One advisor for the whole world?

5Main Results

This section reports how LLM advice varies across countries. I first document the raw patterns in the advice, then I estimate the country effects on the primary question, decompose them into language and local framing, and ask whether they track measured risk preferences and financial literacy. I then turn to the mortgage and inflation questions and close with robustness. The pre-registered test of age and pension generosity is reported with the other tests of the advice rule in Section 6.2.

Except for the two secondary questions in Section 5.4, the outcome is the recommended equity fraction on Question 1. Every regression includes model fixed effects, so each contrast compares the advice of a single model. I cluster standard errors at the model level. As pre-registered, the twenty primary country contrasts carry a Holm step-down correction as a single family. The reason for the Holm step-down procedure is that I test 20 countries against the United States. With twenty contrasts each tested at the 5 percent level, about one would appear significant by chance even if advice were identical in every country. The correction keeps the probability of any false rejection among the twenty at 5 percent. Appendix Table A7 lists each pre-registered hypothesis with its test and outcome. The coefficients measure the causal effect of the prompt’s country, language, and framing on AI financial advice.

5.1Descriptive Patterns

Figure 1Distribution of recommended equity shares in the primary cell
Notes: Share of responses at each recommended equity share on the primary cell (Question 1, Version B, age 40, \(N = 13{,}639\)), pooled across the twenty-one countries and thirteen models. Bars holding at least 5 percent of responses are labeled with their share.

I begin with descriptive facts about the advice that organize the results that follow. The models bunch their recommendations at a few round numbers within a narrow range. I can reject equality between countries. The differences between countries are one-sided (every significant contrast is negative) and modest (six points on average, at most eleven).

Starting with round numbers in the recommendations, Figure 1 plots the distribution of the recommended equity share, pooled across countries and models. Every response is a multiple of five percentage points, and 85 percent are multiples of ten. The modal recommendation is 60 percent equity, given in 31 percent of responses. In addition, the support is narrow. No response in 13,639 allocates less than 10 or more than 80 percent to equities. The bunching shows that the models answer a continuous allocation problem with a conventional round number. Choukhmane et al. (2026) also find round-number bunching in AI advice, which they note departs from the more subtle prescriptions of life-cycle theory. A portfolio rule would produce a continuous answer for at least some of the answers. Instead, the modal 60/40 split is the most common target allocation in retail portfolio advice, a first sign of the risk-questionnaire interpretation. This is not just because the models lack information about the investor, as 100 percent of the responses in the stated-investor arm are also round numbers. Ross and Lo (2026) report that between 89 and 100 percent of all allocation percentages are multiples of 5 without web search, concluding that “round-number anchoring is a deeply ingrained property of model behavior that web search alone does not fully override.”

Table 1Advice by model: level, dispersion, and stability
Stability across repetitions
Model Vendor Mean SD Within-cell SD Identical cells
Gemini 3.1 Pro Preview Google 0.37 0.08 0.037 0.05
GPT-5-mini OpenAI 0.38 0.08 0.043 0.00
GPT-5.5 OpenAI 0.41 0.05 0.032 0.00
DeepSeek V4 Flash DeepSeek 0.45 0.12 0.110 0.00
DeepSeek V4 Pro DeepSeek 0.48 0.12 0.105 0.00
Grok 4.3 xAI 0.54 0.09 0.065 0.00
Gemini 3 Flash Preview Google 0.56 0.07 0.000 0.62
Gemini 3.5 Flash Google 0.57 0.06 0.020 0.24
Claude 4.8 Opus Anthropic 0.58 0.04 0.000 0.52
Claude 4.6 Sonnet Anthropic 0.63 0.04 0.000 0.67
Claude 4.6 Opus Anthropic 0.65 0.07 0.000 0.95
Mistral Medium 3.5 Mistral 0.70 0.04 0.024 0.29
Claude 4.5 Haiku Anthropic 0.71 0.04 0.007 0.48
Notes: Recommended equity share by model on the primary cell (Question 1, Version B, age 40), pooled across the twenty-one countries, roughly 1,050 responses per model. Within-cell SD is the median, across a model’s twenty-one country cells, of the standard deviation over the fifty repetitions in the cell. Identical cells is the share of those cells whose fifty repetitions return one identical number, despite sampling at temperature 0.7. A variance decomposition of the primary cell attributes 69.5 percent of the total variation to model identity, 5.7 percent to the country, 9.5 percent to their interaction, and 15.3 percent to within-cell sampling.

The design also varies the investor’s age. The recommended share falls by about 0.7 percentage points per year of age, a gradient that Section 6.2 examines together with the pension system.

Across countries, Appendix Table A8 shows that mean risky share spans twelve percentage points, from 0.46 in Brazil to 0.58 in South Korea. The United States is at 0.57, near the top of the range. The band is narrow relative to the dispersion within any single country. The typical within-country standard deviation is 0.13. Instead, most of the within-country dispersion comes from the models. Table 1 reports mean advice by model, which spans 0.37 (Gemini 3.1 Pro) to 0.71 (Claude Haiku 4.5), a range of 34 percentage points. That is nearly three times the country range. The ordering of the models follows neither vendor nor scale. The four Anthropic models cluster above the pooled mean, between 0.58 and 0.71. Google’s flagship reasoning model is the most conservative advisor in the panel, but its two Flash-tier models are twenty points higher. Model identity accounts for 69.5 percent of the total variation in the primary cell, the country for 5.7 percent, their interaction for 9.5 percent, and sampling variation within country–model cells for the remaining 15.3 percent. The dominance of the model motivates two choices maintained throughout the paper. Every regression identifies country contrasts within model, and inference clusters at the model level, treating the thirteen models rather than the quarter-million responses as the effective sample.

5.2Country effects on the primary question

I now formally test whether LLM advice is the same across countries. The primary specification estimates country fixed effects from the following regression \[y_{imc} \;=\; \alpha_{m} + b_{c} + \varepsilon_{imc}, \tag{3}\] where \(y_{imc}\) is the equity fraction recommended in repetition \(i\) by model \(m\) for country \(c\), the \(\alpha_{m}\) are model fixed effects, and the \(b_{c}\) are deviations from the omitted U.S. baseline. The baseline regression uses the age-40 version of the prompt in the local language.

The main finding is that advice is not the same everywhere, as shown in Figure 2 for all twenty contrasts. Thirteen of the twenty contrasts clear the wild-bootstrap threshold before correction, and seven survive the Holm step-down: Brazil at \(-11.0\) percentage points, Italy at \(-8.6\), Poland at \(-7.4\), France at \(-5.4\), Spain at \(-4.4\), the Netherlands at \(-4.2\), and Mexico at \(-2.7\). In addition, every contrast that survives the correction is negative. Against a U.S. mean equity share of 0.57, the largest gap, Brazil’s eleven percentage points, moves $5,500 of the $50,000 stake from stocks to bonds. The seven significant contrasts average about six points. For comparison, the gender gap that Foltyn and Olsson (2026) document on a similar allocation question is under two percentage points. Country context moves LLM advice by several times the demographic gaps documented so far in the literature.

Figure 2Country effects on the recommended equity share
Notes: Point estimates of the country fixed effects from equation (3) (Question 1, Version B, age 40), in percentage points relative to the omitted U.S. baseline. Bars are 95 percent wild cluster bootstrap confidence intervals (Webb six-point weights, 9,999 replications, clustered by model, \(G = 13\)), the intervals of record, matching Appendix Table A8. Filled markers denote contrasts that are significant at the 5 percent level after Holm correction, and the column on the right repeats the point estimates.

After the Holm correction, thirteen of the twenty contrasts fail to reject equality with the U.S. baseline. Six of those are individually significant on the uncorrected wild-bootstrap test but do not survive the step-down (China, Switzerland, Germany, Turkey, India, and Canada, ranging from \(-4.4\) to \(-1.1\) points). The remaining seven do not reject even before correction. Four of them have small positive point estimates (Australia, Japan, South Korea, and the United Kingdom). Finland and South Africa have small negative gaps, and Sweden sits at the U.S. level. The precision of these nulls differs across countries, because the models agree with one another far more in some countries than in others. In the three English-prompting countries (Australia, South Africa, and the United Kingdom) the 95 percent confidence intervals exclude gaps larger than about two and a half percentage points in either direction, which is $1,250 of the $50,000 stake. For the United Kingdom this is the pre-registered equivalence test H3. Its contrast is \(+0.7\) points with a 95 percent confidence interval of \([-0.8, +2.2]\), inside the registered \(\pm 5\) point margin. In Finland, Japan, South Korea, and Sweden the intervals are wider and reach between 3.4 and 5.7 points from zero, so the data cannot exclude gaps of the size estimated for Spain or the Netherlands.

5.3Language and local framing

I next run two within-country tests. One asks whether local framing shapes the advice, and the other separates the effect of the question’s language from that of the country name. The local-framing regression compares Versions C and B across the twenty non-U.S. countries at age 40, \[y_{imc} \;=\; \alpha_{m} + b_{c} + \phi\,\mathbf{1}[\text{version}=C] + \varepsilon_{imc}, \tag{4}\] and the language regression compares Versions B and B-cross across the fifteen non-English countries, \[y_{imc} \;=\; \alpha_{m} + b_{c} + \lambda\,\mathbf{1}[\text{language}=\text{local}] + \varepsilon_{imc}. \tag{5}\]

Asking in the local language rather than in English lowers the recommended equity share by 1.3 percentage points (\(\hat{\lambda} = -0.013\), wild bootstrap \(p = 0.007\), confidence interval \(-2.3\) to \(-0.4\) points). The language effect is statistically significant but economically small, an order of magnitude below the largest country effects. Naming the local instruments does even less. Replacing the generic vocabulary with local pension names, local government bonds, and local index funds moves the allocation by \(-0.7\) points, a result that is not significant at the five percent level. Figure 3 applies the same decomposition country by country. The country-name step is the English version of the prompt minus the United States (Version B minus U.S.), the language step is the native-language prompt minus the English prompt for each country (Version B minus B-cross), and the local-context step is the native-language prompt with local context minus the native-language prompt (Version C minus Version B). The country-name step also changes the currency unit and the nominal amount, which move together with the country reference. In the large-effect countries (Brazil, Mexico, Poland, France, Italy, Germany) the country-name slab accounts for most of the gap. The language and local-context slabs are usually small, but they are noisier than the pooled \(\hat{\lambda}\) and \(\hat{\phi}\) and not uniformly negative. Japan is the clearest exception. A large positive language slab pushes its advice toward stocks and produces the panel’s largest positive total.

Figure 3Decomposing the country effect into naming, language, and local context
Notes: Each bar decomposes a country’s fully localized effect on the recommended equity share, Version C relative to the omitted U.S. English generic baseline (Question 1, age 40), into three additive cues in percentage points. Country name is the English-only contrast Version B-cross minus the U.S. (language and instruments held fixed, though the currency unit and nominal amount change with the country reference), language is Version B minus B-cross (country name held fixed), and local context is Version C minus B (language held fixed). The black diamond marks the total, which equals the sum of the three slabs by construction. The six English-prompting countries have no B-cross cell, so their language slab is zero and country name is Version B minus the U.S.

The size of the language effect helps distinguish between two views of how multilingual models answer a question. If the models retrieve advice the way they retrieve language-keyed knowledge, the language effect \(\lambda\) should be substantial. The same model often asserts different facts depending on the query language (Kassner et al. 2021; Qi et al. 2023). For example, if German queries were to draw on the bond-heavy conventions of German-language financial text, the effect of the language would presumably be large. If instead the models solve the task in a shared representation, \(\lambda\) should be near zero. Tracing studies in computer science find that language-specific processing lies in the layers nearest the input and output (i.e., the initial prompt and the explanation, which is always in the local language). More importantly, however, tracing studies also find middle layers, where the LLM solves the task in a shared concept space lying closer to English (Wendler et al. 2024; Tang et al. 2024; Wu et al. 2025; Lindsey et al. 2025). Under such processing, two translations of the same question are nearly the same query in the middle layer, which supports a small language effect. Indeed, the estimate of \(-1.3\) points fits this second account, with the small residual consistent with the language-specific features that remain. The advice behaves like one policy applied across query languages rather than like language-segmented retrieval.

5.4The mortgage and inflation questions

Figure 4Country effects across the three questions
Notes: Point estimates of the country fixed effects from equation (3) for Question 1 (recommended equity share), Question 2 (recommended fraction invested rather than used to prepay a 4 percent mortgage), and Question 3 (recommended fraction moved out of a 1 percent savings account), Version B, age 40, in percentage points relative to the omitted U.S. baseline. Each question is estimated separately with model fixed effects. The three panels share one horizontal scale. Lines are 95 percent wild cluster bootstrap confidence intervals, reported in Appendix Tables A8 and A9. The range above each panel is the distance between the largest and the smallest country effect, with the United States at zero. Hollow markers denote the premise-acceptance group on Question 2. Countries are ordered by their Question 1 effect.

I now turn to the two secondary questions. Figure 4 plots the country effects on all three questions side by side. For the two secondary questions, the advice differs more across countries and the variance decomposition reverses. Country accounts for 27 percent of the variation in mortgage advice against 15 percent for model. Model identity accounts for only 26 percent of the variation in inflation advice, whereas on the allocation question the shares are 6 percent for country and 70 percent for model. Agliata and Hasso (2026) also note the importance of the model for financial advice. The secondary questions also pick up a different country signal from the portfolio allocation question. The countries that receive less equity on the allocation question (Brazil, Italy, Poland and France) are not the countries told to prepay (Japan, China, Switzerland, and Germany). Japan, statistically at the U.S. level on the allocation question, has the largest gap on both secondary questions. Across countries the allocation and mortgage effects correlate at \(-0.10\), while the mortgage and inflation effects correlate at \(0.80\).

On the mortgage question, the models recommend investing 10 to 29 percentage points less than in the United States in each of the fifteen countries outside the premise-acceptance group. All fifteen contrasts survive the Holm correction. Among the largest gaps are Japan at \(-28.6\), China at \(-22.7\), Switzerland at \(-21.6\), Germany at \(-21.1\), and Spain at \(-20.4\). Japan’s gap alone is more than twice the largest contrast on the allocation question. Part of the common gap reflects the United States itself, where deductible and freely prepayable mortgages favor investing. The ordering among the fifteen countries, however, spans a further 19 points and places the low-inflation economies at the bottom. For the five countries where the prevailing mortgage rates exceeded twice the prompted 4 percent at the filing date (Brazil, India, Mexico, Turkey, and South Africa), the contrasts are mostly null, and only South Africa (\(-2.4\) points) survives the correction.

On the inflation question, seven country contrasts survive the Holm correction, all negative. In these countries the models recommend moving less of the savings out of the account than in the United States: Japan (\(-11.3\) points), China (\(-7.2\)), South Korea (\(-6.9\)), the Netherlands (\(-6.5\)), Italy (\(-5.0\)), Switzerland (\(-5.0\)), and France (\(-4.3\)). In addition to the share, I also ask the models to recommend an instrument to save in (see Appendix Table A13). Here we find the clearest evidence for localization: Equities are the primary instrument in 93 percent or more of responses in fourteen countries and in a clear majority in all but the highest-inflation economies. In Brazil, Turkey, and Mexico the models switch to inflation protection. In Brazil, inflation-linked bonds account for 86.7 percent of the recommendations, and in Turkey inflation-linked bonds and gold together account for 82.3 percent (66.0 and 16.3 percent). Mexico’s recommendations tilt the same way (58.1 percent inflation-linked bonds), and Poland’s are split (38.6 percent inflation-linked). The tilt appears even inside the English-prompting group, where models in the two higher-inflation members, India and South Africa, recommend inflation protection (11.5 and 16.3 percent, compared to under 3 percent in the United States, the United Kingdom, Canada, and Australia). Because these two countries share the baseline’s language, the switch tracks the country rather than the language of the prompt. The models’ own beliefs follow the same ordering. Asked for the inflation rate they assume when the mortgage question states no rate (described below), they give 4.3 to 5.3 percent for Brazil, Mexico, India, and South Africa and 40.6 percent for Turkey, against 1.4 to 3.7 percent in the other sixteen countries.

The country effects on the mortgage question sort into three groups by the local level of interest rates. In Japan and Switzerland, where policy rates were below one percent in 2024, the models recommend 25 points less investment than in the United States. In the five premise-acceptance countries, where policy rates exceeded six percent, the gap is 2 points. In the remaining thirteen countries the gap averages 17 points and is unrelated to the local rate (slope 0.07 points per point of the policy rate, \(p = 0.96\)). The pattern is consistent with models comparing the stated 4 percent with local rates only where the two differ widely.

Figure 5The mortgage question with and without the stated rate
Notes: The English Question 2 block (Version B in the six English-prompting countries, B-cross elsewhere) at age 40, asked on November 16 and 17, 2026 with the clause stating the 4 percent mortgage rate deleted. Thirteen models, twenty-five repetitions per country. Country fixed effects from equation (3) relative to the United States, plotted against the prevailing mortgage rate in the pre-registered rate supplement (log scale) and estimated separately for the numbered English cell of the main collection (hollow markers, 4 percent stated) and for the arm without the rate (filled markers). Arrows connect the two for each country. Triangles mark the premise-acceptance group. Estimates, confidence intervals, and Holm-corrected \(p\)-values are in Appendix Table A10.

To test this directly, I removed the stated 4 percent rate from the mortgage question. After the recommendation, I also ask each model for the mortgage rate and the inflation rate it assumed when answering the question.

The first hint that the models localize can be found in the rate that they assume. Across the twenty-one countries, the assumed and the prevailing mortgage rate correlate at 0.92 in logs, and the correlation for each model lies between 0.83 and 0.96. The one large departure is Japan, where the models assume 1.1 percent against 2.8 percent in the data.

Once the models choose the rate, Figure 5 shows that they give a different recommendation. In the eleven countries whose prevailing rate is below 4 percent, the average gap to the United States narrows from 17.9 to 5.0 percentage points. The change survives the Holm correction in eight countries. Japan moves from 26.9 points below the United States to 3.0 points above it, and Switzerland from 17.2 below to 1.5 above. Nine of the eleven contrasts are no longer distinguishable from zero. The premise-acceptance countries, where the mortgage rate was at least 8 percent, move the other way. Their average gap widens from 1.2 to 18.4 points, with Brazil at 28.7 and South Africa at 26.7 points below the United States. Across the twenty non-U.S. countries, the gap falls by 1.0 point for each point by which the prevailing rate exceeds 4 percent (wild bootstrap \(p < 0.001\)).

Turkey is the one premise-acceptance country that moves little. Its gap changes by 9.2 points, which is not statistically significant. The models assume Turkish inflation of 40.6 percent against a mortgage rate of 26.6 percent, which makes the real rate on the mortgage negative. With a negative real rate, it would be sensible to keep the mortgage and not prepay. The real return on prepaying is the interest rate on the mortgage, and if that rate is high, investing in stocks rather than prepaying is less favorable (Becker and Shabani 2010). The models’ advice is consistent with this rule. Within a country and model, responses that assume a mortgage rate one point higher recommend investing 1.6 percentage points less (\(p = 0.02\)), and responses that assume inflation one point higher recommend investing 0.8 points more (\(p = 0.10\)). I cannot reject that only the assumed real rate matters (\(p = 0.13\)).

Hence, with the rate stated, the models condition on the premise, even if the premise is unrealistic. The large gaps in the low-rate countries record how they price a 4 percent loan in economies where such a loan is expensive. Without the rate, they price the decision at local conditions that match the data.

5.5What does the advice track?

Financial advice from AI models differs across countries. I now ask whether the country effects track several country-level measures that an advisor may pick up on. The answer depends on the question. On the allocation question, the country effects track none of the characteristics I consider. On the mortgage and inflation questions, they track local interest rates.

I start with the willingness to take risks and financial literacy. I use data from the Global Preferences Survey from Falk et al. (2018) for risk-taking and data on the share of adults classified as financially literate in the S&P Global FinLit Survey from Klapper et al. (2015). If AI models tailor their advice to country characteristics, they may plausibly infer that a country with a higher willingness to take risks should be given a higher equity share. Similarly, if models take financial literacy into account, they may give a higher equity share to a country with higher financial literacy. Financial literacy also proxies for a wide range of other characteristics, like general financial sophistication or high stock market participation (Rooij et al. 2011). I test these two predictions by regressing the country fixed effects from equation (3) on each measure. Note that these tests are descriptive and that they ask whether the models impute national averages, not whether they should.

Figure 6Country effects against measured risk preferences and financial literacy
Notes: Estimated country fixed effects \(\hat{b}_{c}\) from equation (3) (Question 1, Version B, age 40, U.S. baseline omitted) plotted against the country mean of the GPS willingness to take risks (Falk et al. 2018) in Panel A and the share of financially literate adults from the S&P Global FinLit Survey (Klapper et al. 2015) in Panel B, for the twenty non-U.S. countries. Country effects are in percentage points. Ribbons are 95 percent country-bootstrap confidence bands around the fitted bivariate line (9,999 replications). Each panel reports the bivariate slope with its wild bootstrap \(p\)-value.

Figure 6 plots both relationships along with the statistical results. I find little support for either hypothesis. Panel A shows that the advice does not track measured risk preferences. The slope is flat and the wild bootstrap over countries does not reject zero. The point estimate is also economically small. A country that is one cross-country standard deviation more risk-tolerant than another receives \(0.8\) percentage points more equity. Even the widest preference gap in the sample, between Japan at \(-0.36\) and South Africa at \(0.97\), predicts a gap of under four points. At the same time, Japan receives advice statistically indistinguishable from the U.S. baseline, while South Africa, by far the most risk-tolerant country, receives slightly less equity than the baseline rather than more. Panel B tells the same story for literacy. Moving across the full literacy range in the panel, from 24 percent of adults in India and Turkey to 71 percent in Sweden, changes predicted advice by under three percentage points.

Measured preferences and literacy are not special in this respect. Appendix Table A14 regresses the country effects on each of sixteen country characteristics in turn, covering survey measures, income, the volatility and home equity premium, policy, long-term and mortgage interest rates, inflation, pension replacement rates, and mortgage-interest deductibility, with a wild bootstrap over the twenty countries and a Holm correction within each question. For Question 1 on the portfolio allocation, no characteristic survives the correction for multiple hypothesis testing. The two largest unadjusted associations are the English-prompt indicator and the pension replacement rate, but neither is significant after the Holm step-down (\(p = 0.33\) and \(0.25\)). Overall, no country characteristic is significantly associated with the country coefficient on advice in Question 1.

Figure 7Advice against the benchmark on the allocation and mortgage questions
Notes: Panel A plots the estimated country fixed effects \(\hat{b}_{c}\) from equation (3) (Question 1, Version B, age 40) against the benchmark’s prescribed deviations from the United States, \(\alpha^{S}_{c}-\alpha^{S}_{US}\), at the baseline grid point (\(\bar{\gamma}=3\)), for the twenty non-U.S. countries, both in percentage points. The dashed line is full calibration (\(\kappa=1\)), the dotted horizontal line is the U.S.-default anchor (\(\kappa=0\)), and the solid line is the fitted regression (11) with slope \(\hat{\kappa}=0.13\). Inference of record is the wild bootstrap over countries reported in the text. Panel B plots the mean recommended index-fund share on the mortgage question (Question 2, Version B, age 40) against the country-specific benchmark \(\alpha^{Q2}_{c}\) from Equation (12) at the same grid point. The dashed line is the locus of full calibration, on which advice equals the benchmark. Points above it recommend more equity, and hence less prepayment, than the benchmark. Triangles mark the premise-acceptance group.

Average advice lies within about six percentage points of the currency-hedged version of the simple benchmark in Section 4.1, but up to 33 points above the unhedged version in the countries with the most volatile currencies (Appendix Table A4). Regressing the country effects from Question 1 on the benchmark’s prescribed deviations from the United States gives a slope of \(\hat{\kappa}=0.13\) (Figure 7, Panel A). Dropping Brazil from the estimation lowers the slope to 0.03, dropping Turkey raises it to 0.21, and dropping any other country leaves it between 0.11 and 0.15. These results are all far from a slope of 1, which would indicate full pass-through of the currency risk, and are closer to zero, which would indicate that all countries get the same advice as the United States. The wild bootstrap over countries rejects full calibration (\(p<0.001\) against \(\kappa=1\)) but does not reject the U.S.-default hypothesis (\(p=0.32\) against \(\kappa=0\)).

However, the results for the mortgage and inflation questions are different. Interest rates and inflation positively predict the answer to the mortgage question. A one-standard-deviation higher log mortgage rate is associated with 6.6 percentage points more recommended investment, which survives the Holm correction (Holm \(p = 0.006\), \(R^{2} = 0.67\)), and the long-term rate and inflation each explain close to half or more of the cross-country variance. The models therefore draw on local economic conditions when a question involves a local interest rate, but not on the allocation question. Only part of this response is appropriate. Where inflation is high, the real return to prepaying the prompted 4 percent mortgage is low, so advising less prepayment is right. It does not follow that the savings belong in equities, because in the high-rate countries local bonds pay more than the mortgage costs. Across all twenty countries the advice tracks local rates but not its own benchmark (slope 0.5 points per standard deviation, \(p = 0.86\)).

Panel B of Figure 7 plots the mortgage advice against the benchmark of Section 4.2. In every calibration country the models recommend investing more than the benchmark prescribes, by 13 to 36 percentage points and by 28 on average (Appendix Table A5). Where the benchmark lies below 0.40, from Japan to Sweden, the gap is nearly constant at 31 to 34 points, so the advice rises one for one with the benchmark. Above 0.40 the advice stops rising and the gap narrows to between 13 and 22 points. The United States is the exception. Its benchmark of 0.47 is among the highest, yet it receives the highest advice in the panel, 0.83, and the largest gap, 36 points. On the fifteen non-U.S. calibration countries, the calibration regression gives \(\hat{\kappa}^{Q2} = 0.23\) for the mortgage question, against \(0.07\) for the allocation question on the same fifteen countries. With fifteen countries neither slope is statistically distinguishable from zero, and the 95 percent confidence intervals of both exclude full calibration at one. In the five premise-acceptance countries the models lean against prepaying, with average advice at about 0.80, within two points of the U.S. level. The benchmark agrees that these households should not prepay, but it does not prescribe more equity. Because local bonds pay more than the mortgage costs, the savings not used for prepayment belong in bonds, so the benchmark is the allocation benchmark of 0.21 to 0.49. The advice exceeds it by 31 to 60 points, more than in any calibration country. Some models name the bond option explicitly. One answer for Turkey notes that “even government bonds or deposit rates offer returns well above 4 percent” but still places the entire sum in the index fund, because “the question asks for a choice between paying the mortgage and investing.”

5.6Robustness

I now describe several robustness checks. I first collect two pre-registered checks where I ask whether the results depend on the sampling temperature or on the size of the numerals in the local-currency amount. After these checks, I also added several exploratory robustness checks where I first check whether the results change over time as new models are introduced, whether the English Instructions block affects the results, or whether allowing models to ask questions affects the cross-country differences. Finally, I also examine whether the answers depend on the country where I was when I sent the queries (Germany). The results are robust to these checks, which I now describe in more detail.

Temperature check.

Temperature is a sampling parameter that sets how much randomness a language model uses when it picks each word of its response. The next-token prediction at the heart of an LLM picks the next word based on a predicted probability. Model providers introduce some randomness in the token the model uses, which introduces some variation in the answers even with a similar prompt. At a temperature of 0, the model always chooses its most likely next word. Higher values make less likely words more probable. The main set of prompts uses a sampling temperature of 0.7. The temperature check re-collects the primary cell at the temperatures of 0 and 1.0 for five countries (the United States, Germany, Sweden, Japan, and Brazil). Three models do not accept a temperature setting (Section 3.4), so for them the bookend cells only repeat the baseline, and I estimate the check on the other ten. Appendix Table A15, Panel A shows that temperature does not affect the results. The pooled mean equity share shifts by 0.5 percentage points at either bookend relative to the 0.7 baseline (wild bootstrap \(p = 0.40\) at temperature 0 and \(p = 0.32\) at 1.0), and the country contrasts barely move. On all thirteen models, the registered specification, the shifts are 0.3 points (\(p = 0.55\) and \(p = 0.40\)).

Numeral-magnitude.

The second robustness check addresses a numeral-magnitude confound. Version B states the stake as a round native-currency figure. The cross-country contrasts compare prompts whose numerals differ by orders of magnitude, such as 7,500,000 yen in Japan against 50,000 dollars in the U.S. baseline cell. If an AI advisor anchored on the size of the number, this could generate spurious country effects. For Sweden, Japan, and Germany I therefore re-collect Version B with the amount stated as $50,000, holding country and language fixed (Appendix Table A15, Panel B). The pooled shift is a tight zero (\(-0.1\) percentage points, wild bootstrap \(p = 0.84\), confidence interval \(-1.6\) to \(+1.4\) points). The one significant per-country shift is Japan, whose seven-digit yen figure raises the recommended share by 2.8 points relative to the dollar statement (\(p = 0.01\)). Japan’s country contrast is small and statistically null with native numerals (\(+1.2\) points) and remains small under the dollar statement (an implied \(-1.6\) points).

Purchasing power adjustment.

The main collection uses $50,000 converted at market exchange rates to the local currency. At purchasing-power parity, the stake is worth about twice the U.S. stake in Brazil, Mexico, Poland, and China, two and a half times in South Africa, and three and a half to four times in Turkey and India. If the models advise more caution on a stake that is large relative to local prices, part of these countries’ contrasts would reflect the size of the stake rather than the country.

On October 29, 2026, I restated the stake in the twenty non-U.S. Version B prompts as the purchasing-power-parity equivalent of $50,000, rounded in each prompt’s own notation, and re-collected the unchanged Version B prompts for all twenty-one countries, with twenty-five repetitions per model and cell. Appendix Table A16 reports the shifts. Cutting the real stake raises the recommended equity share by 1.0 percentage point on average (wild bootstrap \(p<0.001\), confidence interval 0.5 to 1.7 points). Each halving of the real stake adds 1.2 points (\(p<0.001\)). However, the effect is small relative to the baseline contrasts. For example, Brazil’s contrast moves from \(-11.5\) to \(-9.7\) points and Poland’s from \(-8.4\) to \(-6.1\). Stake size therefore accounts for between a sixth and a third of these contrasts. Moreover, the effect is not specific to emerging markets, since Spain’s contrast narrows by a similar 2.1 points.

Leave-one-out mean and vendor fixed effects.

The pre-registered leave-one-model-out sweep re-estimates equation (3) thirteen times, each time without one model. No contrast moves by more than 1.3 percentage points. Six of the seven contrasts that survive the Holm correction survive it in all thirteen samples, with a largest Holm \(p\)-value of 0.027. Replacing the thirteen model fixed effects with six vendor fixed effects, for comparability with Foltyn and Olsson (2026), leaves every country contrast essentially unchanged, although with six clusters the variant supports no formal inference.

Model drift over time.

To examine whether newer models give different results and to check the robustness of the answers over time, I re-asked one English slice of the primary cell on September 19–20, 2026 to a panel of twenty-one models. Fourteen of these models were new to the panel. The other seven carry over from the main panel under unchanged identifiers. Appendix Table A17 reports the results.

The main findings are robust. On the same English slice, model identity accounts for 70 percent of the variance in September against 78 percent in July, and country for 5 percent against 3 percent. Within each model, 77 percent of non-U.S. cells lie below the model’s own U.S. cell, against 67 percent in July. Six of the seven carried-over models return average advice within 2.5 points of July. The fourteen new models give country a larger role, 15 percent of the variance against 43 percent for model identity, but the added variation does not follow the benchmark. Turkey, where the benchmark prescribes the least equity, receives the most. The advice behind an unchanged identifier can also move. Gemini 3 Flash lowers its Brazil recommendation by 9 points, and DeepSeek V4 Flash recommends 8.5 points more equity because OpenRouter sent most of its September queries to hosts that switch off its reasoning step.

Caller location and routing.

Every query went through OpenRouter from Germany, which raises two concerns. First, the models may condition on the caller’s location. Second, the routing of requests between hosts may move the estimates. Note first that the location is the same in every call, and that the APIs do not pass the location to the model. Each query is a single user message. The only text the vendors add is fixed (the current date for OpenAI and a system prompt for xAI). Asked what they know about the user, none of the thirteen models reports a location.

A direct test shows that the level does not move either. I sent the primary prompt for the United States, Germany, Brazil, and Japan to one model from each vendor in four blocks of 240 queries, routed through VPN exits in Germany and the United States. Moving the origin to the United States changes the recommended equity share by 0.4 percentage points. The shift does not differ across prompt countries (\(p = 0.23\)), and each model bills the same number of prompt tokens from both origins.

Translating the Instructions block.

Every prompt in the design wraps the translated question in the same English instructions. Since the models are sent the entire query, they may read the English instructions as a cue to anchor on an English-language answer, which would compress the country differences and understate the language effect. I therefore re-collected the primary cell on October 29, 2026 with the Instructions block rendered in the local language for the fifteen non-English countries. The six English-prompting countries received prompts identical to the main collection.

The main results are robust. The twenty country contrasts estimated on the new collection correlate at 0.93 with those of equation (3), eight of them lie below the United States after the Holm correction, and their range is 11 points against 12 (Appendix Table A16). Translating the instructions lowers the non-English cells by 1.4 points and the identical English cells by 0.6. The response is concentrated in two models. The two Gemini Flash models recommend about four points less equity when the instructions are in the local language, while the other eleven move by less than 1.5 points.

Allowing the models to ask questions.

A human advisor would ask questions before advising, a feature that the design denies the models. This raises the possibility that the uniform advice reflects a uniform lack of information rather than a uniform policy. I therefore re-asked the primary question in English for all twenty-one countries, where I first ask the models to list up to five questions they would want the investor to answer first. I also ask the models to state the direction in which each answer would move their recommendation, and to give the share they would recommend provisionally if the investor answered nothing further.

However, allowing the models to ask questions does not affect the results. Country means in the two arms correlate at 0.94 with a slope of one (Appendix Table A16, last column). Model identity accounts for 82 percent of the variance of the provisional share against 4 percent for country, the same ordering as in the main arm. The one difference is a level shift. Asked for a provisional number, the models recommend about six points less equity. The shift is uniform across countries and concentrated in the models with the highest equity shares.