A comparison of professional financial advisers and large language models found an awkward trade-off: the AI systems showed less tendency to project an adviser’s own investment preferences onto a client, but they were also generally more conservative and could leave substantial long-term returns on the table. The peer-reviewed study, published in the Journal of Corporate Finance, compared portfolio recommendations from human advisers with recommendations generated by large language models under different prompts.
The result is a useful warning against reducing the debate to whether AI is better or worse than a person. Advice can fail in different ways. A human adviser can bring experience, context and judgement, but can also be influenced by personal preferences. A model can apply a prompt consistently and avoid some forms of personal projection, while still choosing allocations that are too cautious for a particular client or reacting strongly to the wording of the instructions.
Human advisers projected more of themselves into the portfolio
The researchers describe a vignette-based experiment in which professional advisers and generative AI systems were asked to recommend portfolios for hypothetical clients. A central finding was what the authors call projection: human advisers were more likely to recommend portfolios resembling their own investments. The King’s College London research record confirms that the article is peer reviewed and published in the September 2026 volume of the journal.
Projection is not automatically irrational. An adviser’s own portfolio may reflect genuine expertise and beliefs about risk. The concern is that a client with different goals, age, income, time horizon or tolerance for losses should not simply inherit the adviser’s personal preferences. Personal financial advice is supposed to begin with the client rather than the adviser.
The AI systems were not free from bias either. Their behaviour depended on the model family and the prompt. The study reports that ChatGPT showed the least projection in the tested settings, while a deliberately biased Gemini prompt produced stronger projection that largely disappeared when adviser demographic information was removed. That is an important clue: some apparent model bias may be created or amplified by the context placed in the prompt.
Less projection did not mean higher returns
The models were generally more conservative than the professional advisers. According to the paper, the recommended portfolios had lower risk-adjusted returns and could produce up to 18% lower terminal wealth over a 20-year horizon in the scenarios studied. That figure is a modelled outcome based on the recommended allocations and assumptions used by the researchers, not a prediction that somebody following AI advice will lose 18% compared with a human adviser.
Conservatism can be desirable for a client who genuinely has a low tolerance for risk. It becomes a problem when caution is automatic rather than personalised. Over long periods, small changes in the proportion held in growth assets can compound into large differences in final wealth. A recommendation that feels safer in the short term can therefore carry a hidden opportunity cost.
Fees changed the comparison again
The researchers also modelled the effect of human advisory fees. They calculated a 20-year break-even fee of 1.03% per year in their comparison. In other words, the higher expected terminal wealth associated with the human recommendations could be eroded by ongoing fees at around that level. That does not mean 1.03% is a universal fair fee, because real advice includes services beyond portfolio selection and client circumstances differ widely.
This is where AI could become useful without replacing professional advice. LiveAIWire has covered the expansion of AI into money decisions, including AI in small-business lending and the risk of people treating AI predictions as more certain than they are. A low-cost model might be valuable as a second opinion, a consistency check or a way to explore alternatives before a person makes a decision.
A prompt is part of the advice system
One of the most practical findings is that prompt design affected model behaviour. That matters because a consumer asking a chatbot for help is not using the same system as a regulated advice firm with a carefully designed questionnaire, constraints and review process. The model name alone does not determine the recommendation. The information supplied, the system instructions, the order of questions and the way risk is described can all influence the result.
This also creates a governance problem. A firm deploying AI-assisted advice needs to test the whole interaction, not only the underlying model. If one version of a prompt produces materially different asset allocations from another, the organisation needs to know why and whether the change is appropriate for clients.
The study is not a trial of real client outcomes
The research compares recommendations in experimental scenarios. It does not follow clients for 20 years, measure actual realised investment returns or capture every part of the adviser-client relationship. Professional advisers can discuss tax, family circumstances, behaviour during market falls and changing life goals. A chatbot can be available instantly and at low cost, but its output can be sensitive to framing and may not recognise missing information unless the system is designed to ask for it.
The models tested are also a snapshot in time. AI systems change quickly, and providers alter both underlying models and the consumer interfaces wrapped around them. A result from one model family or prompt should therefore not be treated as a permanent ranking.
What a sensible role for AI could look like
The study points towards a hybrid use that is more interesting than the usual replacement question. AI can challenge an adviser’s first instinct, show how a recommendation changes under different assumptions and make basic portfolio analysis cheaper. A professional can then apply context, regulation and responsibility to the decision. In the opposite direction, a consumer using AI first could bring the output to a human adviser and ask why the two recommendations differ.
That kind of comparison may be particularly valuable because financial decisions are vulnerable to both overconfidence and persuasive presentation. LiveAIWire has also examined systemic concerns about AI in finance. The safest conclusion from this study is not that the machine wins or the adviser wins. It is that each can introduce a different pattern of bias and cost, and those differences can be measured.
For consumers, the headline lesson is simple: an AI recommendation that looks neutral may still be shaped by the prompt and may be more cautious than expected. A human recommendation that feels personalised may still contain the adviser’s own preferences. Good advice needs a process that can detect both problems rather than assuming either source is automatically objective.
What to ask before following either recommendation
The study points to a useful discipline for anyone comparing human and AI financial advice: ask for the assumptions rather than focusing only on the final portfolio. A recommendation can change substantially with the investor’s time horizon, capacity for loss, tax position, expected withdrawals, other assets and need for emergency cash. An AI system may not know those facts unless they are supplied clearly, while a human adviser can also make assumptions that the client never notices. Putting the inputs on the table makes both forms of advice easier to challenge.
Costs deserve the same treatment. The paper’s fee comparison illustrates how a recurring percentage charge can compound over a long period, but a single break-even figure is not a universal price at which human advice becomes poor value. Advice can include tax planning, behavioural coaching, estate considerations and ongoing changes that are not represented by a simple investment-allocation exercise. Equally, a free or cheap AI answer is not automatically good value if it produces a portfolio that is unsuitable for the person using it.
Consumers can also stress-test the recommendation. What happens if returns are weaker, inflation is higher, the investor needs money earlier or the chosen asset mix falls sharply? A robust process should make those trade-offs visible rather than present one allocation as inevitable. The study is valuable because it shows that different advisers can have systematic tendencies. It does not remove the need to judge the recommendation against the individual’s real circumstances.
Regulation remains important too. A general-purpose language model is not automatically a regulated financial adviser simply because it can produce convincing financial explanations. Where a decision has material consequences, users should understand what service they are receiving, who is responsible for it and what protections apply. The most useful role for AI may often be to help people interrogate advice more effectively, compare assumptions and arrive at a professional conversation better prepared.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.