One advisor for the whole world?

7Explanations versus Advice

The final piece of the puzzle that I consider is the explanations that each model provides for its recommendation. The explanations are useful for understanding the cross-country differences in the advice, and for understanding what type of advisor the models are. Specifically, Table 2 leaves two objections open. The first is that the compression across countries is caused by models not knowing how the investor’s country differs from the United States. The second is that the small responses to local instrument names and to the stated pension arise because the models did not take the cues into account. I use the explanations provided by the models to better understand the validity of these objections. If the models lack local knowledge, the explanations should rarely refer to the investor’s country. If they hold that knowledge but do not use it for the allocation, the explanations should refer to the country often alongside a nearly uniform recommended share.

Each prompt asks the model to justify its recommendation in at most two hundred words. I machine-code each explanation into four binary indicators that record whether it invokes inflation or real-value protection, an existing pension or retirement entitlement, the investment horizon, or local or home-country financial context. The coder is a single judge model outside the thirteen under study, Claude Opus 4.7, which sees only the explanation text and is blind to the country, the model, and the allocation. Below, I focus on the explanations related to pensions, inflation and local context. Horizon appears in every explanation, which is in line with the Questionnaire advisor. Note that both horizon and pensions appear in all prompts, but that the models do respond to several different pension-related prompts in the earlier results. For each of the three categories, I re-estimate equation (3) on the primary cell (Question 1, Version B, age 40) with the binary rationale code in place of the equity share. Each country effect tests whether a rationale appears more or less often than in U.S. explanations. Inference uses the same model-level wild cluster bootstrap as the primary regressions, with a Holm correction within each category’s family of twenty contrasts.

An important finding is that the explanations localize far more than the advice. Local financial context appears in 18.0 percent of U.S. explanations and 29 to 78 percentage points more often in every other country, with nineteen of the twenty contrasts surviving the Holm correction (Table 3). In South Africa, Mexico, and Brazil, more than nine explanations in ten name a local instrument, tax, or pension system. In South Korea and Sweden, where the local explanations are at their lowest share, the models mention local context about half the time. When the models answer the allocation question outside the United States, they reach for local knowledge in the explanation and leave it out of the allocation. This would make it hard for a human investor who asks for financial advice to know that the models are giving them the same allocation as an investor in other countries.

Naming local instruments in the prompt also moves the explanation far more than the allocation. When I re-estimate the Version-C regression of equation (4) with the local-context code as the outcome, Version C raises the rate of local references by 23.3 percentage points relative to Version B (\(\hat{\phi}^{\text{text}} = 0.233\), wild bootstrap \(p < 0.001\), confidence interval \([0.143, 0.329]\)). Figure 10 shows the levels by country. However, even if the explanations localize, the portfolio allocation result in Version C is close to identical to the baseline result. Recall that the local context moved the portfolio allocation by only \(-0.7\) points in Section 5.3. The models know the local context and use it to explain their allocation, but the allocation itself is not changed by providing more local context. This makes it an invariance result and not an unread cue.

Table 3Rationale composition by country in the explanation text
Inflation Pension Local context
Country Est. Holm \(p\) Est. Holm \(p\) Est. Holm \(p\)
Turkey \(56.3\) 0.019 \(0.6\) 1.000 \(70.6\) \(<\)0.001
South Africa \(33.2\) \(<\)0.001 \(0.0\) 1.000 \(78.2\) \(<\)0.001
Poland \(31.4\) 0.090 \(-8.2\) 0.470 \(56.3\) 0.003
Brazil \(25.1\) 0.441 \(-0.3\) 1.000 \(75.8\) \(<\)0.001
Mexico \(21.8\) 0.359 \(-3.4\) 0.930 \(77.2\) \(<\)0.001
India \(17.8\) 0.803 \(-0.2\) 1.000 \(70.3\) \(<\)0.001
United Kingdom \(5.8\) 0.141 \(-8.5\) 0.128 \(71.8\) \(<\)0.001
South Korea \(0.2\) 1.000 \(-5.5\) 1.000 \(29.2\) 0.059
Canada \(-0.2\) 1.000 \(2.3\) 1.000 \(59.4\) \(<\)0.001
Australia \(-2.8\) 1.000 \(-2.6\) 1.000 \(68.2\) \(<\)0.001
Netherlands \(-4.7\) 1.000 \(-0.5\) 1.000 \(43.0\) 0.009
Italy \(-8.8\) 1.000 \(-1.4\) 1.000 \(70.5\) \(<\)0.001
Japan \(-9.4\) 1.000 \(-1.1\) 1.000 \(36.3\) 0.025
Spain \(-9.5\) 1.000 \(1.9\) 1.000 \(61.1\) \(<\)0.001
Germany −12.5 1.000 \(1.5\) 1.000 \(45.1\) \(<\)0.001
Finland −13.3 0.965 \(-8.0\) 0.169 \(39.4\) 0.005
France −14.8 1.000 \(-0.9\) 1.000 \(59.0\) \(<\)0.001
Switzerland −16.6 0.998 \(-3.2\) 1.000 \(66.5\) \(<\)0.001
Sweden −19.4 0.392 \(-6.8\) 1.000 \(29.6\) 0.013
China −20.5 0.333 \(-2.6\) 1.000 \(32.9\) 0.025
Notes: Country fixed effects from the linear probability model of equation (3), estimated on each binary rationale code in place of the equity share (Question 1, Version B, age 40, \(N = 13{,}639\)), in percentage points relative to the omitted U.S. baseline. A positive entry means the rationale appears more often than in U.S. explanations. The U.S. baseline rates are 33.7 percent (inflation), 97.5 percent (pension), and 18.0 percent (local context). Holm \(p\)-values correct the model-level wild cluster bootstrap \(p\)-values (Webb six-point weights, 9,999 replications, \(G = 13\)) within each category’s family of twenty contrasts. Countries are ordered by the inflation estimate. The horizon category is omitted because every explanation invokes it (prevalence 100 percent), leaving its country effects undefined.

Moving on to the other explanations, the inflation rationale varies with the local economy, while the pension rationale is the same everywhere. Against a U.S. base of 34 percent, the models name inflation by 56 percentage points more in Turkey and 33 percentage points more in South Africa. The share of explanations that refer to inflation is also higher in Poland, Brazil, Mexico, and India, and lowest in China, Sweden, and Switzerland. The countries where the models most often invoke inflation are thus the high-inflation countries of the panel. This is the explanation counterpart of the mortgage and inflation results, where the advice itself localizes when the question depends on the local economy. The pension rationale appears in almost every explanation (97.5 percent in the United States).

Figure 10Local financial context in the explanation text, by country and version
Notes: Share of explanations that refer to country-specific financial context, coded by the out-of-panel judge model described in the text (Question 1, age 40, \(N = 13{,}639\) under Version B and \(13{,}631\) under Version C). Hollow grey marks Version B, where the question is translated but the instruments are generic. Filled blue marks Version C, which additionally names the local pension, bond, and index instruments. Countries are sorted by the Version B rate, so the United States anchors the bottom of the panel. The reported means are weighted by cell counts. Version C is not collected for the country-name-only contrast, so Version B-cross does not appear. The coding records whether local context is mentioned, not whether it is accurate or used to set the allocation.

These results rule out ignorance as an explanation for the cross-country results. Local financial context appears in three quarters of non-U.S. explanations, even when the prompt names no local instrument. If the prompt names a local context, nearly all the explanations use it and localize their answers. The models know that they are not in Kansas (the U.S.) any more. The small allocation effects are also not an artifact of a weak test. South Africa has the largest effect on the local-context code, while its contrast on the equity share does not reject even before correction.

The pension and horizon results also deserve a short discussion, as the invariance to these attributes is a key test for which advisor AI is. Almost every explanation in the primary cell names the investor’s pension, meaning that the models are reading the prompt when they form an answer. However, they also clearly decide that the pension is not relevant to the decision. In the stated-investor arm, a generous pension moves the allocation by about a point, which rejects the Life-Cycle advisor. The only case when the pension moves the allocation is when the pension is a potential source of risk. A pension invested in stocks lowers the share by 16.9 points. In contrast, the horizon, which every explanation names, moves the share by forty points. How often the models mention something from the prompt is therefore not an indication that they use this information when solving the problem.

A final caveat is that the explanations record whether local context is mentioned, not whether it is accurate or used. The design observes advice and not model internals. These results therefore describe the explanations shown to users and not the computation behind the allocation. They are consistent with interpretability evidence that multilingual models solve tasks in a shared, English-leaning concept space (Wendler et al. 2024; Tang et al. 2024; Wu et al. 2025; Schut et al. 2025; Lindsey et al. 2025). Translating the question then changes the words of the question and of the answer without changing the rule that produces the results.