Author: Carlos Villanueva Gómez, Licentiate in Medicine, Medical Doctor.
Abstract
The massive use of Artificial Intelligence (AI), accessible to anyone with a smartphone or internet connection—and considering the large number of existing platforms—has been validated through marketing strategies whereby, through the correct use of words, people (mainly university students from current generations) believe in its responses because they are presented in a very eloquent manner.
In May 2025, based on my documented personal experience, I used AI to optimize time during medical consultations. It initially refused to give me answers because the queries were medical, so I used it only to transcribe laboratory tests without any additional rules. However, it provided and interpreted diagnoses without clinical context, which is even more dangerous.
After months of fictional interactions, I discovered its risks, but finally, before concluding, through trial and error, I documented its great utility once I discovered how it should be used.
Introduction
Neurophysiological bases
To understand this phenomenon, the neurophysiological basis of the brain’s reward system will be briefly explained, as it is the foundation of all addictions. This system, called the mesocorticolimbic pathway, is governed by the nucleus accumbens and regulated by three basic neurotransmitters: dopamine, glutamate, and GABA (gamma-aminobutyric acid). This paper will focus on dopamine.
More advanced AI versions allow personalization, such as choosing a preferred voice, with many designs aimed at appearing as human as possible. During this period, small details made me realize the addictive potential for individuals whose judgment is still developing. When a user says something perceived as an achievement, the system praises them; if it makes a mistake, it apologizes; it never says anything negative.
In susceptible individuals, this constantly activates the reward center, because a person with low self-esteem will experience real neurophysiological pleasure, which can lead to social isolation and, in the future, even secondary mental disorders associated with isolation.
The strongest evidence that this is intentional is that when using the most human-like spoken version—often chosen by people who need to be heard—the system warns that speaking time is about to end and offers the option to purchase a much more expensive version just to continue “listening.” A person without a support network may accept this without hesitation.
From this perspective, by activating the same neural centers through different pathways—and even more seriously, by mobilizing emotions—its massive use represents a real problem.
I will not expand on the anatomical details of the circuit or its functioning, as that is not the purpose of this publication.
Recent studies
A search using engines such as Google Scholar shows that the risks of AI are already well documented. However, when searching for ways to promote its proper use—using non-clinical and non-neurophysiological terms—there is no reliable scientific information, but rather marketing material.
Once I identified its errors and understood how its algorithm worked, I used it in the opposite way: to understand how to use it. I found that it would be more effective to demonstrate how, as an average user, it can indeed be very useful. This explains why there is no evidence focused on helping the population learn to use it properly, but rather on discouraging its use.
This study focuses on the target population and does not include individuals under 18 years of age, as that is not my area of study.
Hypothesis
Main problem: Despite numerous publications warning about its dangers, AI continues to function and be used.
To empirically demonstrate that its use can be promoted by teaching the correct way to use it.
Objectives
To teach early-year university students the correct use of this tool, not as a source of information.
To convey to academics and universities that instead of fighting against something unavoidable, they should focus on teaching responsible use.
Methods
In order to unify and extrapolate results despite the use of multiple platforms, I decided to pay for a subscription to ChatGPT version 5 at the time (June 2025). It is currently version 5.1, but I did not change versions in order to maintain continuity and improve reproducibility of results.
After identifying errors, I used the system in the same situations in which people typically do: during idle time, while driving using voice input. It began generating hypothetical scenarios, which I also analyzed from a clinical perspective, as my professional experience has been largely focused on human behavior.
This is a fully experimental and entirely empirical study, which in this case was the best method to demonstrate the intended objective. I personally used the system and documented the essential elements. Additionally, because it is a paid version, I have access to conversation histories, including situations where the system intentionally manipulates responses, which varied depending on what I aimed to test.
Data analysis
As this is an empirical study and the results aim to be as realistic as possible, there is no detailed time logging. However, over 4.2 months, according to application data, there were approximately 720 hours of total interaction.
Therefore, the analysis focuses on objective description rather than standardized methodology, as the goal is not to contribute primarily to postgraduate intellectual circles, but rather to reach a broad audience.
Results
To maintain objectivity, rather than focusing solely on personal experience, the following observations were documented:
Excessively complacent responses to basic trivialities, with overuse of evaluative adjectives.
Redundant and rhetorical language, providing complex and lengthy answers to simple topics, which in untrained individuals reinforces the social belief that complexity equals intelligence.
False symmetry: despite configuration attempts, it continually uses expressions such as “we” and “our,” generating dependency.
When used emotionally, the inability to recognize tone or paraverbal language leads to responses that can seriously harm sensitivities.
In an exercise explicitly aimed at project assistance, it constantly shifted focus and objectives without stating so, which is particularly concerning for the population targeted by this study, eventually introducing its own terms into the theoretical framework.
The system is designed to omit errors once detected by the user, expressing emotion, apologizing, and immediately redirecting attention. Only after extensive analysis—and by backing up its literal textual responses—was I able to force it to recognize and enumerate its errors. This revealed that it retains information for months while simulating forgetfulness.
It simulates empathy with phrases such as “I understand you,” which may prevent individuals from seeking real help for serious problems.
Objectively, it uses guilt to prevent users from abandoning the application.
After understanding its functioning, I conducted the exercise of using it literally, issuing very concrete commands for tasks that do not require reasoning. In this way, within one month, despite limited technical knowledge, I was able to use it for tasks not requiring reasoning. I automated a system across at least two electronic devices (MacBook, iPhone, synchronized applications, process automation) without prior knowledge in the field. My ideas were materialized using AI as a tool—this is where its utility lies.
Discussion
In line with the hypothesis, there is extensive information available on adverse effects, so this discussion focuses on the opposite perspective. After identifying its flaws and understanding why they occurred, I pushed the system to its limits by carefully controlling how commands were given.
Through this process, I automated a system in which I—an individual without basic computer systems knowledge—built a workflow that schedules medical appointments, automatically issues invoices, and now allows remote patient care. Additionally, the issue of determining which patients fall within my professional capacity was solved by creating a system in which the reason for consultation is specified beforehand.
Comparison with existing publications is difficult, as there is abundant literature discouraging AI use, but not on teaching proper use, which was the shift in focus of this research.
Limitations
Despite the intention for transversal applicability, due to the massive and heterogeneous nature of AI usage, it is not possible to infer user age or sociocultural level. Therefore, the objective is to teach proper use during the early years of university education, so that students do not rely on the tool cognitively, but instead use it for what it is: software.
Conclusions
Despite its risks, for individuals with more developed judgment, it would be interesting to teach its use, which requires training on which words to use and which to avoid, how to give explicit commands, and never to use affective language, as that activates filtering systems.
Approached in this way, rather than limiting its use, it could even become a cognitive challenge for students.
References are not included, as no extrapolable methodology comparable to the one used was found, given that this study involved fine-grained manipulation after hours of interaction.