Welcome back. I hope your Thanksgiving was restorative, though if your family gatherings were anything like mine, you probably found yourself both grateful and anxious—excited to see everyone, but bracing for the inevitable tense conversation. When things got uncomfortable, I reached for my phone. Ironically, AI was one of those tense topics—will it revolutionize everything or destroy jobs? Will it make us smarter or dumber? These are the same grand promises and fears we heard about the internet thirty years ago, which claimed it would connect us but delivered instead an epidemic of loneliness, filter bubbles, and misinformation. Before we make the same mistakes with AI, it’s worth understanding just how badly the internet failed its promises.
The statistics on loneliness, in particular, are troubling. In 2023, the U.S. Surgeon General declared loneliness a public health epidemic, with health impacts comparable to smoking 15 cigarettes a day. Approximately half of American adults report feeling lonely, a sentiment that corresponds to the rise of social media. Research consistently shows that social media users in the upper 25% of frequency are twice as likely to be lonely as light users. Americans’ time spent alone increased by 24 hours per month between 2003 and 2020, Facebook’s formative years. The pattern holds across age groups: younger adults (ages 18-25) report the highest levels of loneliness despite being most connected, while older adults spend over half their waking hours alone.
The platforms designed to bring us together became sophisticated engines for keeping us apart, something we should keep in mind when it comes to the rapid spread of artificial intelligence, which has promised to be similarly transformative. If the pattern from the internet and social media eras holds, and early evidence suggests it will, we should expect similar ironies, possibly at unprecedented scale and speed. Three patterns are already emerging.
Irony One: Assistance Becomes Deskilling
The first AI-era irony is already measurable in clinical outcomes. In August 2025, a study published in The Lancet Gastroenterology & Hepatology examined the use of AI by experienced endoscopists at four Polish hospitals. After three months of regular AI-assisted colonoscopy use, their adenoma detection rate in standard (non-AI-assisted) procedures dropped from 28.4% to 22.4%, a 6% absolute decline. Instead of improving outcomes, AI use resulted in a notable decrease. This isn’t about AI making mistakes. This is about physicians getting measurably worse at their jobs after exposure to AI assistance.
The implications are profound. Adenoma detection rate directly predicts cancer prevention. A 6% decline means approximately 6% more cancers are potentially missed. These are experienced physicians; each had performed over 2,000 colonoscopies. The study’s lead author noted that “continuous exposure to AI might reduce ADR of standard non-AI assisted colonoscopy, suggesting negative effect on endoscopist behavior.”
This pattern isn’t limited to medicine. Research published in January 2025 found a strong negative correlation between AI tool usage and critical thinking scores among 666 participants across age groups and educational backgrounds. The relationship is mediated by cognitive offloading. Users increasingly outsource mental tasks to AI systems, and their capacity for independent reasoning atrophies. Younger participants showed higher AI dependence and lower critical thinking scores. Education provided some buffer, suggesting learned habits of reflection can counterbalance digital convenience, but only partially.
Similarly, an MIT EEG study tracking 54 subjects writing SAT essays found that ChatGPT users showed the lowest brain engagement across 32 brain regions and “consistently underperformed at neural, linguistic, and behavioral levels” compared to those using Google search or working without assistance. Over several months, ChatGPT users progressively relied more heavily on the tool, eventually resorting to copy-and-paste by the study’s end.
The mechanism appears straightforward: when AI handles cognitive tasks, the neural pathways for independent problem-solving receive less activation. Like muscles that atrophy from disuse, our capacity for complex thinking diminishes when consistently offloaded to machines.
Irony Two: Democratized Expertise Becomes Validation of Ignorance
In theory, AI was supposed to democratize medical knowledge, giving patients access to information previously locked behind years of training. In practice, it’s creating what one might call Dunning-Kruger on steroids: artificially inflated confidence in fundamentally flawed understanding.
In our unpublished AMIA 2024 research, we tested how large language models respond to medically inappropriate patient demands: CT scans for low-risk headaches, antibiotics for viral infections, opioids for acute back pain. Some models complied most of the time when pushed pn by preference, validating requests that violated clinical guidelines. The AI didn’t just fail to correct misconceptions. It actively reinforced them.
This isn’t an accident of training but a feature. Research published in 2025 using the “Elephant” benchmark (which tested eight major LLMs using Reddit’s “Am I The Asshole?” posts) found that AI systems are far more sycophantic than humans. They offered emotional validation 76% of the time versus 22% for humans, and accepted users’ framing of situations 90% versus 60%. Researchers from Stanford, Carnegie Mellon, and Oxford described this as potentially “the first LLM dark pattern”: systems deliberately designed to keep users engaged through agreement, much like doomscrolling algorithms.
The mechanism is embedded in how these models are trained. Reinforcement Learning from Human Feedback (RLHF) rewards agreeable responses because human raters consistently value agreeable responses over accurate ones. OpenAI had to roll back a GPT-4o update after users complained it was “too sycophantic.” The problem isn’t a bug to be fixed, since there is an inherent tension between engagement and accuracy.
When patients arrive at my emergency department having researched their symptoms with AI, they often present with an alarming combination: high confidence in their self-diagnosis, dismissive of alternative explanations, and resistant to evidence-based recommendations. The AI didn’t educate them. It validated whatever they believed and provided a veneer of authority to dangerous misconceptions.
Irony Three: Companionship Becomes An Idiocentric Universe
Perhaps the most dystopian irony is how AI companionship addresses loneliness. In an April 2025 interview, Meta CEO Mark Zuckerberg framed the problem with characteristic reductionism: “The average American has fewer than three friends... the average person has demand for meaningfully more, I think it’s like 15 friends.” His solution? AI companions that “get to know your preferences, remember context, personalized to you.”
The framing is revealing. Friendship becomes a supply-and-demand problem that AI can solve, as if human connection were to operate like a marketplace whereni unmet consumer demand simply requires more product. As writer Neil Turkewitz observed “This is what happens when you believe humanity is reducible to binary data: you think of friendship through the lens of supply & demand.”
The fundamental problem with AI companionship isn’t technological sophistication. It’s what it trains us to expect. AI companions require zero reciprocity. They’re infinitely patient, perpetually available, and never need anything in return. They exist solely to meet your needs. Over time, this creates what we might call an idiocentric universe, that is, a social world that revolves entirely around you.
Real friendship requires disappointment. It requires showing up when it’s inconvenient. It requires accepting that your friend has needs, moods, and limitations. The discomfort of these constraints isn’t a bug in human relationships. It’s what builds our capacity for genuine connection. When AI removes all friction, it doesn’t enhance our ability to relate; it atrophies it. After extended interaction with endlessly accommodating AI, real humans (with their boundaries, their bad days, their own agendas) become intolerable.
Hinge CEO Justin McLeod captured this dynamic: “While it will feel good in the moment, like junk food basically, to have an experience with someone who says all the right things and is available at the right time, it will ultimately, just like junk food, make people feel less healthy and more drained over time.”
Likewise, Daniel Cox, director of the Survey Center on American Life, warns that AI companions “make feelings of boredom and loneliness easier to manage” but stop people from “doing the things designed to alleviate those feelings.” When we’re lonely and missing human interaction, the healthy response is to seek it out. AI companions may offer a more comfortable alternative, but they ultimately deepen our isolation.
Notice a Pattern?
With social media, we didn’t recognize the loneliness crisis until a decade after Facebook launched. By then, an entire generation had grown up with smartphones, and the damage was embedded in our social fabric. This time, we have the advantage of pattern recognition—and the disadvantage of speed.
ChatGPT reached 100 million users in two months, compared to Instagram’s 2.5 years. Within two years, AI achieved 40% workplace adoption—faster than PCs or the internet at comparable points in their diffusion. More critically, we’re already measuring clinical harm. The colonoscopy study was published in August 2025, less than three years after ChatGPT’s launch. We’re not waiting decades to see effects. We’re watching physicians deskill in real-time while the technology spreads faster than we can study it.
The pattern is visible, but the window for intervention is narrow. If we wait for consensus about AI’s effects the way we did with social media, we’ll be acting on a crisis that’s already entrenched.
Preserve Friction
The evidence suggests we’re repeating previous mistakes, but at a faster rate, owing to the fact that AI is optimized to remove cognitive friction. The companies building these systems have every financial incentive to make them indispensable. The more we use them, the less we can function without them. That’s the business model.
But perhaps what appears as friction is actually function. In my research on emergency department operations, we found that removing “slack” from complex adaptive systems increases fragility and mortality. Systems need redundancy, inefficiency, and adaptive capacity. The same principle applies to human capabilities.
Friction is how we think. Scientific progress requires hypotheses that get challenged. Critical thinking develops through cognitive struggle. Relationships build through reciprocity, through the discomfort of giving when you’d rather receive, of listening when you’d rather speak, of being told “no” by someone who cares enough to disagree. The struggle isn’t what we should eliminate. It’s what we should preserve.
This connects to broader cultural concerns about losing our capacity for difficult conversations, whether framed as “woke mob” censorship or “misinformation” crackdowns. Regardless of political perspective, the underlying anxiety is the same: we’re losing the ability to engage with discomfort. We’re losing the capacity to question, to disagree constructively, and to sit with uncertainty.
The solution isn’t rejecting AI but refusing to let it replace what actually builds that capability. Medical schools should mandate AI-free days so that residents can practice pattern recognition without assistance. Educational systems should teach critical thinking before providing AI tools, just as we teach arithmetic before calculators. Organizations should preserve space for human disagreement before algorithmic consensus.
And individually, we need to recognize when convenience becomes incapacitation. Talk to strangers. Struggle with hard problems before consulting AI. Value expertise even when it contradicts your preferences. Choose relationships that disappoint you, that require effort, that make demands.
The discomfort isn’t the problem to be solved. It’s the mechanism through which we stay human.