In the Butlerian Jihad, humanity defeats the thinking machines, and the Orange Catholic Bible lays down its second commandment: “Thou shalt not make a machine in the likeness of a human mind.” Frank Herbert wrote Dune in 1965, when this was fiction.

Today, in early 2026, artificial intelligence (AI) is no longer science-fiction; it is embedded in our daily life, and the idea that such systems might one day, not in the so distant future, be able to act against us feels far less abstract.

The pace of AI is unmistakably accelerating. While public narratives still promise cures for disease and breakthroughs across science, the largest investments continue to flow into military applications—autonomous surveillance systems, tactical decision-making engines, and, increasingly, weaponized autonomous platforms. It is worth noting that the EU AI ACT regulates all aspects of AI expect for military use.

AI, however, is more dangerous than nuclear weapons. A nuclear device cannot design a more advanced nuclear device; an AI system can, and already does, contribute to the improvement of successor models. A nuclear device cannot copy itself; advanced AI, if given the wrong affordances, can replicate, migrate, and persist.

Despite all this, I remain curious about AI. These systems offer unprecedented tools for understanding computation, cognition, and the brain. I am optimistic that AI could fix the global healthcare system. But the central question persists: do we truly need to build something with agency that mirrors our own minds? Or would humanity be better off by focusing on narrow, domain-specific AI?

Narrow AI already delivers. These systems do not need human-like agency to be transformative. The insistence on human-level intelligence or even beyond human-level intelligence seems motivated less by scientific necessity than by a mixture of ambition, competition, and myth. And myths, historically, have been poor guides for technological restraint.

The challenge ahead is philosophical and ethical. We must decide what kinds of minds we are willing to construct, and for what purposes. Narrow systems can amplify human capability. General systems, if pursued recklessly, may compete with it. The choice is still ours but for how long?

We know by now that social media can be addictive, not in the DSM-5 yet though. As we march into the world of artificial intelligence (AI) being used more in our daily lives, could it become our next digital dependency?

Generative AI tools such as ChatGPT and Gemini are powerful. Need to write an email, solve a coding problem, compose music, or generate artwork? AI delivers on demand. This ease feels almost magical. You cast a spell, and there it is.

Unlike social media dependency, AI dependence is sneaky. AI produces text, code, presentations, and analyses. It saves time. It boosts confidence. It makes us feel more innovative.

But AI is not a calculator, a spell-checker, a GPS, or a search engine; all of these offload specific cognitive functions. Generative AI is unprecedented in its ability to offload cognitive tasks.

I suspect it could lead to dependence because generative AI outputs are probabilistic and occasionally surprising. This unpredictability, the gap between what we expect and what we receive, creates variable rewards similar to those in gambling, a mechanism known to maximize engagement. Second, conversational interfaces used in Chatbots encourage anthropomorphism. We tend to attribute human qualities to such systems and even trust systems that use social cues.

The line between human and machine was never thinner.

We are at a crossroads. Do we want to outsource our thinking entirely, or amplify our own cognitive work? The stakes are high. It is about preserving expertise, critical thinking, and human agency in an increasingly automated world.

As artificial intelligence (AI) advances, comparisons between human learning and machine learning become increasingly relevant. While it might be tempting to think that machines learn similarly to humans, the reality is far more complex.

In the 1940s, Donald Hebb proposed that, “neurons that fire together, wire together,” illustrating how experiences strengthen connections within the brain, enabling learning. This foundational understanding of human learning contrasts with the development of AI.

The perceptron, introduced by Frank Rosenblatt, a psychologist, in 1958, marked a seminal advancement, simulating a basic form of learning in machines by adjusting responses based on outcomes.

In around that time, the term “machine learning” became popular. Arthur Samuel, worked in 1959 on programming a computer to improve at the game checkers. His work coined the term “machine learning” and he wrote: “A computer can be programmed so that it will learn to play a better game of checkers than can be played by the person who wrote the program.”

Later, and in 1960s and 1970s, neural networks and deep learning started to emerge as subsets of the field AI. These are described as attempts to mimic how the brain works by their developers.

However, the debate between Karl Friston, a neuroscientist, and Yann LeCun, a computer scientist, at Davos 2024 underscores a critical divide in understanding AI’s capabilities.

Friston criticizes deep learning for lacking a framework to account for the uncertainties inherent in human cognition, a fundamental aspect of our learning process.

“So, I think Deep Learning is rubbish, largely because it doesn’t have the calculus of the machinery or the physics to have a calculus of beliefs, a calculus of inference, of planning, of situational awareness. And crucially, if you don’t have math or an engineering toolkit that can quantify what you know and what you don’t know, namely, encode the uncertainty that underwrites your sense making, you’ll never know how to act in terms of getting the right kind of information”.

Friston proposes, in his free-energy principle theory, that the human brain operates on predictive coding, constantly making hypotheses about future events to minimize surprises. This is akin to learning through trial and error, where success and mistakes refine our predictions and responses.

On the other hand, LeCun highlights the immense data processing and learning capabilities of AI, yet acknowledges its shortfall in emulating the depth of human learning, which integrates experience, observation, and interaction.

This divergence points to a core issue: the term “machine learning”, while practical, might suggest a cognitive engagement and understanding similar to human learning, is misleading.

Machines excel at identifying patterns and making predictions from data, but they lack the ability to derive meaning, purpose, or undergo experiential growth, all are essential components of human cognition.

Acknowledging these differences is vital for setting realistic expectations for AI’s capabilities while avoiding delusional beliefs. It also guides the ethical development and deployment of AI technologies. The quest to create machines that emulate human learning continues, but recognizing the distinct nature of human cognition is crucial for the progress of AI.

This article was published originally online in 2024 at theyuan.com.

Since the introduction of ChatGPT and large language models, I will refer to here as AI, they have made a grand entry into nearly everyone's life. You have probably noticed they come with a disclaimer: they make mistakes. For some reason, these mistakes have been labeled "hallucinations." As usual, the field of AI naively borrows from neuroscience and psychology. In fact, calling these mistakes "hallucinations" is more poetic than precise; the proper term is confabulations.

According to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, Text Revision (DSM-5-TR, 2022), hallucinations are defined as perceptual experiences that occur in the absence of an external stimulus and are experienced as real. People with conditions like schizophrenia may hear voices that no one else hears. Others with Parkinson's disease or certain types of dementia might see things that aren't really present. Some people going through drug withdrawal might feel bugs crawling on their skin, and a strange smell with no source can sometimes be an early sign of a seizure disorder.

The DSM-5-TR does not provide a formal definition of confabulation. It is not listed as a standalone diagnostic criterion or disorder. Confabulation is primarily a neuropsychological concept rather than a psychiatric diagnostic term, and it is more extensively discussed in clinical neuropsychology literature than in DSM manuals. In neuropsychology, Schnider in 2005 defined confabulations as the unintentional fabrication of memories, typically to fill in gaps caused by amnesia. Korsakoff's syndrome is a brain disorder often caused by long-term alcohol misuse, leading to severe memory problems. A person with this condition might firmly believe they had breakfast that morning at a favorite café, which actually closed years ago. They are not trying to lie. Their brain is unknowingly filling in the blanks. This is known as confabulation. It is when someone creates false memories without realizing it, to make sense of gaps in their memory.

AI does not know anything in the human sense. What it does is predict. When a chatbot generates a false but plausible answer, it is not sensing a ghost or hearing voices that are not there. It is predicting an output from an input based on patterns in its training data. It is trying to fill in what should come next, even if it makes it up, just as a confabulating brain fills in what must have happened.

Using "hallucination" to describe AI errors encourages an idea of machines, as if they had inner worlds and private perceptions. "Confabulation" is more accurate and more useful. It reminds us these systems generate text, not truth as they do not know what is real in a human sense. They stitch together words based on statistical echoes, not reality checks. Understanding this helps developers build better tools. Ones that flag uncertainty or backtrack when unsure. And it helps users stay clear-eyed: the machine is not dreaming or lying. It is guessing, with confidence but no conscience. So let us drop the romance of "hallucination." AI does not perceive. It predicts. When it gets things wrong, it is not deluded. It is making things up. It is confabulating. And just as in medicine, proper treatment begins with an accurate diagnosis. Therefore, naming the problem correctly is the first step to addressing it.

Is obesity a new phenomenon? How was it historically managed? And what’s behind the new medications for weight loss?

In ancient times, obesity was rarely an issue. In fact, throughout much of history and in many societies, being overweight was a sign of prosperity, as it signaled access to food that was primarily restricted to the elites. This view persists, up until today, in some communities in Mauritania, where a female beauty is determined by higher weight. Throughout history, having low weight was often viewed negatively, as an indication of poverty, illness, or both. But what about the new world?

In the new world, obesity remained a rarity even in the late nineteenth century. Chauncey Morlan, reported to be the "world's fattest man" in 1890, weighed 875 pounds (397 kg) at his heaviest, and traveled with a circus as a human attraction. He was introduced as the "Human Freight Car," drawing audiences from across the country who came specifically to see him.

We knew the health risks associated with overeating for thousands of years. Prophets spoke of the virtue of eating in moderation. Hippocrates, the father of medicine, acknowledged the health risks of being obese and recommended moderation in food intake and regular exercise.

However, when moderation and exercise failed, we resorted to invasive methods. Historical accounts suggest that the earliest form of invasive intervention to induce weight loss occurred in 10th-century Spain. King Sancho of Leon, known as Sancho the Fat, a reference to his extreme obesity, which hindered his mobility and ability to rule, lost his throne as a result.

In pursuit of treatment, he was taken by his grandmother to Cordoba to visit the physician Hasdai Ibn Shaprut. Ibn Shaprut stitched the king’s lips, restricting him to a liquid diet delivered through a straw. Remarkably, King Sancho shed half his body weight, enabling him to return to Leon, ride a horse again, and reclaim his throne.

In the early 20th century, as medical science advanced, a more scientific approach to understanding and managing obesity started to emerge. We began to understand concepts of calorie counting and energy balance. In parallel, the mid to late 20th century saw a significant increase in obesity rates, propelled by increased consumption of high-fat, sugar-rich, processed foods, leading to more focused research on obesity management.

This period witnessed the advent of modern surgical approaches to induce weight loss. The first modern metabolic surgery, known as the jejunoileal bypass, was introduced by Kremen in 1954. This surgery involves connecting the upper part of the small intestine (jejunum) directly to its lower part (ileum), skipping a large section of the intestine that absorbs nutrients. Instead of stitching the mouth shut—as practiced a millennium ago by Ibn Sharput to reduce food intake—modern surgeons bypass nutrient-absorbing segments of the gastrointestinal tract, a strategy that can lead to malnutrition and, in some cases, surgery-related complications and death due to the invasive nature of the procedure.

Traditionally, Body Mass Index (BMI) is the metric used to determine if someone is obese. BMI is calculated by dividing a person's weight in kilograms by their height in meters squared.

However, BMI is limited for personalized intervention, despite its importance as a statistical tool. For example, consider a bodybuilder who weighs 120 kg, has a body fat percentage of 5%, and is 170 cm tall; his BMI is 41.5, and under BMI criteria, this individual would be classified as morbidly obese. This example highlights why BMI alone may not always be the best measure of obesity, given the 'fat but fit' phenomenon.

Nowadays, obesity is increasingly recognized as a disease. According to the World Obesity Federation, obesity is recognized as a chronic, relapsing, progressive disease. However, this definition is not universally accepted, indicating ongoing debates in the medical community about how best to define and understand obesity.

In an effort to define obesity, the 2020 Canadian Clinical Practice Guidelines for adults state: "Obesity is a prevalent, complex, progressive, and relapsing chronic disease characterized by the presence of excess or abnormal body fat (adiposity) that impairs health". The key point is the increase in abnormal fat tissue that impairs health.

Regardless of the ongoing debate around its definition, obesity is primarily regarded as a preventable global pandemic at the population level with complex etiologies, including genetic, environmental, psychological, and social factors.

As such, to tackle the global pandemic of obesity, several medications were developed to induce weight loss, reflecting the challenge between ensuring safety and efficacy. For example, in the 1990s, the fenfluramine and phentermine (fen-phen) combination was popular but withdrawn due to heart risks. Sibutramine was U.S. Food and Drug Administration (FDA)-approved but withdrawn in 2010 for heart risks. Ephedra, causing heart problems, was banned. Rimonabant was initially approved in Europe but withdrawn due to psychiatric risks and was never FDA-approved. Consequently, orlistat, an inhibitor of intestinal fat absorption, remained for many years the only long-term pharmacological option for obesity.

Since 2005, however, we have new players in the game. The first Glucagon-like Peptide-1 (GLP-1 agonist) used for diabetes management Exenatide was approved by the FDA in 2005. While initially developed for type 2 diabetes, GLP-1 agonists' role in weight loss was observed, leading to further investigation of their obesity management and subsequent FDA approvals of additional GLP-1 agonists indicated explicitly for chronic weight management, such as liraglutide (approved 2014) and semaglutide (approved 2021), as well as dual GLP-1/GIP agonist tirzepatide for obesity (approved in 2023), and most recently an oral semaglutide (approved in 2025) for weight loss.

GLP-1 agonists work by mimicking the natural glucagon-like peptide-1 hormone in the body. They promote insulin secretion in a glucose-dependent manner, enhancing blood sugar control. Additionally, they slow gastric emptying, leading to prolonged feelings of fullness, and directly act on the brain to reduce appetite. This dual effect on both glucose metabolism and appetite regulation makes them effective for both diabetes and obesity management. Emerging evidence suggests that their positive effects extend to reducing inflammation and protecting the brain.

These medications are changing the way we are approaching metabolic disorders, however, they have some serious side effects, most commonly gastrointestinal adverse effects. For example, a retrospective observational study found increased risks of pancreatitis (pancreas inflammation), gastroparesis (paralysis of the gut), and bowel obstruction with GLP-1 agonists.

Our understanding of metabolic disorders including obesity is evolving and as we are entering the era of artificial intelligence we might finally not only manage chronic medical disorders but cure them. Until then at the individual level, obesity often behaves as a chronic, relapsing condition for which lifestyle interventions alone may be insufficient. In such cases, pharmacotherapy can serve as an evidence-based adjunct to lifestyle modification, initiated under medical supervision in appropriately selected patients.

---

Disclaimer: This is for educational purposes only and does not constitute medical advice. Always consult your treating physician.