The Human Variable

Artificial intelligence tools such as ChatGPT have quickly become part of everyday student life. For many young people, AI is no longer something futuristic or distant; it is already present in homework, writing assignments, studying, translating, and summarizing. At first glance, this seems like a major educational advantage. A student who does not understand a difficult concept can ask ChatGPT for an explanation. A teenager who struggles with writing can ask for help organizing an essay. A university student can save time by asking AI to summarize a long article. In this sense, AI appears to make learning faster, easier, and more accessible.

But what matters more is not only what AI helps young people do, but also what it may stop them from developing. If a student uses ChatGPT to understand a topic, that may support learning. But if the same student uses it to generate answers, arguments, ideas, and conclusions without thinking through the process, then AI may become a shortcut that replaces cognitive effort. This raises a significant concern: youth may appear more productive while becoming less independent in their thinking.

In this context, a study conducted by Nidugonda and Shaik (2025) presents ChatGPT as having a dual influence on youth cognitive development. On the one hand, the authors argue that generative AI tools can support learning by improving access to information, helping students understand complex material, assisting with academic tasks, and strengthening problem-solving processes. On the other hand, they emphasize that excessive or unbalanced use of AI may produce important cognitive drawbacks, including overreliance, diminished critical thinking, reduced originality, and weaker independent decision-making.

This paper argues that AI should not be treated only as a tool that provides students with answers, but as a technology capable of reshaping the way young people think and learn. The danger is not that AI is inherently harmful, but that it is often used in ways that reduce effort, uncertainty, and intellectual struggle. Yet these uncomfortable aspects of learning are precisely what help young people develop independent reasoning, creativity, and judgment. When AI removes too much of the struggle, it may also remove part of the learning process itself.

My own experience as a student also informs this argument. I have used ChatGPT for studying, organizing lecture notes, summarizing academic articles, improving academic writing, and clarifying difficult concepts. In many cases, this has made learning more efficient and accessible. At the same time, it has also made me more aware of a possible tradeoff. When a summary is available instantly, there is less motivation to struggle through a difficult article. When AI can quickly generate ideas or organize an argument, there is less pressure to construct the answer independently. This experience led me to see AI not only as a useful learning tool, but also as a technology that can influence how students approach thinking, effort, and learning.

If AI has the potential to reshape the way young people think and learn, it is important to examine the specific mechanisms through which this may occur. The following discussion focuses on three hidden costs of AI tool use among youth that appear in the literature.

The first concern is overreliance and cognitive passivity. AI tools are attractive because they provide quick, fluent, and organized answers. However, this convenience can encourage students to depend on AI before attempting to think independently. While ChatGPT can support academic learning by improving access to information and assisting with problem-solving, excessive use may reduce students' mental effort and engagement. Instead of struggling with a difficult text, forming their own interpretation, or developing an argument step by step, young users may become passive receivers of AI-generated knowledge (Nidugonda & Shaik, 2025). This is a hidden cost because the student may appear more productive on the surface, while actually practicing fewer of the cognitive skills that education is meant to develop.

Another issue involves weakened critical thinking and reduced skepticism. Critical thinking requires students to question information, compare sources, evaluate evidence, and recognize weak or misleading arguments. However, AI-generated answers are often written in a confident and authoritative style, which may make them seem more reliable than they actually are. Young users may develop misplaced trust in AI-generated information and become less skeptical of its limitations. This is especially concerning because youth are still developing information literacy and academic judgment. If students accept AI responses without checking them, they may lose opportunities to practice evaluating evidence and building independent conclusions (Nidugonda & Shaik, 2025). Kasneci et al. (2023) similarly argue that the use of large language models in education requires fact-checking strategies, human oversight, and a strong focus on critical thinking. Without these safeguards, AI may support efficiency but weaken deeper learning.

The third concern is reduced originality and independent decision-making. Learning is not only about producing correct answers; it is also about developing one's own voice, creativity, and judgment. When students repeatedly use AI to generate ideas, essays, or explanations, they may become less confident in their own thinking. Overdependence on AI may lead to reduced originality and weaker independent reasoning (Nidugonda & Shaik, 2025). This means that students may begin to rely on AI not only for information, but also for shaping their ideas, judgments, and decisions.

Overall, these concerns show that AI use among youth is not only a question of convenience or academic honesty, but also a question of cognitive development. AI tools can definitely be helpful for students because they make information easier to access, support schoolwork, and give quick explanations when something is confusing, but when they start relying on them too much, they may end up doing less independent thinking, less problem-solving on their own, and less reflection on their own ideas. These downsides shouldn't be ignored.

Ultimately, AI should not replace the effort involved in learning. It can be a really useful tool, but only if students use it to support their thinking instead of doing the thinking for them. Therefore, schools and universities should teach students how to use AI responsibly by checking facts, thinking critically about AI-generated answers, and understanding where these tools can go wrong. In the end, AI should help students think better, not think less.

Every time we ask an AI to help us write, decide, or create — whose voice actually responds? AI systems are trained on the statistical patterns of human behavior at massive scale. The unexpected effect: they don't just assist us; they pull us toward the average. When billions of people interact daily with the same few models, the statistical center of human thought stops being a description and starts becoming a prescription. We think we're guiding AI — increasingly, the mean is guiding us.

The unexpected cost of AI isn't losing control to a superintelligence. It's quietly becoming more average. AI homogenization operates on two levels simultaneously: collectively, erasing the differences between us; individually, narrowing who each of us is over time. Both effects stem from the same root — LLMs are optimized to reproduce what is statistically common, not what is distinctly human.

Drawing on two key studies, this paper examines how AI homogenization manifests across two levels:

1. Collective level — Sourati et al. (2026) show that LLMs flatten linguistic style, perspective, and reasoning across users. Because models are trained to reproduce statistical regularities, they favor dominant — typically Western, educated, male — voices, while marginalizing minority expressions. A recursive feedback loop accelerates this: AI output re-enters human discourse, shapes how people write and think, and eventually becomes training data for the next generation of models.
2. Individual level — the personalization paradox — Matz et al. (2025) reveal that even personalized AI agents, designed to reflect your own preferences, narrow the range of who you are over time. Generic AI pulls you toward the population average; personalized AI locks you into a smaller, more rigid version of yourself. There is no escape from the gravitational pull of the mean.

AI does not need to be malicious to reshape who we are. Trained on the statistical average of humanity and deployed to billions daily, it quietly redefines what counts as normal speech, correct reasoning, and good taste. The take-home message: awareness of this loop is the first step to resisting it. Preserving cognitive diversity — in how we write, think, and choose — must become a design goal, not an afterthought.

Imagine what would happen if Wi-Fi suddenly disappeared while you were writing your essay. This sounds like a simple technical problem. But maybe it reveals something deeper: how much of our thinking have we started to place outside ourselves?

AI tools give us quick answers, organized ideas, and polished sentences, but relying on them may also make us skip the slow and uncomfortable process of forming our own thoughts. This raises an important question: what do we lose when AI begins to think, feel, search for, and find answers instead of us?

Used carefully, AI can support our thinking process. But when we rely on it too early and too often, it may weaken the very abilities it is supposed to help strengthen independent thinking, personal expression, critical judgment, and the ability to sit with discomfort.

One hidden cost of AI is that it can give us the final product without requiring us to go through the process. When we write an essay, solve a problem, or try to understand a difficult idea, the answer itself is not the only important part. The process matters too. Through the process, we struggle, make mistakes, delete sentences, change our minds, and slowly discover what we actually think. A weak first draft is not useless. Confusion is not always a failure. Sometimes, confusion is the beginning of real understanding.

This concern is connected to the concept of cognitive offloading. Gerlich (2025) discusses cognitive offloading as the use of external tools to reduce the mental effort required for tasks such as remembering, analyzing, or making decisions. This is not always negative. Humans have always used tools to support thinking, such as notebooks, calculators, dictionaries, and search engines. The problem begins when the tool stops supporting the thinking process and starts replacing it. If we ask AI to write before we try, summarize before we read, or decide before we think, then we may slowly stop practicing the skills we need.

Gerlich’s study found that frequent AI tool use was associated with lower critical thinking abilities, and that this relationship was connected to cognitive offloading. In other words, the more people relied on AI to reduce their mental effort, the less they seemed to engage in deep, independent thinking. This supports the idea that the danger is not simply using AI but using it in a way that makes us less mentally active.

This concern becomes especially important when we think about students and children. Children are still developing skills such as creativity, patience, critical thinking, problem-solving, and self-expression. If they use AI before these skills are strong, they may learn to skip the hard parts of learning. School assignments are not only meant to measure knowledge; they are also meant to train the mind. When a student struggles to explain something in their own words, that struggle has value. If AI becomes the first solution for every difficult task, we may raise students who can submit good-looking answers but feel helpless without external support.

Another hidden cost of AI is the possible loss of personal voice. Human beings are not unique only because they can give correct answers. We are unique because of our memories, emotions, experiences, and personal way of connecting ideas. Two people can write about the same topic and still sound completely different. Their writing carries something of who they are.

AI, however, often produces answers that are smooth, organized, and acceptable, but also similar. It can help us sound more academic and fluent, especially if we struggle with language or structure. But if we rely on it too much, we may slowly stop sounding like ourselves. Our writing may become cleaner, but less personal and less connected to our real inner world.

This can create a kind of “false self.” A person may present a version of themselves that sounds more intelligent, organized, or emotionally balanced than they actually feel. At first, this may feel helpful, but over time it may create distance between the person and their own voice. The deeper danger is that we ourselves may forget how we sound without help.

AI may also weaken critical judgment. One of the powerful things about AI is that it often sounds confident. It gives clear answers, even when those answers may be incomplete, biased, or wrong. Because of this, people may start treating AI as an authority. Instead of asking, “Where did this information come from?” or “Does this actually make sense?” they may accept the answer because it sounds convincing.

This is dangerous because critical thinking is not only about finding information. It is about questioning it. It means asking: What is the source? Is there evidence? Could there be another explanation? Gerlich’s findings are important here because they show that higher AI use was linked to more cognitive offloading and lower critical thinking. If AI gives us answers too easily, we may practice these questions less. We may begin to confuse fluent language with truth. A sentence can sound intelligent and still be wrong.

There is also an emotional side to this issue. Not knowing is uncomfortable. Starting an essay with a weak sentence is uncomfortable. Feeling confused, lonely, or uncertain is uncomfortable. But these moments are not meaningless. They are part of emotional growth. They teach us patience, flexibility, and the ability to stay with difficulty instead of escaping it immediately.

AI can become a way to escape discomfort. If I do not know what to write, AI writes. If I am confused, AI explains. If I feel alone, AI responds. If I need advice, AI gives it instantly. This can be comforting and useful in some moments, but it can also become a habit. If every uncomfortable feeling is immediately solved by AI, we may risk becoming less able to tolerate silence, uncertainty, and frustration.

At the same time, it would be unfair to say that AI is only harmful. AI can help students who struggle with language, people with learning difficulties, or anyone overwhelmed by a task. It can make information more accessible and help people begin when they feel stuck. The issue, then, is not whether we should use AI or avoid it completely. The issue is how we use it.

A healthier relationship with AI would mean using it as support, not surrender. A student could first write rough ideas, even if they are messy, and only then ask AI to help organize them. A person could ask AI for feedback but still decide what feels true to their voice. Someone could use AI to understand a difficult concept but still check sources and think critically.

In the end, the hidden cost of AI may not be that machines become too intelligent. It may be that humans become too willing to step away from their own inner process. The Wi-Fi going off may seem like a small technical problem. But maybe it asks us a deeper question: who are we without the tool? AI should help us think, not think instead of us. It should support our voice, not erase it. The challenge is to use AI in a way that keeps our minds active, our judgment awake, and our voices still our own.

The human creature is defined by its inherently social nature. Throughout evolutionary history, the ability to form and maintain complex interpersonal bonds has been a primary requirement for survival and psychological development. Recently, the promise of artificial intelligence has led to its integration into multiple fields of human experience, including the sensitive domain of providing mental help and social mediation. Developers often promote the power and usefulness of these technologies as substitutes for human judgment. This article addresses a critical open question: can a non-biological agent truly facilitate the resolution of interpersonal conflict. Despite the increasing prevalence of large language models used as social referees, this discussion posits that AI is fundamentally ill suited for this role. While these tools offer specific cognitive benefits, they ultimately represent a categorical failure to address the core of human connection because humans are social creatures possessing genuine empathy, which no algorithm can replicate.

Proponents of AI integration in social contexts highlight several potential benefits. A central phenomenon is algorithm appreciation, where individuals demonstrate a preference for algorithmic judgment over human judgment in complex scenarios. Research by Logg, Minson, and Moore (2019) suggests that humans perceive algorithms as more objective and less influenced by the noise of human ego. In a relationship crisis, a chatbot can act as a neutral mirror, helping individuals regulate their emotional systems before they cause real-world damage.

Furthermore, AI may serve as a cognitive scaffold for perspective-taking. According to Cross and colleagues (2024), large language models can scaffold a user's Theory of Mind, which is the ability to attribute mental states and intentions to others. This allows for a form of emotional rehearsal in a safe environment where a user can simulate a friend's or parent's viewpoint. Additionally, chatbots can provide social health benefits by acting as non-judgmental social companions, offering support when human networks are strained.

Despite these cognitive supports, the dangers and limitations of using AI for social mediation far outweigh the benefits. The primary challenge is not merely a technical limitation but a fundamental divergence between machine behavior and human sociality. While humans operate through empathy and shared biological experience, AI models operate through probabilistic data processing.

The appearance of support provided by a chatbot is often a form of transmission without truth. Research by Kosoy and Gopnik (2023) identifies a profound distinction between imitation and innovation. Large language models function as stochastic parrots, producing socially plausible text based on vector space semantics without any underlying intent or empathy. Empathy requires a shared affective state and a biological understanding of emotion, which an algorithm cannot possess. Relying on an objective algorithm for a subjective social crisis ignores the requirement for mutual recognition of feelings in human bonds.

Furthermore, the integration of AI into friendships introduces significant ethical risks. Algorithmic bias, as explored by Friedman and Nissenbaum (1996), remains a pervasive issue where systems may replicate or amplify societal prejudices. If an algorithm's social advice is rooted in biased data, it may provide culturally inappropriate guidance that exacerbates a conflict rather than resolving it. Additionally, as these models become larger and more complex, researchers lose the ability to interpret how they make decisions. Entrusting the delicate balance of a human bond to a black box model is a significant risk because the model's decisions may not be aligned with human goals and values.

Finally, the use of AI in relationships risks the erosion of essential social labor. Individuals should not relegate their obligation to think, learn, and synthesize knowledge to a machine. Social interaction is a form of continuous synthesis and emotional work. By delegating the resolution of a conflict to an AI, individuals avoid the necessary emotional labor that leads to genuine growth and stronger social bonds. This relegation prevents the human from performing the very tasks that define their social nature.

Conclusion

The application of artificial intelligence to human relationships is a misguided attempt to apply mathematical optimization to the fluid world of empathy. While AI can imitate the structure of a supportive conversation and provide a neutral space for rehearsal, it lacks the biological and psychological components necessary for authentic connection.

Human sociality is non-algorithmic. We must not relegate the responsibility of maintaining our bonds to machines that prioritize imitation over truth. To preserve our friendships and family connections, we must engage in the difficult work of feeling and understanding one another directly. Empathy is a human faculty that no silicon mediator can replicate or replace.

Artificial intelligence has become a common part of higher education, especially for students involved in independent learning and research. Large language models such as ChatGPT are increasingly used to summarize scientific literature, explain complex concepts, help with programming, and troubleshoot research software. For many university students, AI is no longer just another search engine. Instead, it has become an interactive study companion that provides immediate, personalized support. As a Cognitive Science student involved in neuroscience research, I experienced this while working in my laboratory. During my research, I frequently used ChatGPT while learning to operate software such as SLEAP and B-Soid, which we used for animal pose estimation and behavioral analysis. I often encountered technical errors that were difficult to interpret, and much of the available documentation assumed prior programming experience. Rather than replacing the learning process, ChatGPT helped me understand unfamiliar terminology, identify possible causes of errors, and find relevant resources. Many times, it felt more like a tutor than a chatbot, guiding me through technical problems that would otherwise have taken hours to solve. This experience raised an important question: When does ChatGPT support learning, and when does it begin to replace the cognitive effort that meaningful learning requires? Although AI has become a valuable educational tool, its growing presence in higher education has sparked ongoing debate among educators, researchers, and students. While some view ChatGPT as a resource that makes learning more accessible, others worry that relying on it too heavily may weaken the very skills universities are meant to develop. One of ChatGPT's greatest strengths is its ability to provide immediate, personalized assistance. Unlike traditional search engines such as Google, which often require students to search through multiple websites before finding useful information, ChatGPT explains concepts in a conversational way and adapts its explanations to the user's level of understanding. A student struggling with statistics, programming, or a challenging scientific article can continue asking follow-up questions until the concept becomes clear. This interactive approach allows students to engage with material in a way that textbooks and online documentation often cannot. For independent learners and undergraduate researchers, this type of support can be especially valuable. Research projects often require students to develop skills that go far beyond what they learn in class. It is common to have to learn new software, understand unfamiliar research methods, analyze data, or read scientific papers outside one's area of expertise. Although supervisors and other researchers are usually willing to help, they are often busy with their own experiments, deadlines, and responsibilities. As a result, students frequently have to solve problems independently, making the learning process both time-consuming and frustrating.

This was certainly true in my own experience. When I first started working in the lab, learning to use SLEAP was one of the biggest challenges I faced. The software itself is extremely useful, but many of the error messages were highly technical, and the documentation often assumed users already had programming experience. At one point, I spent almost an entire afternoon trying to understand a single error message because I did not yet know enough programming to recognize what was actually causing it. Looking back, I realized that understanding the problem was often harder than fixing it. Once I understood what the error actually meant, finding a solution became much more manageable. Experiences like this made me appreciate how valuable ChatGPT could be as a learning tool rather than simply a problem-solving tool. During this process, ChatGPT became a valuable learning resource. Instead of simply giving me an answer, it helped me break technical concepts down into language I could understand. It explained unfamiliar programming terms, helped me interpret error messages, suggested different troubleshooting approaches, and pointed me toward relevant documentation or online resources. Being able to ask follow-up questions and receive explanations tailored to my level of understanding made the learning process much more manageable than relying on

Fraud predates artificial intelligence, and it was never easy to detect. Some time ago a supposed airline representative called my friend about a refund for a cancelled flight; everything he said sounded wrong, yet he was so confident that we briefly doubted ourselves rather than him. If a stranger with a weak story could almost succeed, the relevant question is what happens when the caller is someone we know and trust. Artificial intelligence makes that scenario realistic, because it can now reproduce a specific person's voice convincingly. This paper argues that the central threat of AI-generated media is not technological but evidentiary: it dismantles our ability to treat a digital evidence as proof, and because detection consistently fails, the only proportionate response is to mark synthetic content at the moment of creation.

A case study: The Kim Soo-hyun affair

The South Korean actor Kim Soo-hyun offers a documented example. A YouTube channel, HoverLab, published an audio recording it presented as evidence that he had dated the late actress Kim Sae-ron while she was a minor - a criminal offence in Korea. Sponsors withdrew, a streaming series was shelved, and his career collapsed within weeks. More than a year later, a police forensic investigation concluded that the key audio had been generated or altered using AI and that accompanying chat messages could not be reliably attributed to him. The channel's operator was arrested on defamation and related charges; he denies wrongdoing, and proceedings remain ongoing, so the matter is neither closed nor a formal acquittal. The case is instructive precisely because the fabricated evidence functioned as real evidence for over a year, and the institutions tasked with verifying it could not do so quickly enough to prevent the harm.

How voice cloning works

Understanding the threat requires understanding the tool. Modern voice cloning is a form of neural text-to-speech synthesis, and it typically proceeds in three stages (Jemine, 2019). First, the system analyses a short audio sample of the target and extracts a speaker embedding: a compact mathematical representation of the acoustic features that make a voice distinctive, including pitch distribution, formant patterns, speaking rhythm, and tonal range. Second, this embedding is used to condition a neural model already trained on many speakers. The model does not learn speech from scratch but adapts a pre-existing, high-quality speech model to the target's specific signature. Third, a vocoder converts the model's output into an audible waveform. Because the system only adapts an existing model rather than building one, cloning is fast and data-efficient: usable results can be produced from roughly sixty seconds of audio, and some implementations report average cloning times of only a few seconds.

Existing harms

These capabilities are already being exploited. In the United States, the FBI has warned of "family in distress" scams in which a victim receives a call in the cloned voice of a relative requesting urgent money; in one reported case a man in Houston lost 15,000 dollars to a voice he believed was his son's. The same technology scales beyond individuals into the economic and corporate sphere, where synthetic voices and faces are used for industrial espionage and to defeat remote identity checks. A further, low-cost variant is emotional manipulation at scale: fabricated images of staged suffering - a person claiming to be alone and uncelebrated on their birthday, for instance - are posted to social platforms to harvest engagement. Such accounts now mostly collect likes, but the open question is when audiences who cannot tell real from fabricated suffering will begin donating in response.

The deeper harm is evidentiary. A recording has traditionally functioned as something close to proof, capable of settling disputes that would otherwise reduce one person's word against another's. Voice cloning removes that function. A friend of mine was once the subject of a false accusation before a military commander; it failed only because the commander judged the people involved accurately and another officer already knew the true account. Had the accuser possessed a fabricated recording in her voice, denial would have been far harder, because forged audio carries an evidentiary weight that testimony alone lacks. When any voice can be synthesized, a recording no longer proves anything on its own.

Why detection is insufficient

A natural response is to rely on detection, but the evidence indicates that detection fails at both the human and the technical level. Mai et al. (2023) found that people cannot reliably identify speech deepfakes even after being trained to do so, and that listeners remain confident in judgements that are frequently wrong. One might expect forensic systems to compensate, yet Univaso and San Segundo (2025) demonstrate that automated forensic voice-comparison tools also degrade under realistic conditions such as background noise, compression, and unfamiliar cloning algorithms. Both lines of defense - the human ear and the forensic laboratory - therefore give way. This produces a structural asymmetry: fabrication is fast while verification is slow, so confirmation tends to arrive only after the damage is complete. If neither listeners nor laboratories can reliably identify fakes after the fact, post-hoc detection cannot serve as the primary safeguard.

A structural solution: marking at the source

If detection arrives too late, the intervention must move to the point of creation. The proposal is that AI-generated output should carry a mark from the moment it is produced. Two complementary mechanisms exist. A watermark embeds a hidden, machine-detectable pattern within the content itself; it survives downloading but can be weakened by editing. A provenance signature, by contrast, is metadata that travels with the file and records how it was made; it is transparent and informative but can be removed, for instance by taking a screenshot. Used together, they allow platforms to verify automatically whether content is synthetic and to present the public with a clear, visible indicator - comparable to a nutrition label on food, which informs rather than censors.

This approach is already entering law. The European Union's AI Act establishes transparency obligations under Article 50, enforceable from August 2026, requiring that AI-generated or manipulated content be marked in a machine-readable format and that deepfakes be disclosed even where no deception is intended. The obligations bind both the providers who build the systems and the deployers who publish content, so neither can rely on the other to disclose. The Act also distinguishes fully AI-generated content from merely AI-assisted work, applying lighter requirements to the latter and exempting clearly artistic or fictional creations - a distinction that matters for proportionality and for protecting legitimate creative use.

Discussion: transparency and its limits

Marking at the source is not without difficulty. Enforcement is hard, particularly across borders: a cloned-voice call may be placed from one jurisdiction and received in another, and meaningful cross-border enforcement remains limited without international cooperation (Pacheco, 2025). The principal advantage of provenance, however, is that a signature attached to a file does not depend on locating a perpetrator after the fact.

A more subtle problem concerns trust itself. If all content is labelled, audiences may come to doubt everything, including genuine material. As the public grows more aware that media can be fabricated, those who have committed an offence may dismiss genuine evidence against them as a deepfake, while in other cases the reverse occurs, as in the Kim Soo-hyun affair. Greater public awareness can therefore deepen the problem rather than resolve it, which is one reason media literacy alone is an insufficient response.

Conclusion

The central argument holds: human judgement has proven unable to separate authentic from fabricated media, and so has forensic analysis; where judgement fails, structure must take its place. Marking synthetic content at creation will not eliminate every hard case or misuse, but it offers a scalable means of preserving a shared evidentiary basis for public life. People have already been harmed by these tools, whether through voice-cloning scams or the reputational damage in the Kim Soo-hyun case, yet the harm so far remains largely reparable, and the actor himself is reportedly recovering his career.

This invites a broader question for further inquiry: how significant are the dangers of these tools relative to the advantages they offer, and how far will they carry us along both scales? As forgery becomes effortless, authenticity will depend less on our ability to detect fakes and more on whether content carries a verifiable mark of origin.

A few weeks ago, I found myself experiencing an instance of cognitive and emotional block while attempting to write an entry in my journal. I knew I needed to externalize my feelings to process them. Struggling to find the precise vocabulary to describe my internal state, frustrated, I turned to the default solution: ChatGPT. I typed my chaotic, unedited, and raw thoughts into a prompt box, and added a specific instruction: “I'm finding it hard to explain what this is, can you just format this and make it make sense?” Within three seconds, the model delivered a structured, coherent paragraph that neatly articulated my emotions. However, when I read the paragraph it didn't quite contain the element of “Me” and unfortunately it did not help with truly processing my underlying emotions which made me wonder: When we let AI do the primary cognitive labor of translating and structuring our subjective emotional chaos, do we systematically undermine our own capacity to understand, integrate, and author our own life story? There is a subtle yet dangerous hidden cost of generative AI: the systemic erosion of narrative identity. By replacing the vulnerable, and often painful process of human translation with algorithmic outputs, we are outsourcing the very mechanics by which the human mind constructs meaning, leading to psychological and cultural flattening. In my opinion, there is an inherent beauty in human psychological “messiness”, and by offloading creativity to predictive models, we subconsciously strip away the core elements of the self. Psychologist Dan McAdams established that human identity is not a static set of traits, but an internalized, evolving story and a narrative identity (McAdams, 2001, McAdams & McLean, 2013). We weave our reconstructed past and imagined future into a coherent biographical narrative that provides our lives with unity and purpose. The psychological value of a life story does not come from its structural perfection, but from the sometimes irrational ways we negotiate its contradictions and traumas. When individuals use LLMs to draft personal essays, write journal entries, generate dating profiles, or script films, they insert a predictive mathematical buffer into this internal loop. Because LLMs are trained to predict the most statistically probable next word, they optimize for an idealized, collective average and in doing so we are undergoing a process of narrative erasure. The human brain utilizes narrative writing as a core mechanism for emotional regulation and cognitive integration. Work in expressive writing by James Pennebaker (1997) demonstrates that transforming difficult, chaotic, or traumatic experiences into a structured linguistic story actively alters immune function, neural connectivity, and psychological wellbeing.

This therapeutic effect requires intense cognitive “heavy lifting”. The individual must personally struggle with language to reconcile their internal chaos with external reality. For example, when an individual uses an AI to “polish” their raw thoughts, that vital cognitive labor is bypassed. While Pennebaker’s paradigm focused on the structured processing of trauma, the same psychological mechanisms apply to everyday emotional navigation. The AI provides an articulate response instantly, eliminating the uncomfortable, necessary silence where psychological insight and emotional resolution are actually born. Furthermore, McAdams and McLean (2013) emphasize that narrative identity has a critical developmental trajectory. During adolescence and emerging adulthood, individuals must acquire the capacity for autobiographical reasoning which is the cognitive skill required to derive self reflective lessons from hardships. By embedding algorithmic intermediaries into early writing processes, we risk raising a generation of communicators who possess high digital productivity but suffer from severe emotional atrophy. However, this psychological flattening does not stop at individual reflection, it projects outward onto our cultural narrative. For example, cinema and literature serve as external outlets for internal struggles, feelings, and opinions. Historically, the boundary pushing shifts in film history, such as the French New Wave or American New Hollywood, emerged precisely because directors rejected traditional, polished structures in favor of radical, fragmented forms. When filmmakers rely on these systems, the resulting cinema may look flawless and perform well on streaming algorithms, but it loses its capacity to challenge or expand human empathy. It creates a closed loop where AI trains on human data, and humans consume that narrative, unconsciously recalibrating their own emotional expectations to match the algorithmic output. This directly mirrors the emotional detachment one can feel when consuming highly polished, contemporary media, where the structural perfection of the narrative feels clinically corporate, and leaves the observer entirely unimpacted. In conclusion, generative AI tools are incredible mirrors, reflecting back to us the sum of what we have already written. But a mirror cannot walk forward into unknown territory. The ultimate message for both psychologists and artists is that “meaning making” cannot be automated without being destroyed. To preserve our individual and collective identity, we must deliberately guard the spaces where our narratives are unpolished, messy, and unresolved. We must protect the silence of the blank page, the awkward journal entries, and the jagged

cinematic scripts that make no sense to an algorithm. It is within those exact zones of human creativity that our true identity resides.

I still remember the day I received my first coding assignment, back to 2021, in my first academic journey at Ben Gurion University of The Negev. I remember sitting with my friend in the library of the university's campus at Beer Sheva', each with his own computer, a pen and paper, focusing on solving what needed to be solved, step by step, in order to write the code properly. I also remember it took us hours to complete each section. We had to think carefully, make mistakes, erase, try again and discuss possible solutions. After we solved each section, we compared our codes and talked about the differences. Looking back, I realize that the learning did not come only from reaching the correct answer, but from the long process of struggling with the problem. Today, only a few years later, solving the same assignment might look very different. Instead of spending hours trying to understand the logic of the assignment, I could paste the instructions into any AI tool and receive a possible working solution and a runnable code within seconds. At first, this seems like a clear advantage, AI can save time, reduce frustration and help students move forward when they feel stuck. Studies on AI chatbots in education suggest that these tools can improve students' learning outcomes, especially when they provide immediate feedback and personalized support (Wu & Yu, 2024). However, this change also raises an important deep question about what learning programming really means: to what extent does using AI tools for code writing assignments support students' learning and becoming better programmers, and when does it begin to replace their independent problemsolving process? Programming is not only about producing code that runs, but also about learning how to think in a structured and logical way. When students work on coding assignments, they practice breaking a large problem into smaller parts, identifying patterns, searching for ideas, choosing the right functions, testing possible solutions and correcting mistakes. Debugging, although often so frustrating, is one of the most important parts of learning programming. An error output forces the student to slow down and ask deeper questions about the

code he's writing. Such questions develop the student's computational thinking skills. I remember we used to create trace tables1 that track the flow of the code, each row represented a function that showed what the output would be in a given input. We used to mark down the outputs in the table, which made us sometimes realize a mistake we made through the code.

• a trace table:

In this sense, the long hours spent solving a programming assignment are not simply a wasted time. They are a part of the learning process itself. When I sat with my friend in the library, we were not only trying to finish the task, but we were also learning how to explain our reasoning to each other, how to compare different approaches and how to recognize that there can be more than one correct solution. This social and cognitive process that we were involved in helped us understand the material more deeply than simply receiving a finished answer. Hassen (2025) emphasizes that AI may affect students' reading, critical thinking, and problem-solving skills, which are exactly the abilities involved in learning to code. From this perspective, the concern is not only whether students submit correct code, but whether they develop the thinking skills behind it. At the same time, AI tools can support students in many meaningful ways. Many beginner students experience programming as stressful because even a small mistake can cause the entire code to fail. In these situations, AI can act almost like a private tutor. It can explain unfamiliar concepts, translate technical language into simpler words, suggest why an error might occur or provide an example that helps the student understand the logic behind a solution. Klimova and Pikhart (2025) note that AI in higher education can support academic experiences by making learning more personalized and accessible.

For students who are afraid to ask basic questions in class, AI may offer a more comfortable space for learning. The educational value of AI, especially in programming, depends on the student's role in the interaction. If the student remains active, AI can strengthen learning. For example, a student can first try to solve the assignment independently and then ask AI to explain a specific error. They can ask for a hint instead of a full solution or ask AI to compare different possible approaches. They can also ask AI to explain each line of the code, generate practice questions to check whether their understanding is correct. In these cases, AI doesn’t replace thinking, it supports the student while they are still doing the intellectual work and going through the whole thinking process that needs tracking inputs and outputs and understanding extreme cases. The risk begins when AI is used too early or too passively. If a student gives the assignment to AI before attempting to solve it by himself, he may receive a correct solution without understanding. Because the AI explanation is fluent and confident, the student may feel that they understand the code. However, when asked to solve a similar problem without help, they may not know where and how to begin. In this case, AI doesn’t function as a learning booster, it becomes a thinking replacement. Prather et al. (2024) describe a similar concern; generative AI may widen the gap between stronger and weaker students by giving struggling students a false sense of progress or competence. This risk is especially important in code writing because programming skills are built through practice. A student can't become confident in coding only by reading finished solutions. They need to experience confusion, test possible ideas, make mistakes and repair them. If AI removes this process completely, students may become dependent on automated answers. They may become better at prompting AI, but not necessarily better at programming. The assignment may be completed but the skill may not be developed. Li, Cui, and Hagedorn (2026) discuss the cognitive impact of ChatGPT in higher education and show that AI-assisted learning can have both supportive and harmful effects on critical and creative thinking. For this reason, the question shouldn't be whether students should use AI writing code or avoid it completely. A more useful question would be how students can use AI responsibly. Students should be encouraged to use AI after they have made an initial attempt. They should ask AI to guide and expand their thinking rather than replace it. For example, instead of asking "write the full code for me", a student should ask "what is wrong with my logic?" or "can you give me a hint for the next step?". Another helpful practice is to explain the AI-generated

code in one's own words. If a student is not able to explain what the code does, then they have not truly learned from it. responsible AI use, which is already happening in my university. Instead of only banning AI, courses can include clear guidelines about when and how AI may be used. Assignments can require students to submit short reflections explaining their process, the prompts they used, and what they learned from the compare their own solution with an AI-generated one. These methods may shift the focus from the final product to the learning process and support responsible use of AI (Li et al., 2026; Klimova & Pikhart, 2025). In conclusion, AI-assisted code writing among programming students may change the meaning of programming assignments. In one hand, it can make learning faster, more accessible and less frustrating. In the other hand, it may also weaken learning by allowing students to skip the struggle that helps them develop real understanding. Looking back at my first coding assignment, I now realize that the difficulty was not separate from the learning experience. The mistakes, the discussions and the slow process of solving the problem were all parts of learning how to think like a programmer. AI can be a powerful learning booster, but only when it helps students think better, not when it thinks instead of them.