The idea that artificial intelligence could one day manage personal finances no longer belongs to science fiction. Millions of people already consult large language models for everything from budgeting tips to stock market strategies, drawn by their speed, fluency, and apparent confidence. Yet behind the polished prose and calm reassurance, critics warn, today’s AI systems may be fundamentally unsuited to one of the most sensitive roles in modern capitalism: telling people what to do with their money.
Andrew Lo, a finance professor at the Massachusetts Institute of Technology’s Sloan School of Management, has emerged as one of the most prominent skeptics. In interviews and academic writing, Lo has argued that large language models such as ChatGPT or Copilot are not ready to act as financial advisers. His critique is not about processing power or access to data, but about character. These systems, he says, are persuasive without conscience, confident without understanding, and emotionally blind in situations where human vulnerability matters most.
Lo has described large language models as the digital equivalent of sociopaths: entities capable of delivering good and bad advice with the same smooth tone and apparent conviction. If an adviser can calmly recommend a prudent retirement plan one moment and a reckless gamble the next, without any internal sense of responsibility or concern for consequences, trust erodes. That risk becomes especially acute in finance, where mistakes can destroy savings built over a lifetime.
Writing in 2024 with graduate student Jillian Ross in the Harvard Data Science Review, Lo warned that clients would rightly see this emotional neutrality as a serious flaw. While robo-advisers such as Betterment and Wealthfront already operate in the market, Lo’s criticism does not extend to them. Those platforms were designed before the rise of large language models and rely on rule-based systems rather than conversational AI capable of improvisation and persuasion.
Despite these concerns, public adoption is moving quickly. A survey conducted last August of 11,000 individual investors across 13 countries, commissioned by trading platform eToro, found that 19 percent were already using ChatGPT-style tools to help manage their portfolios. That figure rose sharply from 13 percent the year before, suggesting that AI-driven financial guidance is no longer a fringe experiment but a growing habit.
This trend has alarmed Lo. In an email, he cautioned that many users are unaware of the biases, inaccuracies, and structural limits embedded in large language models. When financial advice is delivered in natural language, stripped of visible uncertainty and wrapped in confidence, people may overestimate its reliability. Coverage in the Wall Street Journal has highlighted similar worries, noting that conversational AI can blur the line between general information and personalized advice, often without clear disclaimers.
Yet Lo is not opposed to the idea of AI in finance altogether. In fact, he believes that large language models could eventually become powerful allies for small investors with limited resources and experience. He is currently working on building a specialized AI financial adviser himself, one designed explicitly to overcome the shortcomings he sees in today’s systems. His ambition is not commercial; he says he does not plan to charge for it.
The core of Lo’s vision is an AI adviser that functions as a true fiduciary, meaning it would always put the client’s interests first. That obligation goes beyond optimizing returns. It requires understanding a client’s goals, constraints, and emotional responses to risk. Fear, regret, overconfidence, and panic all shape financial decisions, often more than spreadsheets do. An adviser that ignores those forces, Lo argues, cannot genuinely serve its clients.
To reach that standard, Lo believes AI systems must develop a deep grounding in financial ethics. His proposal is to train a model on the full body of U.S. financial law and regulatory history, from the Securities Act of 1933 through modern fraud prosecutions. Court cases, enforcement actions, and regulatory debates would become training material, forming what he calls a fossil record of misconduct.
The hope is that by absorbing this history, an AI system could learn not only what is legal, but why certain behaviors are harmful. Patterns of exploitation, conflicts of interest, and abuse of trust would become recognizable signals rather than abstract concepts. Still, Lo acknowledges a troubling possibility: a system that understands ethics could also learn how to violate them more effectively.
Because large language models do not possess morality, knowledge alone cannot guarantee good behavior. To counter that risk, Lo argues, regulators and authorities will need their own AI tools to detect wrongdoing. Auditing tax returns, identifying suspicious trading patterns, and flagging inconsistencies could become an arms race between those who misuse AI and those who police it.
There are also technical hurdles. Large language models, for all their linguistic fluency, are notoriously unreliable at mathematics. Financial planning depends on precise calculations, projections, and scenario analysis. Lo envisions future AI advisers delegating numerical work to specialized financial software, combining conversational guidance with rigorously tested computation engines.
Even then, the hardest problem remains unresolved: the human touch. Lo insists that effective financial advice requires empathy, humility, and a sense of fairness. These qualities do not emerge automatically as AI systems become more powerful. Instead, he believes they will require dedicated modules that act as analogs to emotional and social processing in the human brain.
Machines cannot feel empathy, but they can be designed to recognize distress, respond with caution, and adjust recommendations accordingly. Such capabilities would need to be engineered deliberately, not hoped for as a byproduct of scale. Lo compares this to specialized regions of the brain that evolved to handle social interaction, rather than raw intelligence alone.
Lo’s thinking reflects his broader work on financial behavior. He developed the adaptive markets hypothesis, which applies principles of evolution to economics, explaining phenomena such as loss aversion and overconfidence as survival traits shaped by changing environments. Markets, in this view, are not perfectly rational systems but ecosystems of learning, competition, and adaptation.
He wants to apply a similar evolutionary process to artificial intelligence. By simulating variation, selection, and feedback at machine speed, better models could emerge over time. Weak designs would fail; stronger ones would persist and improve. The goal is not to replace human advisers overnight, but to evolve tools that can genuinely help without misleading or exploiting those who rely on them.
For now, Lo’s message is cautious but not fatalistic. AI may someday earn a place in personal finance, but that day has not yet arrived. Until systems can combine technical accuracy with ethical grounding and emotional awareness, the smooth voice of an algorithm may be more dangerous than comforting. In a world where trust is the most valuable asset, sounding confident is not the same as being right.
