The $50,000 prompt: MIT on AI financial advice
MIT Sloan simulated 1,000 people following AI money advice from age 22 to 89. The models were fine. The prompts weren't, and that gap cost some users $100,000 in retirement wealth.
AI financial advice turns out to be decent, and the interesting part is who it's decent for. MIT Sloan researchers ran an experiment where 1,000 adults wrote their own prompts asking an LLM where to put their money, then simulated what would happen if those people actually followed the answers from age 22 to 89. The advice was mostly sound. The outcomes were not evenly distributed.
I'm commenting on MIT Sloan's write-up of the study, not reproducing it. Go read theirs. What I want to argue is the part that matters for anyone reading this from Sri Lanka: the model wasn't the bottleneck. The prompt was.
🔍 The gap wasn't in the model, it was in the question
The researchers ran the same exercise twice: once with prompts real people wrote, and once with "academic prompts" that spelled out life expectancy, employment risk, tax assumptions, and expected returns. Same models — GPT-5.2, GPT-5.6, and Gemini 3 Flash. Very different results.
Then they broke down the outcomes by who was asking:
| Group | Simulated shortfall vs. best outcome |
|---|---|
| Women and less financially literate users | ~$50,000 lower wealth by age 60 (≈4%) |
| Users with no prior AI experience | ~$100,000 less retirement wealth (≈6%) |
The finding that should stop you: roughly two-thirds of the gender gap came from how the prompt was worded, and only about one-third from bias inside the model itself.
Key takeaway: The single largest source of bad AI financial advice in this study was not the model hallucinating. It was users under-specifying their own situation. That is a fixable problem, and it's fixable for free.
That reframes the whole "can I trust AI with money questions" debate. Most of the risk isn't sitting in the weights. It's sitting in the 14-word question you typed at 11pm.
🇱🇰 A Sri Lankan prompt starts one step further behind
Here's my extension of the finding, and it's mine, not MIT's. If a vague prompt costs a US user 4-6% of lifetime wealth, a vague prompt from Colombo or Kandy costs more — because the model's default assumptions are wrong for you in ways they aren't wrong for a Boston user.
Ask a bare question and the model quietly fills in blanks that don't apply here:
- Retirement vehicle. Its default is a 401(k) or IRA. Yours is EPF and ETF, with different contribution mechanics and a different withdrawal age.
- Tax treatment. It assumes US brackets and capital-gains rules. Sri Lankan APIT and investment income are taxed on a completely different schedule.
- Currency risk. A US saver's home currency is the one they earn and spend in. If you earn in USD and spend in LKR, or the reverse, exchange-rate exposure is a primary variable and the model will treat it as a footnote.
- Instrument availability. "Buy a low-cost total-market index fund" is fine advice that ignores what you can actually access from an LKR account.
- The risk-free baseline. In the US, the default safe asset is a Treasury yielding low single digits. Here, a bank fixed deposit is a completely different anchor, and it changes what "aggressive" even means.
None of this makes the model useless. It means the local context has to come from you, in the prompt, every time.
🛠️ What a properly loaded prompt looks like
The study's "academic prompts" worked because they removed guesswork. You can copy that structure without an economics PhD. Here's the shape I'd use:
Context:
- Age 27, software engineer in Sri Lanka, earn LKR 350,000/month gross.
- Employer EPF/ETF contributions active. APIT deducted at source.
- Emergency fund: 2 months of expenses. No debt except a
vehicle loan, LKR 18,000/month, 3 years remaining.
- I can invest LKR 40,000/month.
Assumptions I want you to use (state them back to me):
- Plan to age 85. Assume LKR depreciation risk exists; say what
rate you're assuming and why.
- Job loss probability is non-trivial in my industry.
- I want to compare: fixed deposits, unit trusts, USD-denominated
assets, and paying the loan down early.
Output:
- Recommend an allocation, show the arithmetic, and tell me which
assumption the answer is most sensitive to.
The last line is the one people skip and it's the most valuable. Forcing the model to name its own load-bearing assumption tells you where the advice breaks.
| Prompt style | What the model has to guess | Advice quality |
|---|---|---|
| "Where should I invest 50,000 rupees?" | Age, income, debt, taxes, horizon, currency, risk appetite | Generic rules of thumb |
| Above + your age and income | Debt, taxes, horizon, currency | Better, still US-shaped defaults |
| Full context + explicit assumptions | Almost nothing | Comparable to the study's best runs |
Then check the arithmetic yourself. That's what our compound interest calculator, EPF calculator, and fixed deposit calculator are for. An LLM is good at structuring a decision and mediocre at arithmetic over 30 compounding periods.
⚠️ Where the advice was still wrong
Better prompting didn't fix everything. The study found consistent weaknesses across models:
- Job loss handling was poor. The advice pushed excessive spending cuts rather than sensible drawdown of a buffer.
- Insufficient rebalancing. Portfolios were set and left, not maintained.
- Over-reliance on rules of thumb. Rounded heuristics stood in for actual optimisation.
On the credit side, the advice genuinely encouraged higher savings rates, more stock market participation, well-diversified portfolios, age-appropriate risk reduction after 45, and real savings buffers for people over 30. Researcher Taha Choukhmane said they "were somewhat surprised by how good the advice was."
Note the shape of those failures. Every one of them is a dynamic mistake — reacting to a shock, adjusting over time. The models were fine at the one-off allocation question and weak at the "what do I do when things change" question. Which is exactly the part a human adviser earns their fee on.
💡 What this means for you
If you're a student, freelancer, or a two-person team here, you were never going to hire a financial adviser at Western hourly rates. An LLM is what you have. The study says that's a reasonable starting point, with conditions:
- Write the prompt like a brief, not a text message. Age, income, debts, horizon, currency, and what you're comparing against. Every time.
- Force the model to state its assumptions back to you. If it assumes 8% real returns and US tax law, you want to see that before you act on it.
- Give it Sri Lankan context explicitly. EPF/ETF, APIT, LKR exposure, what instruments you can actually buy.
- Verify the numbers in a calculator. Structure from the model, arithmetic from a tool you trust.
- Don't trust it on shocks. Job loss, a currency move, a medical bill — that's where it was measurably weakest.
The deeper point is uncomfortable. When the quality of a free tool depends on how well you interrogate it, the tool quietly rewards people who already know what to ask. That's not a new inequality, but it's a newly measurable one, and the fix is unusually cheap: type more.