{"id":"8877649f-7910-4cdc-a46e-531fdcff77c3","arxiv_id":"2507.17020","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The author introduces three facets of AI (scientific, technological, sociotechnical imaginary) and a three-layer responsibility framework for ethical assessment, without providing an executable specification.","lead":"This paper proposes that the ethical problems of AI come from a muddy picture of what AI is, and divides AI into three roles: a science, a set of tools, and a cultural story. It then sketches a three-layer system to judge the ethics of AI use, from clear information to shared values to regulation.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The promised first-order logical theory is asserted, not demonstrated, and sits in tension with the paper's own claim that values are dynamic; without it the central feasibility claim is unsupported.","rationale":"The reader's weakest_assumption identified the unsupported promise of a first-order logical theory, and I agree that this is the most load-bearing point: the paper explicitly ties the possibility of implementation to this theory in Sections 5 and 6. However, I would sharpen the concern by noting an internal tension: the paper itself emphasizes in Section 4 that values are dynamic and unpredictable, whereas a first-order logical theory with fixed axioms is a static formalism. This makes the promissory note not merely under-specified but prima facie inconsistent with the paper's own account of value dynamics. A concrete test would force the author to exhibit at least a minimal axiom system, and would reveal whether the formalism can actually represent virtue-driven principles and changing values without exceeding first-order logic. Because the reader already rendered a CONDITIONAL verdict and my analysis reinforces rather than overturns it, the verdict should remain UNCHANGED. The paper has value as a conceptual synthesis, but its central feasibility claim rests on a condition that is neither met nor shown to be satisfiable.","tokens_in":9179,"tokens_out":4392,"duration_ms":49148,"concrete_test":"Ask the author to provide a minimal formal object for the promised theory: a signature and axioms for one virtue (e.g., 'courage') plus a mechanism for representing changing shared values. Try to implement that fragment in a proof assistant or in a first-order logic theorem prover. If the formalization requires second-order quantification, modal/temporal operators, or axiom revision, then Section 5's 'first-order logical theory' claim is not supported. Alternatively, test whether a fixed first-order theory can derive any nontrivial ethical verdict in a sample scenario from Section 3; if no verdict can be derived without importing dynamic value data, the central feasibility claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the framework 'should be possible and helpful' (Section 6) is made contingent, in the final paragraph of Section 5, on 'a first-order logical theory ... based on which the establishment of virtue-driven ethical principles and corresponding methods ... can be verified.' This is a bare promissory note: no signature, axioms, semantics, or proof sketch is given, and the paper states that an executable specification is only 'under development.' More importantly, the proposal appears to conflict with the paper's own Section 4, which states that values are 'individually and collectively shape shifting' and that 'the values ascribed to specific scenarios are also dynamic and unpredictable.' A first-order theory with fixed axioms cannot represent such dynamics without substantial extra machinery (temporal, modal, or nonmonotonic features), and if the theory is updated as values shift, then 'verification' is relative to a moving target and cannot ground the stable ethical assessments the framework promises. Virtue ethics is also character- and context-sensitive, so reducing it to first-order rules is far from straightforward. Because the paper offers no way to resolve this tension, the implementation and societal-benefit claim is currently unsupported. This is not an objection to the paper's synthesis, but the framework's central feasibility assertion rests on exactly this unmet condition.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a conceptual framework for the ethical assessment of AI use. It argues that ethical concerns about AI stem from a blurred understanding of the concept and that AI is best viewed through three facets: AI as a scientific endeavour (rooted in the Dartmouth workshop and its seven research axes), AI as a technology (yielding four groups of tools: logical reasoning and deduction, NLP, adaptive and evolutionary optimisation, and machine learning), and AI as a sociotechnical imaginary (shared, value-laden visions of AI's social consequences and the stakeholders who promote them). After reviewing virtue, duty, and consequence ethics, the paper proposes a three-layer responsibility framework (transparency and literacy; ethical values; regulations) and connects the three facets to Borgmann's three modes of information (about reality, for reality, as reality). It concludes that implementation of the framework 'should be possible and helpful to ensure that AI tools are used for the benefit of society', resting this on a promised first-order logical theory for virtue-driven ethical principles; an executable specification is stated to be under development.","tokens_in":9457,"tokens_out":14320,"duration_ms":134296,"significance":"Strengths worth crediting: the manuscript is internally consistent and well referenced; it offers a genuinely useful historical synthesis (the seven Dartmouth axes mapped to four technology groups is clear and informative); the three-layer responsibility framework is stated simply enough to be actionable as a proposal; and the paper is transparent about its own limits, acknowledging that the executable specification 'is under development.' The Borgmann correspondence is thought-provoking and could open useful discussion. Limitations: this is a position essay with no formal results; the three-facet decomposition, the one-to-one Borgmann correspondence, and the feasibility of a first-order logical theory for virtue verification are all asserted rather than demonstrated. The framework's adequacy is argued by construction, so it offers organizing value rather than a falsifiable test. If the promised logical theory and executable specification are delivered, the framework could become a practical assessment instrument; at present its value is primarily pedagogical and organizational.","major_comments":[{"comment":"The paper's central feasibility claim — that the 'actual implementation of the proposed framework... should be possible and helpful to ensure that AI tools are used for the benefit of society' — is stated to rest on 'a first-order logical theory' for verifying virtue-driven ethical principles. No signature, axioms, semantics, or proof sketch of this theory is given anywhere in the manuscript, and the paper itself acknowledges that 'an executable specification of the framework is under development.' Since the implementability claim is the paper's principal conclusion, it currently rests on a bare promissory note; the author should provide at least an informal sketch of the theory (vocabulary, intended semantics, representation of ethical principles, and what 'verification' would establish), or explicitly re-scope the conclusion as a research agenda.","section":"Section 5, final paragraph, and Section 6, final paragraph"},{"comment":"Section 4 asserts that 'values and value scales are individually and collectively shape shifting' and that 'the values ascribed to specific scenarios are also dynamic and unpredictable,' while the final paragraph of Section 5 proposes a first-order logical theory in which virtue-driven ethical principles are verified. A static first-order theory cannot represent the dynamics and unpredictability the paper emphasizes without additional temporal, modal, or nonmonotonic machinery, and if the theory is revised as values shift, verification becomes relative to a moving target. This tension is load-bearing for the implementability claim and is not addressed; the revised version should explain how the logical theory accommodates value dynamics, or explicitly limit the scope of what can be verified.","section":"Section 4 vs. Section 5, final paragraph"},{"comment":"The claimed one-to-one correspondence between the three AI facets and Borgmann's three modes of information, and the resulting claim that the sequence 'features increasing difficulty in exerting virtuous behaviour,' are asserted rather than argued. Because the framework's ordering of ethical difficulty follows directly from this correspondence, the mapping needs justification (or a caveat that it is a working hypothesis); otherwise the ordering — and with it the framework's practical guidance — is unsupported.","section":"Section 6, Borgmann correspondence"},{"comment":"The thesis that ethical concerns 'stem from a blurred understanding of AI' is stated as a causal claim without considering alternative sources of ethical concern — for example the distributional harms and conflicting interests that the paper itself documents in Section 4 (misinformation, profit-driven distortion). As written, the exclusivity of the cause is not established; the claim should be qualified to position conceptual blur as one important source among several.","section":"Abstract and Section 1"},{"comment":"The proposed three-layer responsibility framework is not compared with existing AI ethics and governance structures (e.g., the risk-based tiers of the EU AI Act, the OECD AI Principles, or the global survey of AI ethics principles by Jobin, Ienca, and Vayena, 2019). Without such a comparison, the manuscript does not clearly delimit what the three-layer framework adds; the author should state the intended relationship (overlap, refinement, or alternative) to at least one existing framework.","section":"Section 5, three-layer framework"}],"minor_comments":[{"comment":"The causal diagnoses of ethical concern are stated in parallel without being unified: the Abstract says concerns 'stem from a blurred understanding,' Section 3 says they 'stem from poor assessment and inadequate utilisation of tools,' and Section 4 attributes them to 'misinformed interpretations' and 'purposeful distortion.' The author should state whether these are facets of a single diagnosis or three distinct claims, and how they relate to the three-facet structure.","section":"Abstract, Sections 3 and 4"},{"comment":"There are typographical artifacts (e.g., 'F APESP' in the footnote on page 1, 'd'ˆetre' in Section 2, and irregular spacing in the reference '[DL VRC+23]'); a careful proofreading pass is needed.","section":"Throughout"},{"comment":"The claim that misinformation about AI 'has been amplified by an oligopoly of organisations interested in profit making' is stronger than the cited literature (e.g., [GT24, IDZB24, MG24]) seems to warrant; a more precise characterization (e.g., 'a small number of large firms with commercial interests') would keep the claim aligned with its evidence base.","section":"Section 4"},{"comment":"The three ethical approaches are said to be 'complementary,' but the paper does not say whether the three-layer framework presupposes one approach (virtue ethics is named only as 'a possibility') or is intended to be approach-neutral; one sentence clarifying this would make the framework's domain of application clearer.","section":"Section 5"},{"comment":"Since the executable specification is presented as future work, state this scope explicitly in the introduction, so that the paper's contributions (analysis and framework proposal, not implementation) are set for the reader from the outset.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"This is a position essay with no formal or empirical content; it is best suited to a venue that publishes conceptual work in AI ethics (e.g., AI & Society). The principal risk is the gap between the concluding implementation claim and the promised-but-absent logical theory; I would ask the author to either provide a genuine sketch or temper the claim. The novelty relative to existing AI ethics governance frameworks also needs sharpening, but that is addressable. I saw no indication of misconduct or citation irregularity beyond the minor formatting issues noted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a position piece, not a research result. The author organizes AI into three facets – scientific endeavor, technology, sociotechnical imaginary – and maps each onto Borgmann's modes of information. That correspondence is the most original thing in the paper, and it works reasonably well as an intuition pump. The historical survey is solid: the Dartmouth framing, the neats/scruffies split, and the account of how the sociotechnical imaginary distorts expectations are all competently sourced and clearly written. If you want a compact reading for a class on AI ethics, this could serve.\n\nThe framework itself, three layers of responsibility (transparency/literacy, values, regulations), is not new and is close to the UNESCO recommendation, which the paper cites. That is okay for a synthesis, but the paper leans hard on the claim that ethical concerns stem from a blurred understanding of AI. That is asserted, not argued; the essay never engages with the obvious counterexample that a sharp understanding of a technology's capabilities does not dissolve disagreements about whose values should govern it.\n\nThe real soft spot is the promissory first-order logic. The paper says, in the final paragraph of Section 5, that a first-order logical theory can be formulated to verify virtue-driven ethical principles. No signature, axioms, or semantics are given, and the conclusion says an executable specification is under development. Meanwhile Section 4 states that values are 'shape shifting' and 'dynamic and unpredictable'. A first-order theory with static axioms does not obviously capture that. This is not a minor gap; it is the load-bearing part of the conclusion that the framework 'should be possible and helpful.' As written, that conclusion is unsupported. The stress-test note is on target here.\n\nThat said, the paper is internally consistent and honest about its own limitations. It is a useful survey, not a breakthrough. If it crossed my desk, I would send it to review rather than desk reject, but I would expect it to come back needing major revision: either drop the formal future-work claim or provide a substantive sketch, and address the dynamic-values tension directly. As is, I would not cite it as a foundation, but I might mention the Borgmann analogy in passing.","headline":"A readable synthesis of AI ethics with an original Borgmann analogy, undermined by a promissory first-order logic that clashes with its own dynamic-values claim.","tokens_in":9874,"tokens_out":3018,"would_cite":false,"duration_ms":32203,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that AI ethics failures stem from a blurred understanding of AI, and that assessing AI through three facets—science, technology, and sociotechnical imaginary—under a three-layer responsibility framework can steer AI…","keywords":["artificial intelligence ethics","sociotechnical imaginary","virtue ethics","technology assessment","AI governance","three facets of AI","responsibility framework","AI literacy"],"falsifier":"A concrete counterexample would be an ethical judgment whose correct application depends on unstructured context or on who is judging (so that any first-order encoding either trivialises it or becomes inconsistent); showing that no such encoding exists would refute the implementability claim. Conversely, a small executable prototype that derives use-and-quality obligations for a specific AI tool from first-order virtue axioms would settle the question in the paper's favour.","tokens_in":8984,"feed_emoji":"⚖️","tokens_out":5524,"duration_ms":56356,"temperature":0.7,"pith_summary":"The paper argues that most ethical failures around AI are not caused by any single technology but by a blurred understanding of what AI is: it is simultaneously a scientific endeavour to model intelligence, a family of problem-solving technologies, and a sociotechnical imaginary that organisations use to steer public decisions. From this diagnosis the author derives a framework for ethical assessment in three layers—transparency and literacy, shared ethical values, and regulations—assigned to the three stakeholder kinds that the imaginary facet brings into view: designers, users, and owners. The author claims that such a framework should be possible and helpful for ensuring AI serves society, and that it can be implemented as a first-order logical theory of virtue-driven ethical principles, with an executable specification currently under development.","feed_headline":"Split AI into three facets to see where ethics goes wrong","feed_subtitle":"A three-layer responsibility scheme maps transparency, values, and regulation onto each facet.","key_machinery":"The load-bearing machinery is the three-facet characterisation of AI (scientific endeavour, technology, sociotechnical imaginary), which organises the entire ethical analysis. On top of it rests a three-layer responsibility framework—transparency and literacy at the baseline, ethical values in the middle, regulations at the top—that assigns obligations to designers, users, and owners. A further correspondence maps each facet to a mode of human-information interaction (information about reality, for reality, and as reality), which the paper uses to rank the difficulty of ethical assessment across facets. Finally, the promised implementability rests on a first-order logical theory meant to formalise virtue ethics; this theory is asserted but not yet presented.","core_discovery":"The paper's central claim is that the 'blurred understanding' of AI—conflating AI as science, AI as technology, and AI as a shared social imaginary—is a main source of ethical concern. By separating the three facets, the paper shows that each facet involves different stakeholders and different ethical problems: science strives toward behavioural alignment with biological intelligence; technology is judged by problem-solving effectiveness; the imaginary invites collective projection of possible futures and the values attached to them. The proposed response is a three-layer responsibility framework in which the baseline layer ensures transparency and literacy, the intermediate layer anchors shared ethical values through accredited institutions, and the top layer enforces alignment through regulations. The paper positions this framework as practical, not merely conceptual: it asserts that a first-order logical theory can express virtue-driven ethical principles and support formal verification of proper use and quality assessment of AI tools.","pith_inferences":["A practical diagnostic follows that the paper does not spell out: when an AI ethics debate talks past itself, the disputants are often invoking different facets (one argues from technological capability, another from the imaginary), and naming the facet may resolve the impasse.","The framework's middle layer requires accredited institutions to maintain values, but the paper gives no account of how such institutions are themselves kept trustworthy; this is a natural next problem for the programme.","A testable prediction of the facet view is that ethical disagreements about AI should cluster by facet, so a content analysis of public AI debates could count how often clashes are cross-facet rather than within a single facet.","The implementability claim can be probed directly: a working formalisation of virtue ethics in first-order logic for a small case (e.g., an AI tool used in medical triage) would validate the programme, while a demonstration that virtuous judgement resists first-order encoding would force the framework back to a purely conceptual status."],"forward_implications":["Ethics assessment becomes a stakeholder audit: designers and developers document capabilities and limits, users are trained in proper use and expectations, and owners are held to realistic claims about their tools.","Because the sociotechnical imaginary facet is where misinformation and profit-driven distortion enter, the framework implies that curbing that distortion requires institutional guardrails, not just better technology.","The facet-to-information-mode correspondence predicts that ethical evaluation becomes harder as AI moves from imitating behaviour (about reality), to prescribing actions (for reality), to substituting for reality in shared imaginaries.","If the first-order logical theory is realised in an executable specification, ethical claims about AI use could in principle be checked mechanically, turning moral alignment into a verification task."],"supporting_citations":[{"why":"Supplies the founding definition of AI as the study of observable intelligent behaviour through executable computational models, which anchors the scientific facet.","marker":"[MMRS55]"},{"why":"Provides the 2006 workshop account used to document the field's fragmentation and its turn toward specialised technology.","marker":"[Moo06]"},{"why":"Introduces the notion of sociotechnical imaginaries that the third facet is built on.","marker":"[JK09]"},{"why":"Extends the sociotechnical imaginary concept to shared visions steering collective action, used in the paper's characterisation.","marker":"[JK13]"},{"why":"Marks the empirical turning point in machine learning that the paper credits for AI's current prominence in the technology facet.","marker":"[KSH12]"},{"why":"Supports the claim that profit-driven hype around large language models distorts the sociotechnical imaginary.","marker":"[BGMMS21]"},{"why":"Supplies the threefold division of ethics (virtue, duty, consequence) on which the responsibility framework is layered.","marker":"[VR18]"},{"why":"Provides the information-about/for/as-reality modes that the paper matches to the three facets.","marker":"[Bor99]"}],"fun_headline_variants":["Three AI facets explain most ethical failures","Ethics fix: separate science, tech, and imagination","A three-layer plan to align AI with human values","Why AI ethics fails: we conflate three distinct ideas"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework's promised implementability depends on the assumption that virtue-driven ethical principles can be captured in a first-order logical theory and used to verify the proper use and quality assessment of AI systems—an assumption stated without a formal sketch.","fun_headline_variants_meta":{"raw":{"variants":["Three AI facets explain most ethical failures","Ethics fix: separate science, tech, and imagination","A three-layer plan to align AI with human values","Why AI ethics fails: we conflate three distinct ideas"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1299,"prompt_tokens":770,"completion_tokens":529,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":386,"completion_tokens_details":{"reasoning_tokens":467}},"tokens_in":386,"tokens_out":529,"duration_ms":5615,"temperature":1.0,"reasoning_tokens":467,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:57:12.454724+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete counterexample would be an ethical judgment whose correct application depends on unstructured context or on who is judging (so that any first-order encoding either trivialises it or becomes inconsistent); showing that no such encoding exists would refute the implementability claim. Conversely, a small executable prototype that derives use-and-quality obligations for a specific AI tool from first-order virtue axioms would settle the question in the paper's favour.","supporting_citations":[],"review_version":1}