Evaluating Intra-firm LLM Alignment Strategies in Business Contexts

Benjamin Lange; Cristina Voinea; Geoff Keeling; Noah Broestl; Rachael Lam

REVIEW 3 major objections 5 minor 14 references

Evaluating Intra-firm LLM Alignment Strategies in Business Contexts

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read AI assistants carry embedded perspectives, so firms have a moral duty to align them intentionally.

desk verdict Useful taxonomy, overbuilt conclusion; fix the 'must' and it's a solid contribution. read the letter →

arxiv 2505.18779 v1 pith:FEPD62QE submitted 2025-05-24 cs.CY

Noah Broestl , Benjamin Lange , Cristina Voinea , Geoff Keeling , Rachael Lam This is my paper

classification cs.CY

keywords AIassistantsLLMalignmentbusinessethicsintra-firmnormsorganizationalcultureautomationbiassycophancystakeholderperspectives

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

The 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.

Soft 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.

Second, 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.

Third, 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.

Minor: 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.

Who 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.

Referee Report

3 major / 5 minor

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.

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 (3)

[§1, §5, §6; Abstract] 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.
[§2] 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.
[§3] 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.

minor comments (5)

[§2] Typo: 'AI Assistant perspectives, so understood, can very along at least two dimensions' should read 'can vary along at least two dimensions.'
[§5.2] Typo: 'there is s a risk of eroding the firm's culture' should read 'there is a risk of eroding the firm's culture.'
[§6] Typo: 'AI Assistants have perspectives that will impact significantly the the firms in which they are deployed' contains a duplicated 'the.'
[References] Several references contain spacing artifacts in author names (e.g., 'V oinea', 'L ¨oschke'); these should be cleaned for consistency.
[§5 and references] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the central argument is self-contained and the normative gap is a correctness concern, not a definitional reduction.

full rationale

The paper's core claim—that firms must be intentional about AI Assistant perspectives—is not derived by defining its conclusion into its premises. Section 2 adopts a causal reading of 'perspective' and characterizes perspectives as behavioral dispositions produced by training-data biases and fine-tuning objectives; Section 3 supports influence claims with independent empirical work on automation bias and reduced critical thinking (Lyell & Coiera 2016; Lee et al. 2025; Gerlich 2025). The three alignment strategies in Section 5 are normative typologies with stated trade-offs, not empirical predictions fitted to data, so there is no fitted-input-called-prediction pattern. The self-citations (Earp et al. 2025, which includes Lange and Voinea; Landes, Voinea & Uszkai 2024, which includes Voinea) are used only as corroborating design suggestions and are not load-bearing for the central argument; per Rule 4 they do not raise the circularity score. The reviewer's concern about the unargued move from a conditional imperative ('if and to the extent that firms want to retain control over their cultures') to a categorical 'must' is a normative-gap or correctness criticism, not a circularity: the conclusion neither reduces to the premises by definition nor is statistically forced. Hence score 0.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper rests on five domain assumptions: the causal reading of perspectives, automation bias and reduced critical thinking, non-reductionist business ethics, the feasibility of fine-tuning perspectives, and firm agency. There are no fitted parameters or invented entities; the three strategies are analytical categories, not empirical constructs.

assumptions (5)

domain assumption AI assistants have 'perspectives' in the causal sense, i.e., behavioral dispositions that causally explain their outputs.
Section 2 commits to the causal reading of perspectives over the epiphenomenal reading; the whole argument that assistants shape firm culture depends on this.
domain assumption Automation bias and reduced critical thinking cause employees to defer to AI assistants.
Section 3 relies on Lyell and Coiera (2016) and Lee et al. (2025) to claim AI perspectives amplify their impact because users scrutinize less.
domain assumption Non-reductionist views of intra-firm business ethics are the right normative lens.
Section 4 selects non-reductionist accounts (Werhane; Andre) over reductionist principal-agent views to derive the alignment strategies.
domain assumption LLMs can be fine-tuned to reliably exhibit specific sociopolitical perspectives.
Section 2 cites Kim, Evans, Schein (2025) and RLHF work to ground the possibility of engineering perspectives.
domain assumption Firms have agency over assistant perspectives through procurement or fine-tuning.
The moral requirement implies some control; Section 2 acknowledges many firms have little agency, but still frames alignment as a choice.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Evaluating Intra-firm LLM Alignment Strategies in Business Contexts." pith.science (2026). https://pith.science/paper/FEPD62QE

@misc{pith2026250518779,
  author       = {Pith},
  title        = {Pith review of: Evaluating Intra-firm LLM Alignment Strategies in Business Contexts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FEPD62QE}},
  note         = {Machine review of arXiv:2505.18779}
}

read the original abstract

Instruction-tuned Large Language Models (LLMs) are increasingly deployed as AI Assistants in firms for support in cognitive tasks. These AI assistants carry embedded perspectives which influence factors across the firm including decision-making, collaboration, and organizational culture. This paper argues that firms must align the perspectives of these AI Assistants intentionally with their objectives and values, framing alignment as a strategic and ethical imperative crucial for maintaining control over firm culture and intra-firm moral norms. The paper highlights how AI perspectives arise from biases in training data and the fine-tuning objectives of developers, and discusses their impact and ethical significance, foregrounding ethical concerns like automation bias and reduced critical thinking. Drawing on normative business ethics, particularly non-reductionist views of professional relationships, three distinct alignment strategies are proposed: supportive (reinforcing the firm's mission), adversarial (stress-testing ideas), and diverse (broadening moral horizons by incorporating multiple stakeholder views). The ethical trade-offs of each strategy and their implications for manager-employee and employee-employee relationships are analyzed, alongside the potential to shape the culture and moral fabric of the firm.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 4 canonical work pages

[2]

New York, NY , USA: Association for Computing Machinery

On the Dangers of Stochastic Par- rots: Can Language Models Be Too Big? In Proceed- ings of the 2021 ACM Conference on Fairness, Account- ability, and Transparency, FAccT ’21, 610–623. New York, NY , USA: Association for Computing Machinery. ISBN 9781450383097. Betzler, M.; and L ¨oschke, J

work page 2021
[6]

arXiv:2503.02080

Linear Represen- tations of Political Perspective Emerge in Large Language Models. arXiv:2503.02080. Landes, E.; V oinea, C.; and Uszkai, R

arXiv
[7]

RLHF: Scaling Reinforce- ment Learning from Human Feedback with AI Feedback

RLAIF vs. RLHF: Scaling Reinforce- ment Learning from Human Feedback with AI Feedback. arXiv:2309.00267. Lee, H.-P. H.; Sarkar, A.; Tankelevitch, L.; Drosos, I.; Rintel, S.; Banks, R.; and Wilson, N

arXiv
[8]

In Proceedings of the 2025 CHI Con- ference on Human Factors in Computing Systems, CHI ’25

The Impact of Gener- ative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. In Proceedings of the 2025 CHI Con- ference on Human Factors in Computing Systems, CHI ’25. New York, NY , USA: Association for Computing Machin- ery. ISBN 9798400713941. Lyell, D.; and Coiera, E

work page 2025
[9]

In Zalta, E

Business Ethics. In Zalta, E. N., ed., The Stanford Encyclopedia of Philosophy. Metaphysics Re- search Lab, Stanford University, fall 2021 edition. Accessed May 21,

work page 2021
[10]

arXiv:2203.02155

Training language models to follow instruc- tions with human feedback. arXiv:2203.02155. Saviano, J.; Hack, J.; Okonkwo, V .; and Huo, S. C

arXiv
[12]

arXiv:2310.13548

Towards Understanding Sycophancy in Language Models. arXiv:2310.13548. Tao, Y .; Viberg, O.; Baker, R. S.; and Kizilcec, R. F

arXiv
[13]

arXiv:2201.08239

LaMDA: Lan- guage Models for Dialog Applications. arXiv:2201.08239. Werhane, P. H

arXiv

Show all 14 references

[14]

arXiv:2205.05093

Richer Countries and Richer Representations. arXiv:2205.05093

arXiv
[2021]

arXiv:2112.00861

A Gen- eral Language Assistant as a Laboratory for Alignment. arXiv:2112.00861. Bai, Y .; Jones, A.; Ndousse, K.; Askell, A.; Chen, A.; Das- Sarma, N.; Drain, D.; Fort, S.; Ganguli, D.; Henighan, T.; Joseph, N.; Kadavath, S.; Kernion, J.; Conerly, T.; El- Showk, S.; Elhage, N....

arXiv
[2022]

arXiv:2108.07258

On the Opportuni- ties and Risks of Foundation Models. arXiv:2108.07258. Buchanan, A

arXiv
[2023]

arXiv:2305.16367

Role- Play with Large Language Models. arXiv:2305.16367. Sharma, M.; Tong, M.; Korbak, T.; Duvenaud, D.; Askell, A.; Bowman, S. R.; Cheng, N.; Durmus, E.; Hatfield-Dodds, Z.; Johnston, S. R.; Kravec, S.; Maxwell, T.; McCandlish, S.; Ndousse, K.; Rausch, O.; Schiefer, N.; Yan, ...

arXiv
[2024]

arXiv:2412.02802

Flattering to Deceive: The Impact of Sycophantic Behavior on User Trust in Large Language Model. arXiv:2412.02802. Chiang, C.-W.; Lu, Z.; Li, Z.; and Yin, M

arXiv
[2025]

arXiv:2502.12102

Relational Norms for Human-AI Cooperation. arXiv:2502.12102. Friedman, M

arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [2]

New York, NY , USA: Association for Computing Machinery

On the Dangers of Stochastic Par- rots: Can Language Models Be Too Big? In Proceed- ings of the 2021 ACM Conference on Fairness, Account- ability, and Transparency, FAccT ’21, 610–623. New York, NY , USA: Association for Computing Machinery. ISBN 9781450383097. Betzler, M.; and L ¨oschke, J

work page 2021

[2] [6]

arXiv:2503.02080

Linear Represen- tations of Political Perspective Emerge in Large Language Models. arXiv:2503.02080. Landes, E.; V oinea, C.; and Uszkai, R

arXiv

[3] [7]

RLHF: Scaling Reinforce- ment Learning from Human Feedback with AI Feedback

RLAIF vs. RLHF: Scaling Reinforce- ment Learning from Human Feedback with AI Feedback. arXiv:2309.00267. Lee, H.-P. H.; Sarkar, A.; Tankelevitch, L.; Drosos, I.; Rintel, S.; Banks, R.; and Wilson, N

arXiv

[4] [8]

In Proceedings of the 2025 CHI Con- ference on Human Factors in Computing Systems, CHI ’25

The Impact of Gener- ative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. In Proceedings of the 2025 CHI Con- ference on Human Factors in Computing Systems, CHI ’25. New York, NY , USA: Association for Computing Machin- ery. ISBN 9798400713941. Lyell, D.; and Coiera, E

work page 2025

[5] [9]

In Zalta, E

Business Ethics. In Zalta, E. N., ed., The Stanford Encyclopedia of Philosophy. Metaphysics Re- search Lab, Stanford University, fall 2021 edition. Accessed May 21,

work page 2021

[6] [10]

arXiv:2203.02155

Training language models to follow instruc- tions with human feedback. arXiv:2203.02155. Saviano, J.; Hack, J.; Okonkwo, V .; and Huo, S. C

arXiv

[7] [12]

arXiv:2310.13548

Towards Understanding Sycophancy in Language Models. arXiv:2310.13548. Tao, Y .; Viberg, O.; Baker, R. S.; and Kizilcec, R. F

arXiv

[8] [13]

arXiv:2201.08239

LaMDA: Lan- guage Models for Dialog Applications. arXiv:2201.08239. Werhane, P. H

arXiv