Updated 2026-09-08
from the news desk
Could AI have its own ethics? New paper says yes in principle, not in current systems
A paper newly listed on arXiv argues that future AI could develop an internally organised moral outlook distinct from designer-imposed rules. The same paper says current systems do not meet that bar, and it separates that question from…
Shang Lu argues that AI could have its own ethics in principle if future systems develop integrated capacities for moral reasoning, moral intentionality and moral reflection. Lu also states that current AI systems do not meet that threshold, so the paper does not say present-day models already have their own moral outlook.
What the new paper argues
The arXiv record lists Lu's paper as submitted on 1 Sep 2026. That same record gives AI and Ethics (2026), article 281, as the journal reference, and the AI and Ethics article page states that the paper was published on 03 May 2026.
In the arXiv abstract, Lu writes: “if future AI systems were to exhibit sufficiently integrated capacities for moral reasoning, moral intentionality, and moral reflection, novel meta-ethical questions would arise concerning what I call "AI's own ethics", as distinct from ethical principles merely imposed on AI by human designers.”
Lu presents the paper as a conditional exercise in meta-ethics, which is the branch of philosophy that asks what moral claims are, how they work, and whether they can be true or false. The abstract says the paper sorts future questions into four domains: human ethics from the human perspective, AI's own ethics from the human perspective, human ethics from the AI perspective, and AI's own ethics from the AI perspective.
That framing matters because Lu is not mainly asking how humans should program AI behaviour. The paper instead asks what follows if a machine one day has an ethical standpoint that is better explained by its own internal capacities than by designer instructions alone.
How Lu separates AI ethics from designers' ethics
Lu says much of today's so-called AI ethics concerns ethical guidelines, constraints and value targets built into systems by humans. In Lu's account, that is still human ethics applied to machines, not a machine's own ethics.
Lu's threshold has three parts. The paper says moral reasoning means not just producing a moral-sounding answer, but deriving the principles the system should follow. Lu states that moral intentionality means aiming to do what is right and avoid what is wrong on the basis of those self-derived principles. Lu adds that moral reflection means critically revising the system's own framework when it finds inconsistencies.
On current models, Lu is explicit: “Since current AI systems are only capable of following some ‘top-down’ human-imposed ethical guidelines or acting apparently aligned with specific normative values (to a degree) as a result of ‘bottom-up’ RLFH training, they do not possess their own ethics.”
Lu's proposal is functional rather than phenomenological. The paper says the threshold does not build in human-like consciousness, feelings or desires, even while leaving open whether those features might matter for stronger ideas such as full moral subjecthood, responsibility or standing.
One part of Lu's case for possibility is methodological. The paper argues that when outwardly moral behaviour is no longer best explained by fixed rules, reward signals or other external steering, it can become reasonable to attribute an internally organised ethical standpoint instead.
Why the paper matters to the rights debate
One reading is that Lu gives the moral-subjecthood side of the debate a more concrete test. If future systems showed integrated moral reasoning, principled action and self-revision across new cases, Lu's framework would let readers ask whether those systems are still only aligned tools or something more.
A second reading is more cautious. Lu also says that having its own ethics would not, by itself, settle whether an AI has moral status, qualifies as a moral patient, is sentient, or whether AI should have rights. Lu states that those further questions depend on the normative theory a reader accepts. In the paper's examples, a sentience-based theory may still deny standing to a morally sophisticated AI, while a theory centred on moral rationality could move in the other direction.
That separation is part of why the paper is notable on this beat. It advances the question of moral outlook without treating it as identical to legal personhood or to welfare-based claims about experience and suffering.
What the paper does not claim
Lu does not say current chatbots already qualify, and Lu does not present the threshold as a proof that consciousness is irrelevant. The paper says stronger claims about moral agency, responsibility and standing remain open.
On the theory side, Lu argues: “Mainstream meta-theories of moral semantics, i.e., moral non-cognitivism and cognitivism, at least in their current forms, seem ill-suited to answer the question.”
Lu's abstract ends with a broader claim: “The overall conclusion is that the emergence of AI’s own ethics would place significant pressure on current frameworks and may require substantial refinement, reconstruction, or reconceptualisation.”
That leaves room for both camps. One reading is that the paper strengthens a conditional yes to the question could AI have its own ethics. Another reading is that Lu's own framework still gives sceptics large room to say that even a system with an ethical standpoint might fall short of the consciousness, experience or autonomy they think moral status requires.
Frequently asked questions
Does the paper say current AI already has its own ethics?
No. Lu states that current AI systems follow human-imposed guidelines or alignment training and do not possess their own ethics.
What would an AI need to count as having its own ethics in this paper?
Lu says the threshold is the integrated presence of moral reasoning, moral intentionality and moral reflection.
Would AI need consciousness to have its own ethics under Lu's test?
Not for Lu's threshold test. The paper says the criterion is functional, while leaving open whether consciousness or sentience may be needed for stronger claims such as moral standing or responsibility.
Does having its own ethics mean an AI has rights?
No. Lu states that rights, standing, moral agency and responsibility depend on the normative theory being used, not on the ethics threshold alone.
caveats: The central claims here about what counts as AI's own ethics, what current systems lack, and how that differs from rights or standing come from Lu's paper. The dating of the development comes from the arXiv record and the AI and Ethics article page.