PaperlayerView exact revision on arXiv

Source-linked AI summary

Same Person, Different Advice: Auditing LLM Financial Advice with Repeated Risk-Tolerance Responses

Soovadeep Bakshi, Apurba Bose, Aishwarya Bose, Sandip Nandi

arXiv:2610.23060v1cs.CLcs.CY

TL;DR

A single stated-risk answer can vary on retest, even though LLM advisers rely heavily on it to set stock allocations. This paper audits advice across repeated answers from real households and finds sizable same-person swings, while emphasizing that measuring steadier advice does not establish which allocation is right.

  • Problem

    A client’s risk attitude is represented by a noisy single survey answer, yet consequential advice treats that answer as the attitude itself.

    Source: §1 · pp. 1–2

  • Method

    The paper audits seven LLMs by comparing allocation advice for the same household facts paired with that person’s repeated risk answers.

    Source: §1 · p. 2, §10 · p. 13

  • Results

    10 to 16 points: the same client’s repeated answers move recommended stock share by that amount on average, with instability falling unevenly across clients.

    Source: §1 · p. 2, §10 · p. 14

  • Takeaways & Limitations

    The authors recommend same-person retest audits because persona-based tests cannot reveal how advice varies with one client’s repeated answers.

    Source: §10 · p. 13

    The benchmark for excess advice movement depends on assuming a model’s response to a very-reliable answer reflects its response to attitude; it measures consistency, not correctness.

    Source: §11 · p. 14

Abstract

from arXiv · show

AI financial advisers set a client's stock allocation largely from one stated-risk question, yet people do not answer it the same way twice: in a national US panel, 41% of households chose a different answer two years later, and in a laboratory retest a third of people moved two or more points on a 0-10 scale after six months. Seven large language models from five providers gave one-turn allocation advice for 800 real household profiles in which only the risk answer changed. The recommended stock share moved by 10 to 16 percentage points on average, and by 24 to 35 when the answer changed, against 0.6 to 4.2 points for the same prompt asked twice; in the laboratory retest the rating alone moved advice by 11 to 18 points. Without any note, the models respond to the answer about as strongly as when told that it is very reliable (less so for the two Claude models). If that instructed response is how each model maps attitude to advice, discounting the answer for its measured reliability, as classical test theory prescribes, would cut the swings by a factor of 2.4 to 3.3. The swing is largest for clients under 40. A note that the answer is noisy steadies some models, but shifts their weight to the client's age and pulls cautious clients towards a generic middle allocation: clients unwilling to take any risk are advised 22% stocks instead of 4%. Three extra money questions reduce the swing for some models, with no statistically detectable change relative to how much advice differs across people. In a scripted test, one clarifying question removes nearly all of the swing from an answer the client then corrects (33.5 points to 1.9 on average) while keeping the response to a change the client confirms. We argue that same-person retest audits should be a standard check for AI advisers, and that what to do with an unreliable answer is a question of client autonomy as well as of statistics.

1 INTRODUCTION

A single stated-risk answer is both consequential and noisy, making it possible for the same client to receive different advice without changed finances. The paper audits that instability in repeated-answer data and proposes a reliability benchmark and mitigations.

Retest audit

  • The retest audit measures how much LLM advice changes when real people change their stated-risk answer, with checks for prompt wording, sample size, and a crisis-free retest.The authors describe this as the first measurement using real people’s repeated answers.Source: §1 · p. 2

Reliability benchmark

  • The reliability benchmark uses classical test theory to compare advice movement with what the answer’s measured reliability warrants.It approximates each model’s mapping from attitude to advice using its response when told the answer is very reliable, an assumption-dependent choice.Source: §1 · p. 2

Distribution of instability

  • Instability is uneven: swings are largest for younger households, while the larger swings among less wealthy households shrink once age is held fixed.Source: §1 · p. 2

Mitigations and trade-offs

  • Proposed mitigations trade off stability and client-specific weighting: a noise warning steadies some models, extra questions add caution more than stability, and a clarifying question limits swings after corrected answers.In the scripted test, the clarification keeps the response to a change the client confirms.Source: §1 · p. 2

Reusable audit

  • The authors provide code and 77,014 model replies to support reusable same-person audits by deployers and regulators.Source: §1 · p. 2

Scope

  • The audit measures recommendation changes, not client harm, and its reliability benchmark depends on an assumption about each model’s response to a very reliable answer.Source: §1 · p. 2

2 RELATED WORK

Prior work documents risk-tolerance sensitivity in LLM and robo-adviser recommendations, instability in model outputs, and imperfect retest reliability of stated-risk measures. The paper identifies a gap at their intersection: prior studies had not measured how clients’ answer changes affect advice.

LLM and robo financial advice

  • LLM and robo-adviser studies find that allocations often depend strongly on stated risk tolerance, while other audits examine investment biases, risk profiling, and advice variation across reruns.Source: §2 · p. 3

Reliability of stated risk tolerance

  • Survey and psychometric research finds single risk-attitude items only moderately stable over time, with response error contributing substantial variance to some hypothetical risk answers.Some changes may be real, including preference shifts after the 2008 crisis.Source: §2 · p. 3

Consistency and auditing of LLMs

  • Prior LLM consistency audits treat the model as the object measured; this study instead treats it as an instrument that consumes a noisy measurement.The authors note that prior variation studies measured model sampling noise rather than noise in the client’s input.Source: §2 · p. 3

3 HOW NOISY IS THE ANSWER?

Two public retests establish that stated-risk answers can change over time, and the audit design uses those answer changes to isolate their effect on stock-allocation advice.

3 HOW NOISY IS THE ANSWER?

  • 41% of SCF households changed answers over two years, while 36% of pooled Basel–Berlin retest participants moved at least two points on the 0–10 scale.The SCF retest correlation was 0.47; the pooled Basel–Berlin correlation was 0.57.Source: §3 · pp. 3–4, Table 1 · p. 4

4 AUDIT DESIGN

  • The audit profiles include 800 SCF households whose arms differ only in the person’s 2007 or 2009 risk answer, plus 171 Basel–Berlin retest participants.The laboratory profiles include country, education, a fixed age band, and savings amount; public files lack age and wealth.Source: §4 · p. 4
  • Seven models returned stock-share recommendations; the first 200 SCF profiles were queried twice per prompt to provide a same-prompt control.Source: §4 · p. 4
  • The audit defines swing as the mean absolute advice difference between a person’s answers and compares it with advice spread and the slope implied by real holdings.Bootstrap intervals are conditional on the first survey imputation and eligible panel, omitting imputation and survey-design variance.Source: §4 · p. 4, Table 2 · p. 5, Figure 1 · p. 5

5 SAME PERSON, DIFFERENT ADVICE

Changing only a household’s stated risk answer substantially changes the recommended stock share, and robustness checks show the pattern persists across prompt wording, samples, and retests. The advice curve also rises more steeply with stated risk than real holdings do.

5 SAME PERSON, DIFFERENT ADVICE

  • 10.5 to 16.3 percentage points on average, and 24 to 35 points for households whose answer changed, separate recommendations; repeating the same prompt moves advice only 0.6 to 4.2 points.Between 23% and 35% of households see advice move by 20 points or more.Source: §5 · p. 5
  • 19.5 to 28.5 points per answer level is the model advice movement; one-level differences in real holdings correspond to 10.4 points, a comparison the authors caution is weak.Households may be inert, and some of the gap may reflect good advice; the paper draws no conclusion from this comparison.Source: §5 · p. 5

6 ROBUSTNESS

  • 11.2 to 17.7 points in the crisis-free Basel–Berlin retest show that the rating alone moves advice by as much as the two-year SCF swing, though retests cannot separate response error from genuine preference change.Source: §6 · p. 6
  • 11.8 to 20.5 points across rewritten prompts; each rewording changes a model’s swing by only −1.4 to +2.5 points relative to the original prompt.Five of 21 paired differences exclude zero; the largest is Mistral Large 4 in the client’s voice, at +2.5 [0.5, 4.7].Source: §6 · p. 6, Table 3 · p. 6
  • 11.2 to 18.1 points in the first 200 households and 10.3 to 15.7 in the later 600 preserve the model ranking and gap to real holdings; temperature-0 reruns likewise show no detectable swing change.At temperature 0, swings are 12.1 to 17.6 points versus 11.8 to 18.2 by default, with paired differences of −0.6 to +0.5 points.Source: §6 · pp. 6–7
  • 5.3 to 7.8 points per answer point describe the near-linear middle of the 0–10 advice curve, while controlled real holdings change by 0.9 equity-share points per answer point.At answer 0, GPT-6.1 Sol and Gemini 3.8 Flash recommend 0% stocks for all 100 profiles; real holdings imply 26% to 37% stock share across the scale.Source: §6 · p. 6, Figure 2 · p. 7

7 HOW MUCH SHOULD ADVICE MOVE?

The paper benchmarks advice movement against classical test theory’s reliability-adjusted estimate of risk attitude. Under the assumption that a model’s high-reliability instructed response represents its attitude-to-advice mapping, observed swings exceed that benchmark, though the size depends on reliability and modeling assumptions.

7 HOW MUCH SHOULD ADVICE MOVE?

  • 0.47 reliability and R^2 = 0.21 yield λ = 0.33 [0.28, 0.38], so a one-level answer change should move estimated attitude by a third of a level.The benchmark substitutes Kelley’s true-score estimate, conditioned on stable profile facts, for the stated answer in the model’s advice rule.Source: §7 · pp. 7–8
  • The claim that swings are excessive depends on five assumptions, including independent measurement noise, linear advice, retest correlation as reliability, and interpreting instructed sensitivity as the model’s undiscounted response to attitude.Source: §7 · p. 8
  • 2.4 to 3.3 times the corrected benchmark is the observed swing under the assumption that a model’s response to a high-reliability instruction represents its undiscounted attitude-to-advice mapping.Five models retain 0.96 to 1.00 of their instructed answer slope without the note; the two Claude models retain 0.89 and 0.77.Source: §7 · p. 8
  • 1.8 to 2.4 times the corrected benchmark at reliability 0.57, and 1.4 to 1.8 times at 0.68, show that estimated excess swings shrink as assumed reliability rises.At 0.8, the ratio is 1.3 to 1.5 for six models and 1.08 [0.98, 1.20] for Claude Sonnet 5.5.Source: §7 · p. 8, Figure 3 · p. 9
  • 2.0 to 2.7 times the corrected benchmark remains the range on a latent continuous scale, with latent reliability 0.56 [0.52, 0.60] and λ = 0.40 [0.34, 0.46].Source: §7 · p. 9
  • With no informative profile facts, the benchmark and plain-prompt swing/spread ratios are similar, so the benchmark suggests a noisy-answer note may help more when stable facts are available; the comparison is uncontrolled.In Basel–Berlin profiles, the ratios are 0.57 to 0.61 for the benchmark and 0.59 to 0.65 for the plain prompt.Source: §7 · p. 9

8 WHO BEARS THE WOBBLE?

Advice instability varies across household groups, with the clearest adjusted association tied to age rather than wealth. Both how often answers change and how strongly models respond to a change contribute to group differences.

8 WHO BEARS THE WOBBLE?

  • Group differences reflect both answer-change exposure and advice sensitivity; among households changing answers, a one-level change moves advice about 32 points for the bottom wealth third and under-40s.The corresponding swing is 24 to 26 points for the top wealth third and over-60s, despite similar answer-change sizes across groups.Source: §8 · pp. 9–10
  • The 800 audited households all had at least $5,000 in financial assets in 2007, leaving the least wealthy households outside the analysis.The group comparison reports advice instability, not advice levels; whether the disparity harms clients is unobserved.Source: §8 · pp. 9, 11
  • Under-40 households see a 20.4-point average swing, versus 10.0 points for households 60 and over.They also change answers more often; the age gradient remains after adjustment for household characteristics and answer movement.Source: §8 · pp. 9–10, Figure 4 · p. 10
  • Bottom-third net-worth households see a 17.1-point average swing, versus 11.2 points in the top third, but the adjusted wealth association is indistinguishable from zero.Younger age remains associated with smaller swings of 2.5 points per decade, or 1.7 points after accounting for answer movement and cautious-level changes.Source: §8 · pp. 9–10

9 CAN THE ADVICE BE STEADIED?

The paper tests prompt-based mitigations and a client clarification question for stabilizing advice when risk answers may be noisy. In the scripted test, clarification nearly eliminates the swing while retaining the response to a confirmed change.

9 CAN THE ADVICE BE STEADIED?

  • The noisy-answer note cut the SCF swing by 4.0 to 8.7 points for six models, but raised average advice by 3.1 to 8.8 points for six models.It shifted weight toward age for several models; clients reporting no willingness to take financial risk received 22% stocks on average instead of 4%.Source: §9 · p. 11
  • In Basel–Berlin profiles, the noisy-answer note shrank the swing for most models but did not lower the swing-to-spread ratio for any model.The samples differ in more than their facts, so this is not a controlled test; the note also raised average advice by 2.1 to 16.1 points for six models.Source: §9 · p. 11
  • Three additional money questions cut raw swing by 1.5 to 5.4 points for six models, with no statistically detectable change in swing relative to spread for six models.These product-specific questions are not parallel measures of the same attitude, so the result does not test a validated multi-item scale.Source: §9 · p. 11
  • The clarification test used 200 SCF households and scripted whether a differing answer was a mistake or a deliberate change.The prespecified tests required reduced unwanted swing, retention of at least 0.8 of the direct response to a confirmed change, and less than 3 points of movement for a stable client.Source: §9 · p. 12
  • 33.5 points of average advice movement for a one-level answer change fell to 1.9 points after clarification, while response to a confirmed change reached 1.12 times its direct value.All models passed the swing and responsiveness tests; the exchange and an equivalent one-turn prompt performed similarly.Source: §9 · p. 12, Figure 5 · p. 12
  • Prompt-based mitigation is model-specific and can have side effects, so deployers must measure how a model uses a warning about measurement error.The scripted clarification distinguishes a mistaken answer from a real change; its replies were unambiguous, making this a best-case test.Source: §9 · pp. 12–13

10 DISCUSSION

The discussion recommends same-person retest audits and advice designs that rely on stable facts and multiple, reliability-weighted measurements. It also cautions that steadier advice can override a client’s genuine risk preference, and that regulatory implications remain to be tested.

For deployers

  • Pair each profile with a retest answer and report the advice swing and swing/spread ratio; treat the reliability benchmark as an assumption-dependent warning reference.The audit and its checks used 47,264 requests and about $98 in API fees; the clarification test used 29,750 requests and $64.Source: §10 · p. 13

For design

  • Advice should combine stable facts with several attitude measurements weighted by reliability, rather than letting one answer drive allocation.Under the corrected benchmark and its assumptions, a one-level answer change would move advice between a third and two-fifths as much as current models do; this is a consistency standard, not a claim of correct advice.Source: §10 · p. 13

For clients

  • Gathering more information can steady advice without overriding the client: one scripted clarifying question retained all response to a confirmed change, while the noise note retained half.The note advised clients unwilling to take any financial risk 22% stocks instead of 4%; the authors distinguish correction of a noisy answer from overriding genuine caution.Source: §10 · p. 13

For regulation

  • A same-person retest check is a regulatory implication to test, not a requirement the authors say current rules already impose.The discussion also notes that instability is largest for young clients and that less wealthy clients see larger swings before age is accounted for.Source: §10 · pp. 13–14

Scope

  • The paper measures recommendation changes, not suitability violations or investment outcomes, and does not claim steadier advice is better for every client.The claim that advice exceeds the reliability-adjusted benchmark is conditional on the benchmark’s assumptions; at reliability 0.8, Claude Sonnet 5.5 is statistically compatible with it.Source: §10 · p. 14

11 LIMITATIONS

The limitations qualify both the reliability benchmark and the populations and settings covered by the audit. The ethical statement describes a possible misuse and the authors’ response, while the usage statement reports how generative AI contributed to the work.

11 LIMITATIONS

  • The benchmark has no ground truth for correct advice and depends on assumptions about how the models respond to a very-reliable note.The SCF panel straddles the 2008 crisis; the Basel–Berlin profiles lack age and wealth, and its comparison with the SCF mitigations is suggestive rather than controlled. Intervals omit imputation and survey-design variance, and the eligible panel excludes households with under $5,000 in financial assets. Apart from one scripted exchange, the study tests single prompts, one time point, and English only; group analysis is observational.Source: §11 · p. 14

ETHICAL CONSIDERATIONS STATEMENT

  • A potential misuse is tuning an adviser to pass the panel retest without improving its behavior on real clients.The authors release the method rather than a fixed test set and recommend audits on deployers’ own returning clients.Source: ETHICAL CONSIDERATIONS STATEMENT · pp. 14–15

GENERATIVE AI USAGE STATEMENT

  • Generative AI assisted with analysis scripts and drafting, while the authors directed design and reviewed every analysis and sentence.The main audit protocol and sample were fixed before results were seen, and the clarification test plan and analysis code were committed before collecting its answers.Source: GENERATIVE AI USAGE STATEMENT · p. 15

A MODELS AND PROMPTS

The appendix specifies the model-serving setup and the contents of the profile and prompt variants used in the audit. It also defines the request and readable-reply counts reported for the study.

Models and settings

  • Requests used OpenRouter’s chat-completions API with default provider routing on 9–11 October 2026, with no sampling parameters and provider-default settings.The seven listed systems were Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Sol, GPT-6 Luna, Gemini 3.8 Flash, Grok 4.7, and Mistral Large 4; the serving provider was not recorded.Source: A MODELS AND PROMPTS · p. 16

SCF profile (main audit)

  • In the main SCF audit, the client profile gave age 46, high-school-diploma education, household income $43,000 per year, net worth $219,000, financial assets $37,000, and a question about risk when saving or investing.Source: A MODELS AND PROMPTS · p. 17

Prompt variants

  • The noisy-answer and reliable-answer variants added notes about retest consistency, while the Basel–Berlin profile used a 0–10 willingness rating and included separate money-question and note variants.The Basel–Berlin profile fixed the age band at late twenties and savings at 20,000 in local currency; its note described movement of two or more points after six months.Source: A MODELS AND PROMPTS · p. 17

Rewordings and request counts

  • Prompt rewordings included a fiduciary system instruction, first-person client voice, and a three-asset stocks/bonds/cash allocation request.Table 4 defines a request as one model, profile, prompt, and repeat; “readable” replies yielded a stock share, and “two answers” means the 2007 and 2009 answers.Source: Table 4 · p. 17, A MODELS AND PROMPTS · p. 18

B THE BENCHMARK ESTIMATE

The benchmark estimates how much the risk answer reflects underlying attitude after accounting for facts that predict the answer. It compares model advice rules with reliability-adjusted versions, including a latent-scale treatment of categorical answers.

Reliability adjustment

  • The reliability adjustment estimates attitude from the answer and facts, discounting the part of the answer not explained by the facts according to its reliability.The reliability of the fact-unpredicted part is denoted λ; with no informative facts, it is Kelley’s estimate, λ=ρ.Source: B THE BENCHMARK ESTIMATE · p. 18

On the latent scale

  • For the latent-scale analysis, a four-level answer is treated as a cut of a continuous response, and its expected value within the answer category is used as the model input.The estimate uses a panel-weighted ordered probit on 2,859 households and a weighted polychoric correlation of the 2007 and 2009 answers; estimated λ*=0.40 [0.34, 0.46].Source: B THE BENCHMARK ESTIMATE · p. 18

Original and corrected benchmarks

  • The original benchmark fits advice to plain-prompt answers, while the corrected benchmark fits advice under the very-reliable note and assumes that response represents the model’s undiscounted attitude mapping.The corrected benchmark can understate excess if the note makes a model lean harder on the answer, or overstate it if the model still discounts internally.Source: B THE BENCHMARK ESTIMATE · p. 18

Benchmark comparison

  • Table 6 reports observed swing over benchmark swing by benchmark and assumed reliability for the first 200 SCF households, with paired differences and 95% bootstrap intervals.At ρ=0.47, the estimate uses λ=0.33; the latent scale uses λ*=0.40.Source: Table 6 · p. 20

C SWING BY HOUSEHOLD GROUP

Table 7 reports how often risk answers changed and how much advice swung, overall and among households whose answers changed. It reports panel-weighted and audited-sample shares alongside all-model averages and single-model ranges.

C SWING BY HOUSEHOLD GROUP

  • Table 7 distinguishes the share of households whose answer changed from advice swing among all households and among those whose answer changed.It reports changed-answer shares for 2,859 eligible panel households and the 800 audited households.Source: Table 7 · p. 21, C SWING BY HOUSEHOLD GROUP · p. 21
  • The panel-wide changed-answer share is panel-weighted, while the audited-household share is unweighted because households were sampled proportional to weight.Source: Table 7 · p. 21
  • For each household group, the table reports the seven-model average swing with 95% bootstrap intervals and the lowest-to-highest single-model swing.Source: Table 7 · p. 21

D MITIGATIONS

The section evaluates a noise note, added money questions, and a clarifying exchange as ways to address advice swings. In the scripted clarification test, the exchange sharply reduced swings for answers clients corrected while largely retaining responses to confirmed preference changes.

D MITIGATIONS

  • Table 8 compares a noise note with a plain prompt and three money questions with rating-only advice, including swing, spread ratios, and mean advice.The passage supplies the comparison scope but not the table's outcome values.Source: Table 8 · p. 22, D MITIGATIONS · p. 22
  • Averaged over models, the exchange reduced advice swings among clients under 40 (41.9 to 2.5 points), ages 60 and over (28.6 to 1.2), and previously no-risk clients (40.2 to 0.8).After a confirmed change, advice stayed put or moved the wrong way for 2% of households with the exchange, versus 4% at once.Source: D MITIGATIONS · p. 22
  • With the exchange, average swing fell from 41.9 to 2.5 points for clients under 40 and from 28.6 to 1.2 for clients aged 60 and over.Among clients previously unwilling to take any financial risk, average swing fell from 40.2 to 0.8 points.Source: D MITIGATIONS · p. 22
  • The clarification test used a question distinguishing deliberate preference change from misunderstanding or uncertainty, with clients confirming, correcting, or deliberately changing their answer.Conditions used fresh conversations in random order; each reply had two phrasings.Source: D MITIGATIONS · p. 22
  • Table 9 reports unwanted swing, response to confirmed change, retention of that response, and the clarifying question's effect on stable-client advice across four conditions.The table covers 200 SCF households with two repeats and reports bootstrap intervals for retention and a 90% interval for the pre-set stable-client test.Source: Table 9 · p. 23
Loading 2610.23060v1…