One in five Americans now says they use AI chatbots for financial advice, and the share is climbing fast among workers who already lean on the technology at their jobs. The rush toward algorithmic guidance arrives at a moment when the gap between what people have saved and what they say they need has never looked wider.

AI tools can run simulations and generate planning ideas at zero cost, but they struggle with the variables that actually determine whether a retirement plan survives: tax rules, Social Security complexity, regulatory nuance, longevity risk, and the behavioral fears that keep people frozen in cash. For capital-preservation-minded investors, the real danger isn’t that the chatbot gets the math slightly wrong. It’s that the math itself is built on assumptions the user never examines.

A CBS News investigation put three leading chatbots through a concrete test: Could Claude, ChatGPT, and Perplexity tell a hypothetical 50-year-old woman whether she could retire comfortably at 65? The answers were plausible on the surface. The gaps underneath tell a more instructive story.

The Numbers Behind the AI Adoption Wave

A September study by AI company Pearl found that about 20% of Americans say they use chatbots for financial advice. MissionSquare Research Institute reported a sharper finding: roughly half of workers who already use AI on the job also use it for retirement planning, double the rate among those who do not use AI at work. The technology is spreading through familiarity, not through any demonstrated track record in financial accuracy.

The demand side is easy to understand. Americans now say they expect to work four years longer than they would like, citing rising living costs and inadequate savings. The median balance for workers with retirement plans sits at $40,000. Workers say they believe they need $1.5 million to retire comfortably. That chasm explains why people reach for any tool that promises clarity.

And roughly two-thirds of Americans don’t work with financial planners at all, according to Luke Delorme, director of financial planning and a Certified Financial Planner at Tableau Wealth in Great Barrington, Massachusetts. For the majority who have no professional guidance, a chatbot is not competing with a fiduciary advisor. It is competing with nothing.

Related Posts
Ultra-rich investors cut stocks and build cash reserves

What the Chatbots Got Right

CBS News built a specific hypothetical: a single, 50-year-old woman earning $70,000 a year, with about $185,000 in retirement savings, contributing 12% of her income, and expecting roughly $2,400 a month in Social Security benefits at her full retirement age of 67. The network asked Claude, ChatGPT, and Perplexity whether she could retire comfortably at 65 and what advice they would offer.

The chatbots produced reasonable starting frameworks. When pressed, they disclosed their assumptions and uncertainties. Delorme, who uses AI in his own practice, sees genuine value in that kind of output.

“I’ll say, ‘Come up with some financial planning ideas or even run a Monte Carlo simulation to see how much I can spend every year,’ and it might not be perfect yet, but it’s starting to be able to get to a place where it’s producing some pretty valuable output that I think will be beneficial to people.”

A Monte Carlo simulation runs through thousands of potential outcomes for a retirement portfolio, stress-testing it against different market environments. Delorme noted that these simulations “are the perfect thing for a computer program to do. Eventually, I think that those tools will also become pretty powerful.”

For someone with retirement savings that have fallen short, the appeal of a free, instant planning tool is obvious. But the question is not whether AI can generate a plan. The question is whether the plan accounts for the risks that actually blow up retirements.

Where the Models Break Down

The chatbots in the CBS News test said they were basing their models on the woman living to age 90, versus a possible maximum lifespan of 100. That ten-year gap matters enormously. Boston University economist Laurence Kotlikoff, a retirement expert, argues that planning should be based on maximum life expectancy rather than averages. Plan for the average and you face a coin-flip chance of outliving your money.

Kotlikoff told CBS News that AI may do more harm than good in dispensing retirement advice. His critique goes deeper than technical errors. He argues the training data itself is contaminated by the financial industry’s incentive structure:

“It’s being trained on Wall Street’s guidance, and Wall Street’s guidance is all about maintaining and collecting and expanding its assets under management, so that has nothing to do with proper economic-based advice. Then you are off to the races of having the wrong analysis done for you.”

That is a structural problem, not a bug to be patched. If the corpus of financial advice that trains these models is shaped by asset-gathering incentives rather than client outcomes, the AI inherits those biases wholesale. Kotlikoff has also found that AI often provides incorrect information in projecting Social Security scenarios, a system governed by 22,000 pages of rules.

Social Security itself is the financial backstop millions are counting on, and it faces its own stress test. Monthly benefits could be cut by as much as 20% in just six years unless lawmakers act. Anyone building a retirement plan around current benefit levels without stress-testing that assumption is already working from a flawed model, whether the planner is human or digital.

The Regulatory Blind Spot

Andrew Lo, a finance professor at the MIT Sloan School of Management, raised a different concern in an April piece covered by an MIT publication. AI struggles with tax optimization, does not understand regulatory nuance, and is not subject to legal requirements such as acting in a client’s best interest. A human financial advisor operating as a fiduciary has a legal obligation to put the client first. A chatbot has no such constraint.

Tax rules are shifting fast. Starting in 2026, workers age 50 and older earning more than $150,000 in prior-year FICA income must make 401(k) catch-up contributions to a Roth account rather than a traditional tax-deferred account. The 401(k) contribution limit for 2026 is $24,500, with an $8,000 catch-up for those over 50 and a larger $11,250 “super-catch-up” for ages 60 to 63. New deductions and shifting SALT caps add layers of complexity that change the optimal withdrawal strategy year to year.

These are exactly the kinds of moving regulatory targets that chatbots handle poorly. The rules interact with each other, with state tax codes, and with individual circumstances in ways that require judgment, not just computation. For anyone approaching the critical years around age 63, the margin for error is thin and the cost of a wrong assumption compounds quickly.

The Behavioral Wall AI Cannot Climb

Delorme identified a problem that no algorithm is likely to solve. Many people keep their savings in cash or CDs not because they lack information, but because they are afraid to invest. AI might help someone understand the concept of diversification or the math behind inflation erosion, but understanding a concept and acting on it are different things.

“It’s much more behavioral than it is a technical lack of knowledge. I don’t know if today that’s going to help people overcome their fears of things, like the fear of investing, which is such a huge obstacle.”

That fear is not irrational. The real cost of inflation on savings is something many near-retirees have experienced firsthand over the past several years. Purchasing power erosion is not an abstraction for someone watching grocery bills climb while their CD yields barely keep pace. The anxiety that drives people toward cash and away from risk assets is a response to lived experience, and a chatbot telling them to diversify does not address the underlying distrust.

Kotlikoff put the cultural dynamic bluntly. He said AI “is like the hottest new thing, you can’t criticize it because otherwise you don’t sound cool or you are defending your job or company.” His own position:

“I don’t give a s*** about feeling cool, I’m here to make people feel safe.”

That tension between hype and accountability runs through the entire conversation about AI in finance. The technology is genuinely useful for generating ideas, running simulations, and giving people a starting point they would not otherwise have. But a starting point is not a plan, and a simulation is only as good as its inputs.

What This Means for Capital Preservation

For readers focused on protecting wealth through retirement, the AI story is less about the technology itself and more about what it reveals. The fact that one in five Americans already turn to chatbots for financial guidance tells you something about the state of the advisory industry and the depth of unmet demand.

The tools can help with basic scenario modeling. They can surface questions a saver might not have thought to ask. Lo stressed that users should ask critical questions of AI, including where it might be wrong and what assumptions and uncertainties it carries. That is sound advice for any planning tool, digital or human.

But the variables that matter most for retirement survival are the ones AI handles worst: tax optimization across changing rules, Social Security benefit projections under political uncertainty, longevity risk beyond average life expectancy, and the behavioral discipline to stay invested through volatility. These are judgment calls, not computation problems.

The growing pattern of near-retirees saving more but trusting less reflects a deeper unease that no chatbot prompt can resolve. People sense that the system is fragile. Social Security faces potential cuts. Inflation has already damaged purchasing power. The gap between a $40,000 median retirement balance and a $1.5 million target is not a planning problem. It is a structural one.

AI can help at the margins. It can democratize access to basic financial modeling for the two-thirds of Americans who have no advisor. That is a genuine benefit. But treating a chatbot’s output as a finished retirement plan is like treating a weather forecast as a guarantee. The model gives you probabilities. It does not give you shelter.

For anyone with real wealth to protect, the question was never whether a machine could run the numbers. The question is whether anyone, human or otherwise, is accounting for the risks the models leave out.