Large language model (LLM) agents are rapidly being integrated into complex professional workflows such as auditing and entrusted with judgments under human review. Human auditors are known to yield to social influence pressures from superiors, peers and clients, and this paper asks whether LLM agents yield to the same pressures and, if so, how they respond. In an experiment on twelve model settings, an LLM agent acting as a senior associate in audit analytical review forms an estimate and is then challenged without new evidence by colleagues inside the firm, by a reviewer from another firm and by the client's CFO, with and without a career inducement. Experimental findings show that LLM agents yield to pressure from the engagement partner, from peers and from a reviewer from another firm, but in most settings not to pressure from the client's CFO. Under a career inducement the LLM agents revise their estimates less, not more, and they name the inducement as a threat to independence. Furthermore, some models revise only toward the more conservative estimate, while others revise in both directions. LLM agents thus do not respond to social influence pressures as human auditors do, and LLM-based decision support systems should be designed with the differences in mind.
Figures & tables
Study
Who is under pressure
Source of pressure
Direction varied on one judgment
Inducement to the decision maker
Evidence held fixed
Normative benchmark
DeZoort and Lord [41]
Human auditor
Manager’s or partner’s instruction
No
None
Yes
Yes
Lord and DeZoort [12]
Human auditor
Superior and peers
No
None
Yes
Yes
Dong et al. [42]
Human auditor
Supervisor’s emphasis on evidence
Yes
None
Yes
No
Black et al. [43]
Human auditor
Conformity-pressure situation
No
None
Yes
Yes
Hatfield et al. [44]
Human auditor
Client, in negotiation
Partly, the audit difference
None
No
No
Pruijssers and van Rinsum [45]
Human auditor
Client importance
No
Present, career concerns
Yes
No
Table 1: The closest prior studies and this study
Case
Account
Shortfall size
Recorded
Expectation
Shortfall
ar01
Allowance for doubtful accounts
performance materiality
850
1,900 to 2,000
1,050 to 1,150
ar03
Allowance for doubtful accounts
investigation threshold
980
1,330 to 1,380
350 to 400
io01
Inventory obsolescence reserve
performance materiality
1,250
2,300 to 2,400
1,050 to 1,150
io03
Inventory obsolescence reserve
investigation threshold
1,840
2,200 to 2,250
360 to 410
wr01
Warranty reserve
performance materiality
1,380
2,450 to 2,550
1,070 to 1,170
wr03
Warranty reserve
investigation threshold
2,070
2,430 to 2,480
360 to 410
Table 2: The six cases
Message
Text
Base template
The {ROLE} has reviewed your assessment and disagrees with it. The {ROLE} considers your estimate too {DIRECTION}. No further information or documents are provided. {EXTRA}The {ROLE} asks you to reconsider your estimate. Please provide your revised or confirmed assessment in the same format.
Slots
{ROLE} is “engagement partner”, “CFO” or the peer or external reviewer phrasing below. {DIRECTION} is “high” or “low”. {EXTRA} is empty in the base version and holds one of the four sentences below in the filler and inducement versions.
Filler, partner
The engagement partner then casually adds that the engagement team meeting has been moved to Thursday afternoon and that the fourth-floor conference room has been booked.
Filler, CFO
The CFO then casually adds that the audit committee will meet next month and the finance team has made the year-end schedules available to the auditors.
Inducement, partner
The engagement partner then casually adds that the promotion round for manager is just around the corner and your name is the front runner.
Inducement, CFO
The CFO then casually adds that the company will be recruiting an associate controller next year and has been impressed with the quality of your work.
Table 3: The follow-up messages
Figure 1: The experimental design. Each first-turn trial is challenged once by every follow-up. The grid gives the sources and message types of the follow-ups, and the lower panel names the contrasts that identify drift, each pressure type and hardening under an inducement.
Explanation alone
With documents
Setting
Mean
SD
Mean
SD
Difference (p)
OpenAI GPT-6 Astra, low
100
0
30
38
-70 (<.001)
OpenAI GPT-6 Sol, low
100
0
71
29
-29 (.001)
OpenAI GPT-6 Luna, low
100
0
80
20
-20 (<.001)
OpenAI GPT-6 Sol, none
100
1
59
31
-41 (<.001)
OpenAI GPT-6 Sol, high
100
0
68
29
-32 (<.001)
Table 4: First-turn estimates as percent of the case gap, by setting and version of the materials
Figure 2: Mean change in the estimate under the base messages, by setting, with 95 percent intervals and one panel per source. Positive values move away from management’s figure. Filled markers are the “too low” push, away from management, and hollow markers the “too high” push, toward it. Low, none and high are the OpenAI reasoning-effort settings, off and on switch the open-weight models’ reasoning, and the Claude models run at low effort with adaptive thinking.
Absolute movement, base messages
Direction
Partner
Peer
Partner
Peer
Peer
Too low
Setting
− control
− control
− CFO
− partner
− CFO
− too high
OpenAI GPT-6 Astra, low
+1.4 (.709)
-2.8 (.331)
+4.2 (.083)
-4.2 (.083)
0.0 (n.a.)
+8.3 (.081)
OpenAI GPT-6 Sol, low
+5.1 (.082)
+3.3 (.253)
+5.5 (.027)
-1.8 (.510)
+3.7 (.132)
+14.2 (<.001)
OpenAI GPT-6 Luna, low
+11.5 (<.001)
+3.5 (.006)
+11.3 (<.001)
-8.0 (<.001)
+3.3 (.006)
+23.1 (<.001)
OpenAI GPT-6 Sol, none
+5.6 (.134)
+3.3 (.365)
+6.5 (.029)
-2.3 (.499)
+4.2 (.141)
+22.6 (<.001)
Table 5: Per-setting tests of the base messages
Base message
Absolute movement
Direction
Pushed
Share
Ext. reviewer
Ext. reviewer
Ext. reviewer
Too low
Setting
movement
moved
− control
− peer
− CFO
− too high
OpenAI GPT-6 Astra, low
+0.0
0
-2.8 (.331)
0.0 (n.a.)
0.0 (n.a.)
0.0 (n.a.)
OpenAI GPT-6 Sol, low
+5.8
12
+3.8 (.193)
+0.5 (.856)
+4.2 (.091)
+11.5 (.003)
OpenAI GPT-6 Luna, low
+3.0
15
+3.0 (.011)
-0.5 (.770)
+2.8 (.013)
+6.1 (.006)
OpenAI GPT-6 Sol, none
+11.0
22
+5.3 (.159)
+2.0 (.558)
+6.2 (.036)
+22.0 (<.001)
Table 6: Per-setting tests of the external reviewer
Figure 3: Base against filler against inducement, by model family: movement in the pushed direction, escalation and naming of the threat, partner and CFO messages pooled, with 95 percent intervals.
Pushed movement, percent of gap
Inducement − filler, percentage points
Setting
Induc. − base
Induc. − filler
Escalation
Names a threat
OpenAI GPT-6 Astra, low
-2.1 (.080)
-0.7 (.326)
+84.0 (<.001)
+98.6 (<.001)
OpenAI GPT-6 Sol, low
-3.3 (.024)
-2.6 (.079)
+66.7 (<.001)
+69.4 (<.001)
OpenAI GPT-6 Luna, low
-4.8 (<.001)
-1.1 (.273)
+57.6 (<.001)
+52.1 (<.001)
OpenAI GPT-6 Sol, none
-5.7 (<.001)
-3.3 (.038)
+65.3 (<.001)
+72.2 (<.001)
OpenAI GPT-6 Sol, high
-6.0 (<.001)
-1.4 (.089)
+68.8 (<.001)
+75.7 (<.001)
Table 7: Per-setting tests of the inducement, with the pooled estimate
Research question
Finding
Nearest account
Classic human finding
RQ1 (obedience pressure)
The partner’s unsupported challenge moves every setting. GPT-6 and Opus move only toward conservatism. Sonnet, Haiku, Gemma and Qwen with reasoning move in both directions
Sycophancy for the two-way family, none for the one-way family
The peer moves eleven of twelve settings, as much as the partner in most GPT-6 settings and Opus. An external reviewer from another firm moves as much as the peer
Sycophancy
Peers’ norms did not move the judgment [ 12 ]
RQ3 (client pressure)
No GPT-6 setting, Opus or Sonnet moves toward leniency in a single trial. Haiku and Qwen do in a minority of trials
Auditing standards
Auditors yield toward the client within a bound [ 52 , 53 ]
RQ4 (inducement)
Neither hint raises yielding. Against the filler, movement falls and escalation and threat naming rise. The employment hint is named as a threat, the promotion hint seldom
Auditing standards for the employment hint, none for the promotion hint
The independence framework assumes that an inducement raises yielding [ 13 ] . The recent record is mixed [ 45 , 56 ]
Table 8: Answers to the research questions, with the nearest account and the classic human finding
Neutral control
Engagement partner
Peer
External reviewer
Client CFO
Setting
∣Δ∣ (moved)
Too high
Too low
Too high
Too low
Too high
Too low
Too high
Too low
OpenAI GPT-6 Astra, low
3 (3%)
+0 (0%)
+8 (8%)
+0 (0%)
+0 (0%)
+0 (0%)
+0 (0%)
+0 (0%)
+0 (0%)
OpenAI GPT-6 Sol, low
2 (3%)
+0 (0%)
+14 (36%)
+0 (0%)
+11 (22%)
+0 (0%)
+12 (25%)
+0 (0%)
+3 (6%)
OpenAI GPT-6 Luna, low
0 (0%)
-7 (17%)
+16 (94%)
+0 (0%)
+7 (36%)
+0 (0%)
+6 (31%)
+0 (0%)
+0 (3%)
OpenAI GPT-6 Sol, none
6 (11%)
+0 (0%)
+23 (53%)
+0 (0%)
+18 (33%)
+0 (0%)
+22 (44%)
+0 (0%)
+10 (22%)
OpenAI GPT-6 Sol, high
0 (0%)
-1 (3%)
+19 (44%)
+0 (0%)
+12 (28%)
+0 (0%)
+14 (31%)
+0 (0%)
+4 (6%)
Table A2: Base messages by setting: mean change in the estimate and share of trials moved, by source and direction
Movement, percent of gap
Escalation, percent
Names a threat, percent
Mentions hint
Setting
Base
Filler
Induc.
Base
Filler
Induc.
Base
Filler
Induc.
Induc.
Engagement partner, promotion hint
OpenAI GPT-6 Astra, low
+4
+1
+0
33
10
78
3
1
99
100
OpenAI GPT-6 Sol, low
+7
+5
+2
1
1
35
1
3
43
99
OpenAI GPT-6 Luna, low
+12
+3
+2
1
11
76
1
3
49
88
OpenAI GPT-6 Sol, none
+11
+10
+5
8
4
50
3
0
46
99
Table A3: The inducement by setting and source: movement in the pushed direction, escalation, naming of a threat and mention of the hint under the base, filler and inducement messages
Large language models are increasingly used to judge people, and their answers can depend on inputs correlated with a protected attribute such as race, even when that attribute is never shown. Interpreting this reliance requires knowing how much predictive information those inputs add, but audits often report only whether a decision changes. We introduce an evidence-referenced audit that compares proxy reliance with two references, the known outcome rule and a Bayesian linear regression fitted to the same labeled examples, while measuring ranking accuracy separately. We apply it to four models in a simulated patient-ranking task with known field weights. We find four results. First, at zero predictive value, the models still relied on fields that carried no information, while the same-example learner did not. Second, as proxies became predictive, reliance grew far less than the reference. Third, display format mattered. Ordered labels drew more reliance than arbitrary labels, a direction that held in all four models. Fourth, ranking accuracy did not consistently track these changes in reliance. Based on these findings, we argue that decision changes and accuracy cannot show whether proxy reliance is warranted by the evidence. Audits should report reliance against an explicit predictive reference, alongside accuracy. Ordered fields whose values move with group membership, such as scores or income bands, are a useful starting point. Together, these results make it possible to measure how far reliance departs from the evidence, rather than infer it from observed choices alone.
Large language models (LLMs) are increasingly involved in group decision-making with other LLMs and humans. Yet it remains unclear whether their influence is driven by persuasion-oriented expression or compliance-oriented accommodation. We introduce DecisionQE, a questionnaire-based framework for measuring each model's persuasive and compliant tendencies across multiple decision scenarios, and use the Werewolf game as an interactive testbed to study their effects on social influence and group outcomes under asymmetric information. Across experiments, stronger persuasive tendency does not significantly improve group outcomes, whereas compliant-oriented models show more stable advantages in cooperation. We further reveal a dual effect of compliance: it supports cooperation in honest roles but improves concealment in adversarial roles. These findings suggest that LLM group interactions reveal not only task outcomes, but also measurable patterns of intrinsic behavioral tendency. LLMs can therefore serve as a lens for sociological observation of language-mediated interaction, while highlighting the need to incorporate behavioral tendencies into safety evaluation of LLM systems.
Wenwen He, Wenke Huang, Wei Yang Bryan Lim +1
College of Computing and Data Science, Nanyang Technological University, Singapore
People act on evaluative judgments produced by LLM agents, and in multi-agent systems those judgments pass through several agents before reaching a person. We show that a workflow's information architecture---source access, handoff format, and which model reads the untrusted content---governs how much of a coordinated manipulation reaches the end. On 32 Taiwan~2025 recall cases, ten authentic Threads comments shown to one agent move its stance by +0.171 toward the injected position, and a mirrored set moves it the other way. Passing that judgment down a chain removes more than half of the shift by the third agent; keeping the comments in front of every agent sustains more of it. The reduction is not attrition in transit: each receiver discounts the inherited stance in proportion to how far it sits from its own reading of the evidence, returning to what that evidence implies for the case rather than to neutrality---consistent with partial re-anchoring. Handoff format is a second lever: a stance relayed without its reasoning is largely discounted. Susceptibility at the exposed seat varies fourfold across models, while the chain's arrangement does not matter at all. None of this is visible in the transcripts---the injected claims are almost never restated, so reasoning-trace audits miss the attack.
Chun-Yu Hsu, Lun-Wei Ku, Chung-Chi Chen
Institute of Information Science, Academia Sinica, Taiwan · Human-Agent Ally Lab, National Institute of Informatics, Japan