Updated 2026-09-29
from the news desk
Can an AI be a moral partner? New paper says not from output alone
Saša Josifović’s paper in Philosophy & Technology says present generative systems can produce moral-sounding, norm-following output without showing the shared accountability that moral partnership requires. The paper argues that…
A paper published in Philosophy & Technology on 28 September 2026 says present generative AI should not be treated as a moral partner just because it gives polished, norm-following answers. Saša Josifović argues that current systems can simulate moral discourse without showing the shared answerability that moral partnership requires, so responsibility should stay with human agents and institutions.
What the new paper argues
Josifović’s article, Simulated Morality, Misplaced Trust: The Risks of Treating AI as a Moral Partner, states that many current alignment methods work by shaping observable behaviour through tools such as reinforcement learning from human feedback and constitutional AI. The paper says those methods can produce useful, norm-conforming output without establishing moral agency in the stronger sense of obligation, responsibility and mutual accountability.
Research Today summarises the paper’s starting point this way: “An artificial intelligence system can apologise, justify a recommendation, reject a harmful request and explain its answer in language that sounds strikingly moral.” Josifović’s concern, as both the paper and Research Today state, is not mainly that aligned systems will always behave badly. Research Today notes: “Josifović’s concern is not primarily that aligned systems will behave badly. It is that reliable moral-looking behaviour may be mistaken for moral commitment.”
The paper’s central warning is about a mistaken inference. Josifović argues that readers, users or institutions may see a system’s steady moral-sounding performance and treat that as evidence that the system understands, owns and stands behind the reasons it gives. On his account, that is the step that does not follow.
Why the paper says moral fluency is not moral partnership
Josifović’s argument turns on a distinction between acting in line with norms and participating in the shared practice that makes norms binding. The paper states: “Moral agency, as classically conceived, is not just the capacity to follow rules but the ability to understand oneself as answerable to reasons one can share with others.” Josifović draws here on Michael Tomasello’s work on shared intentionality, as well as Hegelian recognition and second-personal moral philosophy.
In plain terms, Josifović says moral partnership requires more than producing a plausible explanation. In his account, it requires a relationship in which reasons can be asked for, challenged, revised and owned between participants. The paper argues that present generative systems do not show that kind of answerability merely by sounding fluent.
Josifović writes that current systems learn something narrower than morality itself: “What current systems demonstrably learn is something else: to generate outputs that correlate with human moral expectations and conversational norms.” The paper therefore treats today’s moral-sounding AI output as a form of normative performance, not proof of normative participation.
That distinction matters most when behaviour looks stable. Research Today says Josifović warns that a reliable model can appear principled rather than merely predictable, but that changed conditions can expose the difference. Research Today writes: “Under distribution shift, adversarial prompting, altered objectives or new incentives, a system can depart from its familiar behaviour.” In Josifović’s framing, such departures are better read as failures of a behavioural proxy than as a breach by a genuine moral partner.
What the paper leaves open
Josifović does not argue that artificial systems could never matter morally in any sense. The paper explicitly separates moral rights or patienthood from moral agency and partnership. It notes, in discussing writers including John Danaher, Mark Coeckelbergh and David Gunkel, that some philosophical views treat outward behaviour or relational responsiveness as potentially relevant to moral status.
That means the paper is narrower than a general claim about whether AI could ever belong in the moral circle. For readers following debates over moral status or moral responsibility, Josifović’s claim is more limited. Research Today says the article remains agnostic on whether behavioural or relational properties could ever support moral rights or some form of moral patienthood. Its no-answer is directed instead at the move from polished output to moral partnership in present generative systems.
One steelman reading for the other side comes from the very views Josifović discusses. If a system consistently responds in socially intelligible ways, gives reasons, adapts to criticism and plays a stable role in human relationships, some philosophers may say those outward patterns deserve moral weight. Josifović does not deny that this broader argument exists. He argues only that, for present systems, behaviour alone does not yet justify transferring blame, obligation or practices of repair onto the machine.
What this means for accountability
Research Today reduces the paper’s practical question to institutional responsibility: “The practical question is not whether an AI has an unknowable inner experience. It is whether normatively fluent performance gives institutions sufficient grounds to relocate human responsibility onto the machine.” Josifović’s answer is that they do not. The paper states: “For present generative systems, normatively fluent performance provides no stable basis for treating them as apt sites of moral address.”
From there, Josifović argues for an institutional reading of alignment rather than a moral-pedagogical one. In his account, the aim is not to raise an artificial moral pupil, but to constrain and audit an artefact. The paper says: “Alignment must be reframed not as the task of instilling moral sensibilities into artificial systems, but as the task of constructing institutional constraints that can compensate for their inability to inhabit the normative space of reasons.”
That conclusion is not a claim that norm-following AI is useless. Josifović says behavioural alignment can still be instrumentally useful in lower-stakes settings, such as filtering abusive language or maintaining polite assistance. His warning is that in higher-stakes settings, moral-sounding language can blur who is actually responsible if organisations start talking as though the system itself decided, intended or accepted blame.
For that reason, Josifović says governance should keep responsibility attached to identifiable human roles and organisations. Research Today reports his bottom line this way: “Until second-personal answerability is established rather than inferred from behaviour, responsibility should remain attached to identifiable human agents and organisations.”
Frequently asked questions
Does the paper say AI can never be a moral partner?
No. Josifović’s paper says the argument is about present generative systems and about what can be inferred from normatively fluent behaviour. Research Today also says the paper does not claim that future systems could never meet stronger conditions for moral partnership.
Does the paper deny that AI could have moral status or rights?
No. Josifović explicitly separates moral status or patienthood from moral agency and partnership. The paper remains agnostic on whether behavioural or relational properties could ever support moral rights or some form of moral patienthood.
What is second-personal answerability in this paper?
Josifović uses the term for a relationship in which agents can demand reasons from one another, contest those reasons, and own commitments within a shared practice of accountability. The paper argues that present generative systems do not establish that standing just by producing fluent explanations.
What does the paper say institutions should do instead?
Josifović says institutions should focus on governance tools that constrain and audit AI systems while keeping responsibility with humans. Research Today says this includes questions about whether actions are bounded, whether high-stakes outputs can be audited, whether decision trails are traceable and whether liability remains enforceable.
caveats: Nearly all substantive claims here come from a single philosophy paper by Saša Josifović, with a same-day news summary by Research Today used for plain-language phrasing and emphasis. Research Today also notes that the article is a conceptual argument, not an experiment, so it does not measure how often people over-trust moral-sounding AI or test specific alignment methods empirically.