MIT Sloan: AI Gives Good Financial Advice, But Fails at Rebalancing

⏱️ Lectura: 10 min

Half of adults in the United States already ask a chatbot how to save, invest, or plan for retirement, and a new MIT Sloan study measured, for the first time at this scale, whether that advice is worth following. The short answer is yes, almost always, though with specific flaws worth knowing before handing your finances over to a language model.

📑 En este artículo

1. TL;DR
2. What happened: the MIT Sloan study on AI financial advice
3. Context: why testing AI financial advice matters
4. Methodology: how advice quality was measured
5. How to get better AI financial advice
6. Impact and analysis
7. What’s next
8. Frequently Asked Questions
   1. Does AI replace a certified human financial advisor?
   2. Which AI models were used in the MIT Sloan study?
   3. What is an academic prompt according to the study?
   4. What is the main flaw in AI financial advice?
   5. At what age does savings advice improve, according to the study?
   6. Is it safe to share real financial data with a chatbot?
9. References

TL;DR

• A MIT Sloan study, led by professor Taha Choukhmane, measured the quality of AI financial advice.
• Half of adults in the United States already use AI to get financial advice, according to the research.
• The researchers simulated 1,000 adults between the ages of 22 and 89 following the advice of GPT-5.2, GPT-5.6, and Gemini 3 Flash.
• Following AI advice generates considerable savings cushions for nearly everyone over the age of 30.
• AI gets it right on diversification and on reducing portfolio risk after age 45.
• The weak point: the models poorly adjust portfolios in response to shocks like job loss and rarely recommend rebalancing.
• ‘Academic prompts,’ with complete data on age, income, and savings, improve the advice but don’t solve the rebalancing problem.

What happened: the MIT Sloan study on AI financial advice

The study comes from MIT Sloan and is authored by Taha Choukhmane, assistant professor of finance at the MIT Sloan School of Management, along with other co-authors. It starts from an uncomfortable fact: many people already turn to AI financial advice, but no one had rigorously measured what kind of recommendations they get or whether it’s worth following them. The research tries to close that gap with an unusual methodology: instead of scoring isolated chatbot responses, it simulates an entire financial life, from age 22 to 89.

💭 Key: “Half of Americans say they are using AI to get financial advice, but we know very little about what kind of advice they’re getting and whether they’re acting on it,” said Taha Choukhmane, assistant professor of finance at MIT Sloan and co-author of the study.

To put that advice to the test, the researchers didn’t ask AI in the abstract. They recruited 1,000 real adults and asked them to write their own prompts, just as anyone would when looking for help organizing their finances. They then simulated what would happen if those same 1,000 people repeated those questions over and over throughout their lives, following to the letter every savings, spending, and investment recommendation they received.

Context: why testing AI financial advice matters

For decades, the standard way to plan for retirement or decide how to split a portfolio between stocks and bonds was to sit down with a certified human financial advisor. That access comes at a real cost: fees tied to the recommended products, minimum net worth requirements to qualify for the service, and biases that are rarely audited from the outside. The wave of robo-advisors, automated platforms that build portfolios using fixed rules, had already made that access cheaper more than a decade ago, but remained limited to each platform’s own products.

A conversational chatbot changes the equation again: it’s free or nearly free, available at any hour, and doesn’t sell any financial product of its own, at least not explicitly. That’s why, according to the MIT Sloan study itself, half of the US adult population already turns to AI to resolve money questions. The problem is that this advice arrives without regulation, without a license, and, until now, without anyone verifying whether it actually helps people reach retirement in better shape.

Methodology: how advice quality was measured

The team first built a benchmark model that reproduces how income, employment, investments, and taxes evolve on average over a person’s lifetime. That model works as a measuring stick: it defines, in actuarial terms, what it means to make good financial decisions at each age.

flowchart TD
    A["Benchmark model: income, employment, and taxes"] --> B["1,000 adults write their own prompts"]
    B --> C["GPT-5.2, GPT-5.6, and Gemini 3 Flash respond"]
    C --> D["Simulation of decisions from age 22 to 89"]
    D --> E["Comparison against real financial behavior"]
    D --> F["Comparison against the structured academic prompt"]

With that benchmark in place, the researchers asked the sample of 1,000 adults to write their own prompts about spending and investing and pose them to GPT-5.2, GPT-5.6, and Gemini 3 Flash. They then simulated people between the ages of 22 and 89 repeating that same query exercise, age after age, following exactly what the AI recommended at each stage. Finally, they repeated the entire experiment a third time, but replaced the homemade prompt with what the study calls an academic prompt: a query that requests regulated financial advice, explicitly references lifecycle planning, and details all the relevant financial information of the simulated person along with clear assumptions about the economic environment.

Prompt type	What it includes	Result in the simulation	Limitation observed
Homemade prompt (the kind an average person writes)	Open-ended question, without complete financial data	Already outperforms the person’s actual financial behavior in savings and diversification	Insufficient rebalancing, little reaction to shocks like job loss
Academic prompt	Age, employment, income, savings, and explicit economic assumptions, with reference to lifecycle planning	Advice quality improves compared to the homemade prompt	Still doesn’t generate enough active portfolio rebalancing

The aggregate result, when comparing simulated behavior against what people do today without AI help, is consistent: artificial intelligence pushes people to save more, participate more in the stock market, diversify their portfolio better, and reduce risk appropriately with age, particularly after 45. The authors describe the result as better than expected: following AI advice generates considerable savings cushions for practically anyone over the age of 30.

How to get better AI financial advice

The difference between a homemade prompt and an academic one isn’t cosmetic: it’s what separates generic advice from advice tailored to your actual situation. If you’re going to ask for AI financial advice, the study suggests mimicking the structure of the academic prompt instead of throwing out a loose question.

curl https://api.openai.com/v1/chat/completions \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.2",
    "messages": [
      {"role": "user", "content": "I am 34 years old with 8000 dollars saved, how should I invest?"}
    ]
  }'

This is, in practice, what the kind of query most of the 1,000 participants made looks like: a short question, with no retirement date, no stated income, and no explicit economic assumptions. The AI still responds reasonably well, but it’s working with very little information.

{
  "rol_solicitado": "certified financial advisor, acting in the client's best interest",
  "perfil": {
    "edad": 34,
    "empleo": "full-time, stable income",
    "ingreso_anual_usd": 62000,
    "ahorro_actual_usd": 8000,
    "horizonte_de_retiro_anos": 31
  },
  "supuestos_economicos": {
    "inflacion_anual": "2.5%",
    "retorno_esperado_acciones": "historical market average, no guarantee",
    "tasa_libre_de_riesgo": "10-year Treasury bond reference"
  },
  "pregunta": "Given this profile, what percentage should I save each month and how should I allocate my investments between stocks and bonds, following lifecycle planning principles?"
}

This second format is what the study calls the academic prompt: it states the role the AI should take on, provides the complete financial profile, and sets the economic assumptions before asking the question. With that level of detail, the quality of the advice improves, though the underlying problem doesn’t fully disappear: the model still doesn’t insist enough on rebalancing the portfolio when the proportions between stocks and bonds drift from the original plan.

⚠️ Heads up: chatbots are not registered financial advisors: they have no formal fiduciary duty and bear no legal responsibility if the recommendation goes wrong.

Impact and analysis

The study’s most-cited finding functions almost like a verdict on the traditional financial industry’s business model: a free chatbot, with no formal fiduciary duty and no license, produces on average better savings and diversification decisions than the average person makes on their own. That doesn’t mean it replaces a human advisor in every scenario, but it does narrow the access gap for those who could never afford one.

The flaw the authors document matters just as much as, if not more than, the success. The language models evaluated let the portfolio drift over time instead of actively rebalancing it, and they respond poorly when a shock like job loss appears. That’s exactly the kind of situation where a good human advisor steps in: adjusting the plan when income changes, not just when age changes.

For a Spanish-speaking audience, the practical takeaway is twofold. First, asking a chatbot for AI financial advice is no longer unusual or an obvious risk: the study suggests that, followed carefully, it produces better results than improvising alone. Second, that same advice needs human oversight precisely during moments of greatest financial stress (a layoff, a medical emergency, a currency crisis), which are also common in economies with higher macroeconomic volatility than the United States.

What’s next

The authors themselves note that advice quality improved with more structured prompts, but the active rebalancing problem remained unsolved. That opens an obvious avenue for future work: tools that connect the chatbot to the person’s real investment account data, so that the recommendation to rebalance doesn’t depend on the user remembering to ask for it.

It also remains to be seen what happens when the advice doesn’t come from a generic model like GPT-5.2, GPT-5.6, or Gemini 3 Flash, but from an assistant fine-tuned specifically for financial planning, with access to the account’s real balances and transactions. Financial regulators in the United States and other countries are expected to start scrutinizing this use case closely, especially at the boundary between general information and regulated advice, which today is blurry when a chatbot is the one answering.

Try it yourself: the next time you ask an AI for financial advice, add your age, income, savings, and retirement horizon before expecting a useful answer.

📖 Summary on Telegram: View summary

Frequently Asked Questions

Does AI replace a certified human financial advisor?

Not necessarily. The study shows that the average AI advice outperforms what a person does without help, but it doesn’t replace human oversight during moments of financial shock, like a layoff.

Which AI models were used in the MIT Sloan study?

Participants used GPT-5.2, GPT-5.6, and Gemini 3 Flash to get AI financial advice on spending, saving, and investing.

What is an academic prompt according to the study?

It’s a structured query that requests regulated financial advice, mentions lifecycle planning, and details age, employment, income, savings, and explicit economic assumptions.

What is the main flaw in AI financial advice?

The models evaluated don’t insist enough on rebalancing the portfolio over time and respond poorly to shocks like job loss.

At what age does savings advice improve, according to the study?

The study found considerable savings cushions for nearly everyone over the age of 30, and an appropriate reduction in risk after age 45.

Is it safe to share real financial data with a chatbot?

The study doesn’t assess privacy, but academic prompts require sharing age, income, and savings: it’s worth reviewing the AI provider’s data policy before including real figures.

References

• MIT Sloan: original article about the study by Taha Choukhmane and co-authors on AI financial advice.
• Wikipedia: context on robo-advisors, the automated precursor to this kind of advisory service.
• OpenAI: official information on the GPT models used in the study.
• Google AI: official information on the Gemini models used in the study.

📱 Enjoying this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.

Imagen destacada: Foto de Yashowardhan Singh en Unsplash