{"id":"36cd2c8e-8be0-463c-a266-e51532061fcb","arxiv_id":"2505.18779","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Firms should intentionally align AI assistants' embedded perspectives using supportive, adversarial, or diverse strategies to protect workplace culture and moral norms.","lead":"This paper argues that AI assistants used inside companies carry built-in perspectives that can change decisions and workplace culture. It proposes three ways firms can deliberately align those perspectives and analyzes the ethical trade-offs of each approach.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 'must' is conditional: the paper's explicit justification is 'if and to the extent that firms want to retain control over their cultures,' and no argument is given for an unconditional duty to want such control.","rationale":"The paper makes a useful conceptual contribution by mapping three alignment strategies onto intra-firm ethical relationships, and the reader's CONDITIONAL verdict is reasonable. However, the single most load-bearing gap is not the metaphysics of perspectives. The causal reading in §2 is explicitly chosen, but the normative strategies are defined behaviorally—fine-tuning on company data, case studies, and examples—so the arguments could be restated under an epiphenomenal reading without losing their force. The more serious gap is that the paper's own stated justification for the moral imperative is explicitly conditional (§1), while the abstract and conclusion assert an unconditional 'must.' The paper never supplies the normative premise that firms have a duty to want control over their culture, nor does it address firms that consciously reject such control. This is an internal argumentative gap, not a dispute with external consensus. The concrete test would expose the hidden premise by formal reconstruction. The reader's weakest assumption is therefore not the one I would focus on; but the recommended verdict is unchanged because the conditional reading already captures the appropriate degree of epistemic caution.","tokens_in":12994,"tokens_out":8422,"duration_ms":87270,"concrete_test":"Reconstruct the argument as a numbered premise-conclusion proof using only claims defended in §§1–5, adding no hidden premises. If the inference to 'firms are morally required to be intentional' requires the unstated premise 'firms have a moral duty to retain control over their culture,' then test the conditional reading by substituting a firm that explicitly waives cultural control: if the conclusion no longer follows, the paper's central claim is conditional rather than categorical, and the verdict should remain CONDITIONAL pending an argument for the missing duty.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's conclusion is categorical: firms are morally required to be intentional about AI Assistant perspectives (Abstract, §5, Conclusion). But the only explicit justification is conditional: 'Alignment is a strategic and ethical imperative if and to the extent that firms want to retain control over their cultures' (§1). Sections 3–5 show that assistant perspectives can influence decisions and that each alignment strategy has trade-offs, but they never defend the missing premise that every firm has an unconditional moral duty to want control over its culture, or to maintain the specific intra-firm norms invoked. Without that premise, the central claim does not follow: a firm that deliberately delegates normative direction to employees or accepts cultural drift is not shown to be morally deficient. The reader's causal-reading concern (§2) is secondary; even on an epiphenomenal reading, output distributions are manageable objects of alignment, so the normative argument can be reconstructed without causal metaphysics. The load-bearing gap is the unargued move from descriptive influence to an unconditional moral requirement.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that instruction-tuned LLMs deployed as general-purpose AI Assistants in firms carry embedded 'perspectives'—behavioral dispositions reflecting social, political, ethical, or cultural biases—that influence decision-making, collaboration, and organizational culture. The authors contend that firms are morally required to align these perspectives intentionally with their objectives and values, and they propose three alignment strategies: supportive (reinforcing the firm's mission), adversarial (stress-testing ideas), and diverse (broadening moral horizons). Drawing on non-reductionist views of intra-firm business ethics, the paper analyzes the ethical trade-offs of each strategy for manager-employee and employee-employee relationships.","tokens_in":13126,"tokens_out":2484,"duration_ms":17710,"significance":"This is a timely and practically relevant contribution to the nascent literature on LLM deployment inside firms. The paper moves beyond individual-level productivity concerns to address organizational culture and intra-firm moral norms, and it offers a clear, usable taxonomy of alignment strategies—supportive, adversarial, and diverse—with a balanced treatment of their ethical trade-offs. The explicit engagement with normative business ethics, particularly non-reductionist accounts of role-specific duties, provides a principled grounding that is often missing in AI-governance discussions. If the normative gap identified below is addressed, the paper could serve as a useful framework for both scholars and practitioners.","major_comments":[{"comment":"The paper's central normative claim is categorical—'firms must be intentional' (Abstract), 'firms must be intentional about aligning AI Assistants' (§5), 'leaders and decision makers in firms must be intentional' (§6)—but the only explicit justification offered is conditional: 'Alignment is a strategic and ethical imperative if and to the extent that firms want to retain control over their cultures' (§1). The paper never argues that every firm has an unconditional moral duty to retain control over its culture, nor does it address the possibility that some firms might legitimately delegate normative direction to employees or accept cultural drift. Without that missing premise, the categorical conclusion does not follow; the argument as stated supports only a conditional imperative for firms that already value such control.","section":"§1, §5, §6; Abstract"},{"comment":"The paper introduces two readings of AI assistant perspectives—causal and epiphenomenal—and then simply asserts: 'Here we interpret perspectives in line with the causal reading.' The argument that assistants have a stable, causally efficacious 'perspective' that shapes firm culture depends on this choice, yet no defense is given against the epiphenomenal reading, where the perspective is merely an interpretive overlay on outputs. The authors should either justify the causal reading or, more promisingly, show that the normative argument can be reconstructed on the epiphenomenal reading, since even on that reading the distribution of outputs is a manageable object of alignment.","section":"§2"},{"comment":"The empirical basis for the claim that AI Assistant perspectives will have a significant, amplified impact on firm culture rests on research on automation bias (Lyell and Coiera 2016) and reduced critical thinking (Lee et al. 2025; Gerlich 2025). These studies are largely self-report or correlational, and the paper does not address effect sizes, generalizability to organizational settings, or the possibility that firms might develop countervailing practices. This matters because the strength of the normative imperative scales with the strength of these empirical claims; as written, the inference from 'users sometimes over-rely' to 'AI assistants will reshape the moral fabric of the firm' is underevidenced.","section":"§3"}],"minor_comments":[{"comment":"Typo: 'AI Assistant perspectives, so understood, can very along at least two dimensions' should read 'can vary along at least two dimensions.'","section":"§2"},{"comment":"Typo: 'there is s a risk of eroding the firm's culture' should read 'there is a risk of eroding the firm's culture.'","section":"§5.2"},{"comment":"Typo: 'AI Assistants have perspectives that will impact significantly the the firms in which they are deployed' contains a duplicated 'the.'","section":"§6"},{"comment":"Several references contain spacing artifacts in author names (e.g., 'V oinea', 'L ¨oschke'); these should be cleaned for consistency.","section":"References"},{"comment":"The paper cites work co-authored by its own authors (Earp et al. 2025; Landes, Voinea, and Uszkai 2024) in support of the alignment strategies; this is not problematic per se, but the authors should ensure that these citations are not doing load-bearing work without independent support, and they may wish to disclose the self-citation more prominently.","section":"§5 and references"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth reading for the taxonomy. The three strategies—supportive, adversarial, diverse—are individually familiar, but I don't know another paper that derives them from non-reductionist business ethics and works through manager-employee and employee-employee relationships for each. The examples are concrete enough to make the trade-offs tangible. That is a real contribution, and it gives practitioners and ethicists a shared vocabulary.\n\nThe paper is honest about trade-offs and includes an ethical and adverse impact statement that names risks of misuse. Citation pattern is mostly fine; the two self-citations (Earp et al., Landes et al.) support peripheral points, not the central claim.\n\nSoft spots, in order. First, the stress-test is right: the categorical 'firms are morally required' is not supported. The only justification offered is conditional: 'if and to the extent that firms want to retain control over their cultures' (§1). Sections 3–5 show influence and trade-offs, but nothing shows every firm has an unconditional duty to want that control, or to preserve the particular norms invoked. The conclusion overreaches. That is fixable: reframe the contribution as a conditional framework plus an argument about why control over culture is typically worth wanting, or defend an unconditional duty head-on.\n\nSecond, the causal reading of 'perspectives' (§2) is an unnecessary commitment. The stress-test is right that even on an epiphenomenal reading, output distributions are objects of alignment; the normative argument does not need the metaphysics. As written it invites a distraction about whether LLMs have causes in the relevant sense. I would advise dropping or deflating that commitment.\n\nThird, the empirical premises—automation bias, reduced critical thinking—are contested and cited from a small number of studies. The paper says 'evidence suggests' and 'plausible,' which is appropriate for a conceptual piece, but the moral conclusion leans on them more than the evidence can carry. A more careful calibration would help.\n\nMinor: the 'neutral' alignment strategy is set aside in a footnote; a sentence on why none of the three strategies reduces to neutrality would tighten the taxonomy.\n\nWho is it for: AI ethics and business ethics readers, plus practitioners choosing among deployment postures. It deserves a serious referee, with a request to fix the scope of the normative claim. I would bring it to reading group and would cite it in work on assistant alignment in organizations.","headline":"Useful taxonomy, overbuilt conclusion; fix the 'must' and it's a solid contribution.","tokens_in":13667,"tokens_out":2704,"would_cite":true,"duration_ms":24631,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"AI assistants carry embedded perspectives, so firms have a moral duty to align them intentionally.","keywords":["AI assistants","LLM alignment","business ethics","intra-firm norms","organizational culture","automation bias","sycophancy","stakeholder perspectives"],"falsifier":"Deploy matched teams in a firm on equivalent cognitive tasks with four assistant configurations, supportive, adversarial, diverse, and default out-of-the-box, over several months, and measure decision quality, dissent frequency, critical-thinking scores, and cultural norms; if no systematic differences emerge, or if the same model can be prompted to exhibit any stable perspective on identical queries, the claim that assistants carry a causally efficacious perspective requiring intentional alignment would be falsified.","tokens_in":12761,"feed_emoji":"🤖","tokens_out":4852,"duration_ms":38595,"temperature":0.7,"pith_summary":"This paper argues that AI assistants used inside firms are not neutral tools; they carry embedded perspectives, stable dispositions to answer in line with particular social, political, or ethical biases, that seep into decision-making, collaboration, and firm culture. Because employees tend to trust automated outputs and engage less critically with them, these perspectives can shape organizational norms largely unnoticed. The paper claims firms therefore have a moral obligation to align assistant perspectives intentionally with their values, and proposes three distinct strategies, supportive, adversarial, and diverse, each with characteristic ethical trade-offs for manager-employee and employee-employee relationships. A sympathetic reader would take the central insight to be that choosing an alignment strategy is itself a moral decision about what kind of workplace culture the firm wants.","feed_headline":"Firms must choose their AI assistant's perspective deliberately","feed_subtitle":"Unchecked out-of-the-box assistants silently shape workplace culture; supportive, adversarial, or diverse alignment is a moral choice.","key_machinery":"The central object is the concept of an AI assistant's 'perspective,' defined as a set of behavioral dispositions to address queries in line with particular social, political, ethical, or cultural biases, and interpreted on the causal reading rather than the epiphenomenal reading. This concept carries the argument because it makes the assistant an active participant whose stable biases can influence firm culture. A second load-bearing element is the non-reductionist view of intra-firm business ethics, which holds that moral duties within firms are role-specific and sui generis, not merely contractual constraints against moral hazard; this view supplies the normative standard against which the supportive, adversarial, and diverse strategies are assessed.","core_discovery":"The paper's central claim is that instruction-tuned LLMs deployed as AI assistants have perspectives, understood on a causal reading as behavioral dispositions that feature in explanations of their outputs, and that these perspectives are shaped by biases in pre-training data and by developers' fine-tuning objectives such as RLHF. Because automation bias and reduced critical thinking make employees less likely to scrutinize assistant output, these perspectives can quietly shape decisions, relationships, and the moral norms of the firm. The paper argues that firms are therefore morally required to be intentional about the perspective their assistant embodies, and it offers three alignment strategies: supportive, which reinforces the firm's mission; adversarial, which stress-tests ideas within the space of the firm's values; and diverse, which broadens the moral horizon by presenting multiple stakeholder perspectives. Drawing on non-reductionist business ethics, the paper evaluates how each strategy reshapes role-specific duties between managers and employees and among colleagues, concluding that no single strategy fits every firm.","pith_inferences":["If the paper's causal reading is right, the same alignment obligation plausibly extends to public agencies, schools, and other organizations that deploy assistants at scale, not just firms.","The three strategies could be sequenced or combined, adversarial during strategy formulation, supportive during execution, diverse during stakeholder review, even though the paper treats them as distinct options.","The framework predicts observable differences: teams with adversarial or diverse assistants should show more divergent thinking and fewer groupthink indicators than teams with supportive or default assistants, a prediction a field experiment could test.","The paper's own adverse impact statement implies a further risk: alignment strategies can be co-opted to justify pre-existing agendas, so intentionality alone is not sufficient and governance and transparency are also needed."],"forward_implications":["Firms that treat an out-of-the-box assistant as neutral are in fact adopting an unexamined perspective, and that default will shape their culture.","Supportive alignment can strengthen mission and collegiality, but at the risk of sycophancy, reinforced power asymmetries, and suppressed dissent.","Adversarial alignment can counter groupthink and improve decision quality, but may erode trust, managerial confidence, and moral deliberation skills.","Diverse alignment broadens the ethical landscape and supports pluralism, but can produce decision paralysis and dilute accountability.","Choosing a strategy is a choice about which intra-firm moral norms to cultivate; there is no one-size-fits-all answer."],"supporting_citations":[{"why":"Supplies the causal versus epiphenomenal distinction of 'perspective' that the paper explicitly commits to.","marker":"Shanahan, McDonell, and Reynolds 2023"},{"why":"Establishes that biases in pre-training data shape the sociopolitical dispositions of language models.","marker":"Bender et al. 2021"},{"why":"Shows how fine-tuning with human feedback encodes developer value objectives into assistants.","marker":"Ouyang et al. 2022"},{"why":"Documents automation bias, the mechanism by which assistant perspectives escape scrutiny.","marker":"Lyell and Coiera 2016"},{"why":"Provides evidence that generative AI reduces critical thinking and cognitive effort in knowledge workers.","marker":"Lee et al. 2025"},{"why":"Defines firm culture as shared values and norms, the target that alignment seeks to control.","marker":"Kaptein 2008"},{"why":"Represents the reductionist principal-agent view of intra-firm duties that the paper contrasts with its non-reductionist approach.","marker":"Heath 2023"},{"why":"Grounds the non-reductionist claim that intra-firm moral norms are sui generis rather than contractual.","marker":"Werhane 1985"},{"why":"Characterizes collegial role responsibilities, grounding the employee-employee analysis.","marker":"Betzler and Löschke 2021"}],"fun_headline_variants":["AI assistants have biases: firms must align them deliberately","Supportive, adversarial, or diverse: three ways to align AI","Unchecked AI assistants shape firm culture silently","Firms that ignore AI assistant perspectives lose control","Align your AI assistant's perspective or lose your firm's culture"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"That an AI assistant has a stable, causally effective perspective, rather than merely being interpreted as having one, so that its influence on firm culture is something the firm can and must manage.","fun_headline_variants_meta":{"raw":{"variants":["AI assistants have biases: firms must align them deliberately","Supportive, adversarial, or diverse: three ways to align AI","Unchecked AI assistants shape firm culture silently","Firms that ignore AI assistant perspectives lose control","Align your AI assistant's perspective or lose your firm's culture"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000382,"raw_usage":{"total_tokens":2016,"prompt_tokens":925,"completion_tokens":1091,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":1013}},"tokens_in":541,"tokens_out":1091,"duration_ms":13901,"temperature":1.0,"reasoning_tokens":1013,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:24:24.954070+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy matched teams in a firm on equivalent cognitive tasks with four assistant configurations, supportive, adversarial, diverse, and default out-of-the-box, over several months, and measure decision quality, dissent frequency, critical-thinking scores, and cultural norms; if no systematic differences emerge, or if the same model can be prompted to exhibit any stable perspective on identical queries, the claim that assistants carry a causally efficacious perspective requiring intentional alignment would be falsified.","supporting_citations":[],"review_version":1}