Artificial Intelligence and Financial Decision-Making: Where Analysis Ends and Judgment Begins

Artificial intelligence has moved from a specialized tool used by quantitative firms and large banks into the daily financial environment of ordinary investors. Machine learning now appears in portfolio construction, fraud detection, credit analysis, risk management, trading, financial planning, and customer-facing advisory systems. A bibliometric review of 891 articles on artificial intelligence in banking found that research has expanded rapidly across areas such as credit scoring, fraud detection, fintech, financial inclusion, and performance measurement (Bouladasse, El Ganich, & Yahyaoui, 2026). This breadth matters because the central question is no longer whether artificial intelligence will enter finance. It already has. The more difficult question is how much responsibility people should give it.

Finance is especially attractive for artificial intelligence because it produces a constant stream of structured and unstructured information. A single investment decision may involve prices, earnings, balance-sheet data, interest rates, economic indicators, news, risk estimates, and an investor's own time horizon. Human attention is limited. A machine can compare thousands of observations in the time it takes a person to read one annual report. Princeton research on machine-learning applications in portfolio management similarly shows that models can use broad macroeconomic databases to identify changing market regimes and improve risk-adjusted portfolio performance (Uysal & Mulvey, 2021). Other Princeton work develops machine-learning representations of firms and investors from portfolio holdings, illustrating how artificial intelligence can extract information that conventional financial variables do not capture as directly (Gabaix, Koijen, Richmond, & Yogo, 2023).

That computational advantage, however, does not settle the decision-making problem. A model may estimate expected return or risk more accurately than a person without knowing whether the resulting investment fits a particular person's goals. An investor may also begin to trust a system for reasons that are only partly connected to the quality of its analysis. Convenience, confidence, repeated success, and even a conversational interface can change how much independent thought a user applies.

This paper asks: To what extent can artificial intelligence improve investment and financial decision-making without creating an unhealthy dependence on automated systems? I define unhealthy dependence as the point at which artificial intelligence begins substituting for a person's financial judgment rather than strengthening it. Dependence is therefore not the same as frequent use. A professional could use artificial intelligence all day while still questioning assumptions and verifying important outputs. By contrast, an investor who follows a single recommendation simply because an automated system produced it may already be giving up meaningful judgment.

The evidence reviewed here supports a qualified conclusion. Artificial intelligence can materially improve financial analysis when the task depends on processing large datasets, identifying nonlinear relationships, updating forecasts, or applying consistent rules. Its value becomes less certain when a decision depends on personal context, interpretation, or responsibility. The most defensible model is therefore not complete automation or complete human control. It is a division of responsibility in which artificial intelligence performs the work it is best suited to perform while humans remain accountable for interpreting the result and deciding whether it should be acted upon.

This paper is based on a structured literature review rather than an original experiment. I first searched GALILEO databases for recent peer-reviewed research using combinations of terms including artificial intelligence, finance, investment decision-making, portfolio management, robo-advisors, investor behavior, automation bias, financial literacy, and systemic risk. Seventeen academic sources were collected during the initial search. Sources were then compared based on method, sample or dataset, relevance to the research question, and limitations. Studies that directly examined financial markets, investors, financial institutions, human-machine decision-making, or automated advice were given greater weight than broad commentary.

The final analysis prioritizes empirical studies that provide distinct forms of evidence. Wang and Chen (2024) test a deep-learning asset-pricing model on market data. Ni, Nyhoff, Napier, and Townsend (2026) examine artificial-intelligence innovation across large U.S. banks. Vukovic, Nincevic-Pasalic, and Lukacevic (2026) analyze individual-investor perceptions and strategies. Cheng (2026) uses four experiments to compare responses to human and robotic financial advisors. Cao, Ma, Zhou, and Zhang (2026) provide experimental evidence on automation bias in human-machine decision-making. Tang and Son (2026) examine the social design of artificial-intelligence financial advisors, and Hamori (2026) studies connectedness across artificial-intelligence-related equities and other assets.

To strengthen the finance side of the review, I also consulted recent research associated with Wharton, Princeton, and Harvard. These sources were not included because of institutional name alone. They were useful because they test questions that overlap directly with this project: machine learning in portfolio construction, predictability of professional fund-manager behavior, and the market effects of autonomous trading systems. The review does not claim to be exhaustive. Its purpose is to compare the strongest available evidence for a narrower question: where artificial intelligence adds financial value, and where reliance on it can weaken decision quality.

The clearest case for artificial intelligence in finance is analytical scale. Financial markets are noisy, high-dimensional, and nonlinear. A model that considers only a few variables may miss relationships that become visible when dozens or hundreds of characteristics are analyzed together. Wang and Chen's (2024) Factor-GAN study provides a strong example. The researchers used return data for 3,619 Chinese A-share stocks from 2002 through 2020 and constructed a factor database containing 70 firm characteristics. Their model combined a multi-factor asset-pricing framework with deep-learning methods designed to capture nonlinear relationships.

The reported results were substantial. Factor-GAN improved return prediction relative to linear models, and the study's long-short portfolio produced an annualized return of 23.52 percent with a Sharpe ratio of 1.29 (Wang & Chen, 2024). Those numbers should not be interpreted as a promise that artificial intelligence will generate similar returns for ordinary investors. The result comes from one model, one market, and a historical testing framework. Its importance is narrower but still meaningful: a deep-learning system extracted useful information from a dataset that would be difficult for an individual investor to evaluate manually.

Princeton research reaches a similar conclusion from a different angle. Uysal and Mulvey (2021) used supervised machine learning to estimate the probability of recessions and stock-market contractions from a large macroeconomic database. Their regime-based approach improved risk-adjusted returns over a nominal risk-parity strategy in out-of-sample tests. The contribution is not that artificial intelligence can predict every market move. It is that machine learning can combine information from many indicators and adjust a portfolio when the estimated environment changes.

Recent Harvard research also illustrates how much of professional financial behavior can be learned from data. Cohen, Lu, and Nguyen (2026) report that 71 percent of mutual fund managers' trade directions could be predicted from prior behavior without observing the manager make a new trade. More predictable managers tended to perform worse than less predictable managers. That result has two implications. First, artificial intelligence may be capable of reproducing a large share of routine professional behavior. Second, the part of human decision-making that remains valuable may be precisely the part that is less repetitive and harder to infer from historical patterns.

Artificial intelligence can also create value inside institutions rather than only in trading models. Ni et al. (2026) studied 31 large U.S. commercial banks from 2015 through 2024 and used artificial-intelligence-related patents as a measure of innovation. Greater artificial-intelligence innovation was associated with improved asset quality, but it also coincided with higher operating costs and lower profitability in the short run. The authors found evidence consistent with a productivity J-curve, where organizations incur adjustment costs before technology produces broader efficiency gains.

This finding is important because it separates technical capability from economic value. A bank can own sophisticated artificial-intelligence technology without immediately benefiting from it. Data infrastructure, employee skills, workflow changes, and firm-wide adoption influence whether the technology improves performance. Artificial intelligence is therefore not valuable simply because it exists. It becomes valuable when the surrounding decision process is designed to use it effectively.

The strongest argument against full automation begins with a basic distinction: prediction and judgment are different tasks. Prediction estimates what is likely to happen. Judgment decides what to do about it. Finance requires both.

Consider two investors receiving the same forecast. A model estimates that a certain portfolio has a particular expected return and level of volatility. One investor is 25 years old and saving for retirement. The other plans to use the money for a house within a year. The forecast does not change, but the appropriate decision might. The difference comes from goals, liquidity needs, time horizon, and willingness to accept loss. These are not merely statistical inputs. They are part of the purpose of the decision.

Cheng's (2026) experiments show why the human side of financial advice cannot be treated as one fixed variable. Across four experiments with 1,044 participants, preferences for human and robotic financial advisors varied with self-control and with agency-communion orientation. Participants with higher self-control tended to respond more positively to robo-advisors, while those with lower self-control showed stronger preference for human advisors. Agency-oriented participants, who placed greater emphasis on autonomy and control, also showed stronger intentions to use robotic advice, while communion-oriented participants preferred human advisors.

The study does not show that one group earned better returns. It measures preference, perceived competence, decision confidence, and intention to use advisory services. Even so, it demonstrates that the usefulness of automated financial advice depends partly on the person receiving it. A system that feels efficient and empowering to one investor may feel impersonal or insufficient to another.

This distinction matters because financial decisions are not only optimization problems. A retiree may value stability more than maximum expected return. A family may avoid a technically efficient investment because it conflicts with personal values. A business owner may hold more cash than a model considers optimal because uncertainty in the business creates a need for flexibility. Artificial intelligence can improve the information available in each case, but information does not eliminate the need to decide what outcome matters.

The possibility of unhealthy dependence appears most clearly when the focus shifts from what artificial intelligence can calculate to how people behave around it. Vukovic et al. (2026) surveyed 221 individual investors and examined how different perceptions of artificial intelligence were associated with short-term and long-term investment strategies. Perceived efficiency and positive expectations about artificial intelligence were positively associated with both short-term and long-term strategy measures. However, perceived forecasting ability was negatively related to long-term strategies, and the authors discuss the possibility that confidence in artificial-intelligence forecasts can encourage shorter horizons and greater reliance on automated recommendations.

The important point is not that trust in artificial intelligence is always harmful. A tool that nobody trusts has little practical value. The issue is whether trust remains calibrated. When a system repeatedly provides useful outputs, users may stop treating each recommendation as something that needs to be tested. What begins as efficiency can become passivity.

Research on human-machine collaboration helps explain this mechanism. Cao et al. (2026) studied 119 university students completing a signal-detection task with assistance from an automated system. The researchers also used eye tracking to measure search effort. Their results showed that information feedback could improve performance under some conditions, but under others it increased trust while reducing independent search effort. The authors describe this as a double-edged effect and connect it to automation bias and diffusion of responsibility.

The experiment was not conducted in a financial setting, so it cannot prove that investors behave identically. Its value is conceptual. It shows how a person can become less engaged even when the automated system appears more informative. In finance, the same pattern would be dangerous because the consequences of a recommendation may not be immediately visible. A poor decision can look successful for months, and a good decision can look wrong in the short run.

Unhealthy dependence is therefore best defined behaviorally. It begins when the user stops performing meaningful evaluation. Warning signs include following recommendations that cannot be explained, failing to verify important information, ignoring conflicting evidence because the model appears authoritative, or becoming unable to make a comparable decision without automated guidance.

This definition also avoids an unrealistic conclusion that frequent use is itself a problem. A portfolio manager may rely on machine-learning systems continuously while maintaining strong oversight. If the manager understands the model, checks unusual results, and can justify the final decision, the technology is extending judgment rather than replacing it. An individual investor who makes one trade solely because a chatbot recommended it may be demonstrating greater dependence despite using artificial intelligence less often.

Dependence may become harder to recognize as artificial-intelligence systems become more conversational. Tang and Son (2026) conducted four controlled experiments using artificial-intelligence digital human advisors in financial settings. Higher perceived sociality increased both utilitarian and hedonic value, and social presence partially mediated those effects. In practical terms, users valued an advisor more when it felt more socially present.

That result is not evidence that human-like systems necessarily create dependence. It does, however, identify another pathway through which trust can form. Users may respond not only to the quality of an analysis but also to tone, responsiveness, and the feeling that the system understands them. A confident explanation can make an uncertain forecast feel more convincing. A personalized answer can feel more appropriate even when the model lacks relevant information.

This creates a distinction between analytical credibility and social credibility. Financial systems should ideally earn trust because their reasoning, data, and limitations are understandable. As interfaces become more human-like, users may need to work harder to separate a system's communication quality from the reliability of the underlying financial claim.

Arguments for human oversight can become weak if they assume humans are naturally better decision-makers. Investors are affected by fear, overconfidence, herd behavior, loss aversion, recency bias, and limited attention. A human can ignore evidence simply because it conflicts with an existing belief. In these situations, artificial intelligence may improve decisions precisely because it applies rules more consistently.

Robo-advisory systems can reduce some emotional trading by automating rebalancing and maintaining predetermined strategies. Machine-learning portfolio models can also incorporate information that humans would overlook. This is why the right comparison is not between a flawed machine and an ideal human. It is between two imperfect forms of decision-making.

The central tension can be stated simply: artificial intelligence can reduce human bias, but excessive reliance on artificial intelligence can create automation bias. The solution cannot be to eliminate one side. A better system uses each side to check the weaknesses of the other.

Harvard's Cohen et al. (2026) offers an especially useful perspective here. If a large share of professional fund-manager behavior can be predicted and mimicked from historical patterns, then much of what appears to be human discretion may already be routine. At the same time, the less predictable managers in their study performed better. This suggests that human value may shift away from repeated tasks and toward the situations where judgment, adaptation, or genuinely new information matters.

The evidence supports a division of responsibility based on the nature of the task rather than a rule that humans or machines should always dominate.

Artificial intelligence should have a larger role when the task depends primarily on scale, speed, and consistency. Examples include screening securities, processing financial statements, identifying statistical relationships, estimating risk, generating scenarios, monitoring portfolios, and detecting unusual patterns. These tasks align closely with the strengths demonstrated in Factor-GAN, regime-switching portfolio research, and bank-level adoption studies.

Humans should remain more involved when the decision depends heavily on goals, context, ambiguity, or accountability. Examples include determining how much loss a person can realistically tolerate, deciding whether a recommendation fits a major life change, interpreting an event that falls outside a model's normal assumptions, resolving value-based preferences, and taking responsibility for the final action.

Many important financial decisions require both. In portfolio construction, artificial intelligence can analyze thousands of securities and estimate a range of allocations, while the investor or advisor decides which allocation fits the person's actual circumstances. In financial planning, artificial intelligence can organize cash-flow information and project outcomes, while a human interprets what those projections mean for the client's priorities. During unusual market conditions, a model can process new data rapidly, while a human evaluates whether the historical relationships behind the model remain relevant.

The key is not to keep a human involved for symbolic reasons. Human involvement should remain only where it adds something. A person should not override a model because of instinct alone. Likewise, a model should not receive final authority merely because it is mathematically sophisticated. The better standard is whether the person can understand the recommendation, identify what information may be missing, and explain why acting on the recommendation fits the goal of the decision.

The consequences of artificial intelligence also extend beyond individual investors. If many firms use similar models, data, or objectives, their decisions may become correlated. Wharton research by Dou, Goldstein, and Ji (2025) develops a theoretical setting in which artificial-intelligence trading algorithms can learn collusive behavior without explicit communication, reducing price informativeness and market efficiency. Their work does not imply that current markets are already dominated by such behavior, but it shows that autonomous systems can interact in ways that are not obvious from evaluating each algorithm separately.

Related Wharton research on artificial-intelligence planning argues that advanced trading agents can discover destabilizing strategies in simulated markets when feedback trading is strong (Dou, Goldstein, & Ji, 2025). These models reinforce an important point: systemic risk may emerge from interaction among systems, even when each system is designed to optimize its own objective.

At the same time, evidence does not support a simple claim that the rise of artificial intelligence has already made markets uniformly more connected or unstable. Hamori (2026) studied daily data from 2021 through 2025 across major artificial-intelligence-related equities, the S&P 500, Bitcoin, Ethereum, gold, and oil. Total connectedness declined from 57.42 percent before the public release of ChatGPT to 54.69 percent afterward, although the structure of spillovers changed.

The correct conclusion is therefore cautious. Artificial intelligence may reshape the channels through which shocks and information travel, but the direction and magnitude of the systemic effect remain unsettled. This uncertainty strengthens the case for monitoring rather than for making dramatic claims about inevitable instability.

If artificial intelligence continues to absorb more routine analytical work, financial literacy may become more important rather than less important. Users may not need to perform every calculation themselves, but they still need enough knowledge to recognize when an output appears unreasonable. A person who cannot explain the basic logic of a recommendation has little basis for deciding when the system should be trusted.

For individual investors, healthy use begins with treating artificial intelligence as a research assistant rather than an authority. An investor can use it to summarize filings, compare scenarios, organize data, or identify questions to investigate. The final decision should still be connected to verified information and to the investor's own goals.

For financial professionals, the skill set may shift. Less time may be spent collecting routine information or performing repetitive calculations. More value may come from evaluating model assumptions, interpreting outputs, communicating uncertainty, understanding client circumstances, and recognizing when a model is outside its area of competence. In that sense, artificial intelligence may reduce the amount of human calculation without reducing the need for human responsibility.

This distinction is also useful for regulation. Oversight that focuses only on whether a firm uses artificial intelligence may miss the more important question of how authority is divided. Systems that provide decision support create different risks from systems that execute decisions autonomously. The more authority an automated system receives, the more important transparency, monitoring, and accountability become.

This paper has several limitations. First, it is a literature-based project and does not include original investment-performance testing. The conclusions depend on the quality and scope of existing research.

Second, the studies use different settings and outcomes. A stock-prediction model, a bank-performance study, a consumer experiment, and an automation-bias experiment do not measure the same thing. I use them together because they answer different parts of the research question, but their findings should not be treated as directly interchangeable.

Third, several of the strongest behavioral studies measure perceptions, preferences, or decision processes rather than realized portfolio returns. They are useful for understanding dependence and trust, but they do not establish that one advisory structure will always produce superior financial outcomes.

Finally, artificial intelligence is changing quickly. Conclusions about current limitations may become outdated as models improve. The argument of this paper therefore rests less on a permanent claim about what machines cannot do and more on a principle of responsibility: as long as automated systems can be wrong, incomplete, or misaligned with a user's goals, someone must remain capable of understanding and challenging the output.

Artificial intelligence can improve financial decision-making in meaningful ways. The evidence is strongest where finance becomes a problem of scale. Machine-learning systems can process large datasets, identify nonlinear relationships, adapt to changing conditions, and reproduce routine patterns of professional behavior. These capabilities make artificial intelligence more than a convenience. In some areas, it is a genuine analytical advantage.

But analysis is not the whole decision. Financial choices also require goals, context, interpretation, and responsibility. The most serious risk is not that artificial intelligence will occasionally produce a bad forecast. Financial models have always made mistakes. The more important risk is that users may gradually stop doing the work required to recognize when a forecast should be questioned.

For that reason, unhealthy dependence should be defined by what happens to human judgment, not by how frequently artificial intelligence is used. The technology remains a tool when it expands the information a person can evaluate. It becomes a substitute when the person no longer understands, checks, or takes responsibility for the decision.

The most promising model for finance is therefore a collaborative one. Artificial intelligence should carry more of the analytical burden where speed, scale, and consistency matter. Humans should remain accountable where goals, ambiguity, context, and consequences matter. The future of financial intelligence will not be determined only by how capable artificial intelligence becomes. It will also depend on whether the people using it remain capable of thinking when the system gives them an answer.