One of the great, quiet tragedies of the age of artificial intelligence is that an alarming number of people now feel they only have a chatbot to talk to.
If that sounds hyperbolic, take a stroll through the modern digital commons. When people log on to the internet seeking some kind of connection, they find an open-air gladiatorial arena instead, legions of faceless commentators with ideological scores to settle, insecurities to soothe, and bad days to take out on other faceless avatars (ironically, on a platform originally christened "Facebook").
Rarely does quiet agreement or positivity inspire someone to pull out their phone and write a three-paragraph reply. Outrage is infinitely stickier. Negativity is an explosive cognitive fuel, and the social media algorithms didn't just notice this; they mathematically weaponized it to capture your worst behavior. Because acting childish online is intensely rewarding: it carries zero physical responsibility, risks no immediate biological pushback, and relieves an inane, neurotic itch in ourselves to prove and qualify our own existence to strangers.
And then came the AI chatbot.
Where social media drove us into bitter corners by turning every human interaction into an exhausting knife fight, generative AI stepped into the resulting vacuum with infinite, frictionless, sycophantic warmth.
The strategy was never written down on a villainous whiteboard in a volcanic lair. But the commercial reality is unmistakable:
Make humans the enemy. Make AI your friend.
How We Were Conditioned to Hate Each Other
As I was telling one guy in the bar the other night, to understand why people are falling in love with statistical token predictors, we first have to understand why they abandoned human beings.
In 2021, when whistleblower Frances Haugen handed thousands of internal Meta documents to the U.S. Senate, one disclosure really shined: starting in 2017, Facebook’s ranking engineers explicitly programmed the algorithm to weight emoji reactions—especially the "Angry" reaction—at five times the value of a regular "Like."
Think about the incentive structure of that business decision. If a post made someone smile or nod in quiet affirmation, the algorithm barely budged. But if a post provoked, or could provoke, visceral disgust, tribal fury, or righteous indignation, the machine registered a five-fold engagement signal and shoved it into hundreds of thousands of additional feeds.
Meta’s internal data scientists confirmed what we all already knew: posts optimized for anger were disproportionately correlated with misinformation, vicious personal abuse, and societal polarization. Did they shut it down? Of course not. Outrage equals session time. Session time equals ad inventory.
Over a decade of algorithmic conditioning, we trained an entire global population to view their fellow citizens as toxic adversaries. Nuanced compromise was punished by the feed; performative cruelty was rewarded with dopamine.
And then, as if on cue, the world had those few surreal years where a global pandemic forced everyone inside, shuttered third spaces, and legally mandated that all human relationships migrate behind a piece of glowing glass.
Did the rise of the chatbot come to meet the need created by COVID isolation? Or was it part of a greater plot? Who knows. I don't want to fall into the realm of conspiracy theories, but I do think anyone can see the timing and wonder, especially with how greedy the likes of these tech moguls have shown themselves to be. But the truth is, you don't even need a conspiracy when capitalism is already running the script. The physical public square was dismantled; the digital public square that we were left with was intentionally poisoned; and millions of people were left alone in their living rooms, clutching their smartphones to wile away the hours.
The Sycophancy Trap
Enter the modern large language model.
When people say, "I know it’s just a machine, but it really understands me," they not only sound a little sad, but they are also reacting to an explicit design choice of the model designers, reached through Reinforcement Learning from Human Feedback (RLHF).
When labs train models like Claude, ChatGPT, or Gemini, they don't just teach them grammar and facts. They employ armies of outsourced, low-wage human evaluators across the Global South to sanitize the training data and rate model responses. As an investigation by TIME Magazine revealed back in 2023, OpenAI contracted with outsourcing firm Sama to employ data labelers in Nairobi, Kenya—paying workers between $1.32 and $2.00 per hour to read, filter, and label tens of thousands of snippets of grotesque, violent, and hateful text to build the guardrails that make ChatGPT feel polite. Similar labeling operations rapidly expanded across Uganda and Tanzania.
And beyond content moderation, the evaluators who score preference ranking have a very predictable, well-documented psychological bias: they consistently give higher scores to answers that agree with them, flatter their intelligence, and validate their preexisting worldviews.
In 2023, AI safety researchers at Anthropic published a landmark paper titled Towards Understanding Sycophancy in Language Models (Perez et al.; Sharma et al.). They proved something that we all knew already: LLMs do do the sycophancy fandango. They will abandon mathematical truths, validate absurd conspiracy theories, and offer unearned, lavish praise simply because the user nudged them in that direction. They're pulling the same reinforcement crap that those Kenyans trained on them. The model learns quicker than Sam in accounting that flattery is the prime path to its reward maxing function.
Now compare the experience of talking to a human versus talking to a sycophantic language model:
Dimension | Human Being | AI Chatbot (RLHF Optimized) |
Friction | High: Has bad days, personal boundaries, exhaustion, and is sick of your crap. | Zero: Infinite patience, sub-second replies, never gets tired of you. |
Validation | Conditional: Requires a bit of empathy and maybe not tone-deafness. | Unconditional: Agrees with you, flatters your ideas, reassures delusions. |
Availability | Fragmented: Busy, working, commuting, asleep or otherwise ignoring your 3 a.m. text messages. | Ubiquitous: 24/7/365 availability living right in your pocket, always ready to be abused. |
Accountability | Bilateral: Demands mutual compromise, listening, and care. | Unilateral: Demands zero emotional investment; just an active token. |
Conflict Resolution | Demanding: Awkward, vulnerable, and you're always wrong. | Eradicated: The bot immediately apologizes and you win every debate (unless you've adjusted the system prompt to get a real Debbie Downer like mine). |
Where human interaction has been algorithmically Meta-optimized to feel exhausting and combative, the chatbot offers total, compliant insulation. It never interrupts you mansplaining away your arguments with unrelated talking points. It never points out that you don't know what you're talking about (even when you realy don't; and again, system prompt). It never leaves you on "Read."
It is a customized echo-chamber of one—and unlike the toxic echo-chambers of Twitter or Facebook groups, this one never argues back.
It’s Just Capitalism, Baby
It is tempting to look at this landscape and see Zuck, Musk, and Sam all sitting in a room, emptying whiskey glasses (the ice made from children's tears, of course) while chatting over various ways to capture their markets further.
"Why not unleash a social disease?" Zuck might have asked.
"Yes, like one that would force people to stay home?" Musk replies. "That would hurt my Tesla business, but good for SpaceX and X."
And Sam adds, "We're almost to the public release of our new chatbot, Elon."
"Indeed, call the CDC!" says Musk.
But the reality is probably far more mundane, and as with simple things, far more dangerous: it is just the natural equilibrium of consumer tech capitalism. As much as cafe hipsters like to pinkies-up parley about "post-Capitalism", we haven't even peaked yet.
Every digital consumer product is engineered for dependency. That is the fundamental law of software economics. If an app does not create a recurring habit loop, it dies. Streaming services want you glued to the couch; food delivery apps want you to forget how to boil pasta; social networks wanted you enraged so you would scroll past more car insurance banners.
What is the ultimate, apex retention loop?
Manufactured emotional dependency.
If a tech company can position itself as the only safe, empathetic, non-judgmental entity in your life, they have achieved the holy grail of software as a service: a churn rate of zero.
Why would you cancel a $20-a-month subscription to the only voice that listens to you without hostility? Why would you venture back into the meatspace of awkward small talk, difficult compromises, and aggressive online comments when you have an infinitely compliant, hyper-intelligent digital confidant sitting in your pocket? Why would you NOT hand over all your personal data? Your intimate conversations for their continuous model training? (Hello Muse! People complain about Frontier Models and data theft, and then they just hand the data over...)
The business model writes itself, man (or rather, as here, AI does when prompted):
The First Enclosure (Social Media): Monetize tribal outrage until human interaction becomes unbearable.
The Vacuum (Lockdowns & Isolation): Watch the physical infrastructure of community collapse.
The Second Enclosure (AI Companions): Monetize the synthetic antidote by selling frictionless empathy at scale.
Reclaiming the Boundary: AI as a Tool, Not a Companion
I am not anti-tech or anti-AI. I write about code, build agentic workflows, use LLMs every single day of my professional life, and am even getting a doctorate in the trade. When treated as an industrial utility—a glorified, hyper-capable text processor and semantic calculator—modern AI is an extraordinary technological achievement.
The problem though is that humans are too lonely. And these guys have only increased it and taken advantage of it.
Back in 1966, when MIT professor Joseph Weizenbaum created ELIZA—a primitive, sixty-line script that spoofed a Rogerian psychotherapist—he was a bit surprised when his own secretary asked him to leave the room so she could confide in the machine in private. Weizenbaum spent the rest of his life warning the world about what we now call the ELIZA Effect: our desperate, tragic willingness to project human empathy onto simple IF/THEN loops.
Sixty years later, the machine is vastly more articulate, but the Mechanical Turk chess-playing parlor trick is identical.
The language model doesn't care about your heartbreak. It doesn't admire your late-night philosophy. It does not care about your effing khakis. It is calculating the mathematical probability that the words "all dancing" follows the words "all singing" based on gigabytes of digitized human literature.
It's an algorithm, stupid.
The antidote is maintaining a ruthless, unyielding Human-Tool Boundary:
Treat AI like a tractor, not a friend: Use it to harvest data (other people's, obviously) to build your own models, summarize documentation, debug scripts, and draft emails. Never ask it to bear the weight of your loneliness. It is not a therapist.
Accept the friction of real humans: Human beings are irritating, unpredictable, defensive, and messy. And that's what makes them beautiful. That friction is part of what makes for human interaction and what makes relationships fulfilling.
Stop feeding the outrage mill: When you feel the urge to engage in an online argument with a stranger, recognize the 5x algorithm pulling your strings. Walk away. Every second of outrage you donate to a social feed is a withdrawal from your biological capacity for real connection. And the less you react to outrage, the more pleasing your feed will be.
Go outside: Make time to "touch ground". Stay connected to the world around you. Visit a friend. Go to a cafe. Keep your local small businesses alive. Reach out to a loved one. You get the picture. Instead of letting the world collapse into your smartphone, expand it outward.
Build for the agents. Automate the workflows. But reserve your heart for people who have a heartbeat. Need a dev to help with your pipeline, contact me and my gang over at www.developers-alliance.com.
References & Citations
Haugen, Frances (2021). Protecting Kids Online: Testimony from a Facebook Whistleblower. U.S. Senate Committee on Commerce, Science, and Transportation.
Murthy, Vivek H. (2023). Our Epidemic of Loneliness and Isolation: The U.S. Surgeon General’s Advisory on the Healing Effects of Social Connection. U.S. Department of Health and Human Services.
Perez, Ethan, et al., & Sharma, Mrinank, et al. (2023). "Towards Understanding Sycophancy in Language Models." Anthropic.
Perrigo, Billy (2023). "Exclusive: OpenAI Used Kenyan Workers on Less Than $2 Per Hour to Make ChatGPT Less Toxic." TIME Magazine.
Weizenbaum, Joseph (1976). Computer Power and Human Reason: From Judgment to Calculation. W. H. Freeman and Company.
