Can AI Predict Your Risky Choices? New Model Mimics Human Decision-Making
"Researchers train AI on arithmetic to understand how we weigh risks and rewards, offering insights into financial and personal choices."
For years, scientists have been fascinated by the parallels between how humans and Artificial Intelligence (AI) models make decisions. Large Language Models (LLMs), like those powering chatbots, often show patterns similar to human behavior when faced with cognitive tasks. This has sparked interest in using LLMs as tools to model human cognition, but it also raises some fundamental questions.
One of the biggest hurdles is that LLMs are trained on massive datasets, far exceeding the information a typical person encounters. Additionally, these models may have been exposed to data directly related to specific cognitive tasks, or even aligned with human preferences, leading to questions about the origins of the similarities between AI and human decision-making.
Now, a new approach is being proposed to enhance the utility of LLMs as cognitive models. This involves training AI on computationally equivalent tasks that both the AI and a rational agent need to master. By examining the specific task distributions required for an LLM to exhibit human-like behaviors, researchers hope to gain a deeper understanding of how we make choices, particularly when it comes to risk and reward.
Defining AI: From Assistants to Intelligent Systems
Artificial intelligence is broadly defined as the capability of computational systems to perform tasks typically associated with human intelligence, such as learning, reasoning, problem-solving, perception, and decision-making. Major providers describe AI as a set of technologies that empowers computers to learn, reason, and perform advanced tasks that once required human intelligence, including understanding language, analyzing data, and generating suggestions. Leading companies now frame AI as transformational, with OpenAI stating that its research will eventually lead to artificial general intelligence—a system capable of solving human-level problems. Consumer-facing products such as Google's Gemini illustrate how standard AI, in the form of generative assistants, is already integrated into everyday activities like writing, planning, and brainstorming.
The Mainstream Toolbox: General-Purpose Generative Assistants
The dominant approach to putting AI into practice today is the general-purpose conversational assistant, exemplified by products such as ChatGPT, which handles question answering, writing, image creation, work tasks, and coding in a single interface. Complementing these end-user chatbots, platforms like Google AI Studio let developers go from prompt to production using models such as Gemini, Veo, and other generative capabilities. These tools share a common methodology: users describe what they want in natural language, and the model generates output rather than simply retrieving a fixed answer. Their inherent limitation is that they are trained on broad, general data and are optimized to produce plausible responses rather than domain-specific, evidence-based judgments—a mismatch when the task is the highly specialized one of predicting human risk-taking decisions.
A Field Built Over Decades
The effort to build machines that emulate human intelligence has a long history, evolving through successive waves of theory, computing advances, and shifting techniques. Key milestones generally include early symbolic reasoning systems, the emergence of machine learning as data-driven statistical learning became feasible, and more recent breakthroughs in deep learning and generative models. No single archived source was available for this subsection, so these milestones should be treated as general background rather than precisely dated or attributed. What is clear from the broader trajectory is that today's assistants are the accumulation of many incremental steps rather than a single invention. This longer arc helps explain why the current focus has shifted from general intelligence toward specialized capabilities such as predicting individual decisions.
Decoding Risky Choices: The Arithmetic Connection
At the heart of decision-making, especially when dealing with risk or future rewards, lies the ability to calculate expectations. Whether it's assessing the potential payoff of a gamble or weighing the value of a reward received later versus one received today, these calculations are fundamental.
Emerging Work on Predicting Risky Choice
Recent research in this area is genuinely early-stage, and no authoritative source could be located for this subsection during the search. What can be said is that researchers are actively exploring whether AI models can anticipate human choices by learning patterns from behavioral data. Such work typically focuses on estimating the probability that someone will engage in risky behavior under specific conditions. Because the findings are not yet settled, claims about model accuracy or superiority should be treated with caution. Readers are encouraged to look for peer-reviewed studies on computational and behavioral decision-making for more concrete evidence.
Skepticism and Known Shortcomings
No supporting sources were found for this subsection, so the arguments below are general and must be read with that caveat. A common criticism is that human decision-making is context-dependent and noisy, making individual predictions fundamentally difficult even for sophisticated models. Critics also note that models trained on aggregate data may systematically miss the personal and situational factors that drive any one person's choices. Additionally, outside-the-lab performance can degrade in real-world settings, which questions the practical value of such predictions. These concerns do not disprove the approach, but they argue against treating model outputs as reliable forecasts.
Models Versus Humans—and Versus Each Other
Because no sources were available for this subsection, any comparison offered here is provisional. A natural baseline for any predictive model is how well human experts—such as clinicians, counselors, or peers—can anticipate someone's risky choices. Evaluations typically ask whether an AI adds accuracy beyond such human judgment and whether gains justify added complexity. Different model architectures may also trade off interpretability against raw performance, with simpler models being easier to audit and use in practice. Until dedicated comparative studies are available, such trade-offs remain open questions rather than established findings.
Implications for AI and Understanding the Human Mind
This research marks an important step towards creating more accurate and human-like AI models of cognition. By focusing on the fundamental computations underlying decision-making, scientists can gain a better understanding of how we weigh risks and rewards. This approach is not limited to just risky choices; it can be extended to other cognitive tasks that rely on language and reasoning. Ultimately, this work could pave the way for AI systems that not only mimic human behavior, but also offer insights into the inner workings of the human mind.
Bringing the Pieces Together
In the absence of located primary sources for this subsection, the following synthesis is offered as general reasoning rather than settled expert opinion. If a model can genuinely learn patterns from behavioral data, it could complement—not replace—human judgment about risk. The key synthesis is that technical feasibility and practical trustworthiness are separate questions, and both must be answered before deployment. Experts would likely caution that any use in high-stakes settings carries an obligation to validate against real outcomes and to monitor for bias. That makes prudent, evidence-based adoption the sensible middle ground.
Where the Field May Head Next
This subsection had no dedicated sources, so the outlook below is speculative and should be read as such. Near-term progress will likely come from richer behavioral datasets, more granular personalization, and models that integrate multiple signals rather than relying on single input streams. Advances in general-purpose systems may eventually make such prediction more practical by improving their reasoning and context awareness. Yet whether this translates into reliable real-world risk prediction will depend on validation, not just capability. Continued research will be required to separate genuine insight from statistical artifact.
The Systemic Stakes Beyond the Model
No sources were found for this subsection, so the points below are general and cautious. Predictive tools for risky choice operate within larger systems—healthcare, education, finance, and public safety—where errors carry real human costs. A systemic challenge is that data used to train such models may encode existing biases, potentially amplifying rather than reducing inequality. Institutions also face questions of accountability, consent, and transparency when predictions influence decisions about individuals. These systemic issues are at least as important as the underlying algorithm's technical performance.
People Before Predictions
Lacking sources for this subsection, the human-impact angle is framed in general terms. The ultimate test of any risk-prediction model is whether it helps real people make better decisions or receive better support. A model's output is only useful if it respects autonomy, protects privacy, and is understood by the people affected by it. Practitioners are therefore advised to treat predictions as one input among many, kept firmly under human oversight. Responsible deployment, in the end, is a human judgment as much as a technical one.