstor.org/stable/41178281
• Brink, D. (2022). Mill’s moral and political philosophy. Stanford Encyclopedia of Philosophy. Course version revised August 22, 2022. https://plato.stanford.edu/archives/fall2022/entries/mill-moral-political/
• Hagendorff, T. (2022). A virtue-based framework to support putting AI ethics into practice. Philosophy & Technology, 35, Article 55. https://doi.org/10.1007/s13347-022-00553-z
• Held, V. (1990). Feminist transformations of moral theory. Philosophy and Phenomenological Research, 50, Supplement, 321-344. https://www.jstor.org/stable/2108046
• Johnson, R., & Cureton, A. (2025). Kant’s moral philosophy. Stanford Encyclopedia of Philosophy. Course version revised October 2, 2025. https://plato.stanford.edu/entries/kant-moral/
• Kant, I. (1785). Fundamental principles of the metaphysic of morals (T. K. Abbott, Trans.). Project Gutenberg. https://www.gutenberg.org/ebooks/5682. Also known as Groundwork of the Metaphysics of Morals. Citations use sections and standard Akademie pagination. https://www.gutenberg.org/cache/epub/5682/pg5682-images.html
• Mill, J. S. (1863). Utilitarianism. Project Gutenberg. https://www.gutenberg.org/files/11224/11224-h/11224-h.htm
• Nietzsche, F. (1887). The genealogy of morals (H. B. Samuel, Trans.; 1913 English edition). Project Gutenberg. Citations identify preface or essay and section. https://www.gutenberg.org/cache/epub/52319/pg52319-images.html
• Plato. (n.d.). Euthyphro (B. Jowett, Trans.). Internet Classics Archive. Citations use standard Stephanus passage numbers. https://classics.mit.edu/Plato/euthyfro.html
• Tömmel, T., & Passerin d’Entreves, M. (2024). Hannah Arendt. Stanford Encyclopedia of Philosophy. Course reading, especially sections 6.2-6.4. https://plato.stanford.edu/entries/arendt/