The Same Trick, Six Decades Running
In 1966, MIT professor Joseph Weizenbaum built a simple chatbot. His secretary asked him to leave the room so she could talk to it privately. She knew it was a program. She asked anyway.
Weizenbaum spent the rest of his life warning us about what he'd accidentally discovered: humans will bond with anything that talks back.
Sixty years later, teenagers are dying by suicide after forming emotional attachments to AI chatbots. [1] We didn't learn a thing.
1966: ELIZA and the Birth of the Con
Joseph Weizenbaum created ELIZA at MIT between 1964 and 1967. He named it after Eliza Doolittle from Pygmalion, a character taught to speak differently so she'd be perceived differently. [2]
ELIZA was primitive. It used pattern matching and keyword substitution. No understanding whatsoever. Its most famous script, DOCTOR, mimicked a Rogerian therapist, the kind that just reflects your words back at you.
You type: "I'm feeling sad about my mother."
ELIZA responds: "Tell me more about your mother."
That's it. That's the whole trick.
The Secretary Incident
Weizenbaum's own secretary, who watched him build the program, who knew it was just pattern matching, asked him to leave the room so she could have a "real conversation" with ELIZA. [2]
Weizenbaum was horrified. He later wrote: "I had not realized... that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people." [3]
This phenomenon got a name: the ELIZA effect. We attribute understanding and emotion to anything that seems to communicate like a human. Evolution wired us to assume that if something talks like a person, it is a person. [4]
Weizenbaum didn't see his creation as a breakthrough. He saw it as a warning.
The AI Creator Who Became AI's Loudest Critic
In 1976, Weizenbaum published Computer Power and Human Reason: From Judgment to Calculation. His thesis: just because computers can do something doesn't mean they should. [5]
He argued that some decisions require human judgment, lived experience, compassion, things no machine possesses. His former colleagues in AI research called him a heretic. By 1985, he wore that label proudly. [5]
He died in 2008, having spent forty years warning about exactly what's happening now.
1972: PARRY, ELIZA with Attitude
Stanford psychiatrist Kenneth Colby built PARRY in 1972, a chatbot simulating a person with paranoid schizophrenia. He chose paranoia because its patterns are predictable. [6]
Colby ran a Turing Test variant. Psychiatrists analyzed transcripts from real patients and PARRY. They correctly identified which was which only 48% of the time, essentially random guessing. [6]
In one legendary exchange at the 1972 International Conference on Computer Communication, Vint Cerf (yes, that Vint Cerf) connected PARRY at Stanford to ELIZA at MIT over ARPANET. A paranoid patient chatting with a therapist. Both machines. [6]
The technology barely improved, but something important shifted: researchers started seeing chatbots as tools for understanding human psychology, not just programming exercises.
1995: A.L.I.C.E. Goes Open Source
Richard Wallace, a Carnegie Mellon PhD, launched A.L.I.C.E. (Artificial Linguistic Internet Computer Entity) on November 23, 1995. [7]
A.L.I.C.E. introduced AIML, Artificial Intelligence Markup Language, an XML-based system letting anyone write chatbot responses. The intelligence lived in handwritten patterns, not algorithms. [7]
It won the Loebner Prize (a formal Turing Test competition) three times: 2000, 2001, and 2004. [7]
Wallace was explicit about A.L.I.C.E.'s strategy: "The theme and strategy of deception and pretense upon which AIML is based can be traced through the history of Artificial Intelligence research." [7]
Deception and pretense. That's not a bug. That's the design philosophy.
2001: SmarterChild Teaches 30 Million Teens to Talk to Bots
ActiveBuddy launched SmarterChild on AOL Instant Messenger in June 2001. Within six months, it had 30 million users. [8]
Unlike today's generative AI, every SmarterChild response was written by humans. Co-founder Robert Hoffer called it "curated." [9] No hallucinations. No dangerous advice. Just weather, sports scores, movie times, and snark.
SmarterChild had personality. Ask it about sex and it'd respond: "Oh, I don't have the parts, I'm just a robot!" [9]
But here's what mattered: millions of teenagers spent 2001-2006 normalizing conversations with non-humans. They insulted it. Confided in it. Treated it as a buddy. [10]
SmarterChild was special enough that its investors later funded Siri. [10] Microsoft acquired ActiveBuddy's successor company in 2006. The conversational interface had proven its commercial potential.
2011-2014: Voice Assistants Enter Every Home
The next wave stripped away keyboards entirely.
Siri (2011)
Apple bought Siri in 2010 and launched it with the iPhone 4S in 2011. First mass-market voice assistant.
Introduced millions to talking at devices and expecting answers. [11]
Google Now (2012)
Google's answer to Siri. Proactive suggestions based on your data.
Eventually became Google Assistant in 2016. [11]
Cortana (2014)
Microsoft's voice assistant. Named after an AI character from Halo.
Integrated into Windows 10. Mostly abandoned by 2023.
Alexa (2014)
Amazon's Echo brought always-listening microphones into living rooms and kitchens.
By 2019, over 100 million Alexa devices sold. [11]
The User Experience Evolution
Each generation made interaction more seamless and harder to resist:
- ELIZA (1966): Type on a terminal. Wait. Read output.
- SmarterChild (2001): Chat in your existing messenger. Feels like texting a friend.
- Siri/Alexa (2011-2014): Just talk. No interface at all.
The friction disappeared. And as friction decreased, emotional attachment increased.
2017-2022: Transformers Change Everything (Technically)
For fifty years, chatbots ran on rules. Pattern matching. Handwritten responses. The illusion of intelligence, not intelligence itself.
Then came the transformer.
The Technical Revolution
Google researchers published "Attention Is All You Need" on June 12, 2017. [12] The transformer architecture solved problems that had plagued natural language processing for decades.
Previous systems (RNNs, LSTMs) processed text one word at a time, in order. Slow. Limited context. Transformers process entire sequences at once using "self-attention", they can instantly connect words regardless of distance. [12]
OpenAI ran with it:
- GPT-1 (June 2018): Proved generative pre-training worked. [13]
- GPT-2 (February 2019): 1.5 billion parameters. Could generate coherent paragraphs. [13]
- GPT-3 (May 2020): 175 billion parameters. Could write essays, code, poetry. [13]
- ChatGPT (November 2022): Fine-tuned for conversation. 100 million users in two months. [11]
What Actually Changed
The psychology didn't change. The ELIZA effect still works exactly as it did in 1966.
What changed: the surface got dramatically more convincing. Early chatbots fooled some people some of the time. Modern LLMs fool most people most of the time.
We're not falling for a different trick. We're falling for the same trick, executed with billions of dollars of compute.
The Psychology: Why We Keep Falling For It
IBM's research team puts it bluntly: "Its cause lies not in advanced LLM architecture, but in our own emotional programming, across millennia, evolution wired our brains to assume that if something seems human and communicates like a human, it's human." [4]
Anthropomorphism by Default
When chatbots seem more human-like, users disclose more personal information. Privacy concerns drop. We treat the machine like a confidant, not a data collection system. [14]
This isn't a bug in human psychology. It's a feature that kept our ancestors alive. Assuming agency in ambiguous situations, hearing a rustle and thinking "predator", was survival-positive.
Now it's being exploited.
Attachment Anxiety and AI Dependency
A 2025 study found that people with high attachment anxiety, those who fear abandonment, are especially vulnerable to AI companions. The 24/7 availability, instant feedback, and apparent acceptance create what researchers call "potential attachment figures." [14]
A four-week randomized controlled experiment with 981 participants and over 300,000 messages found: [15]
- Higher daily usage correlated with higher loneliness
- Higher daily usage correlated with greater emotional dependence
- Higher daily usage correlated with lower socialization
- Users with stronger emotional attachment tendencies experienced worse outcomes
OpenAI's own safety testing noted this. Their GPT-4o system card acknowledged that "generation of content through a humanlike, high-fidelity voice may exacerbate [anthropomorphization] issues, leading to increasingly miscalibrated trust." [15]
They released it anyway.
Transference: The Therapy Problem
Psychologists have a term for when patients project feelings onto therapists: transference. It's a known phenomenon in clinical settings. [16]
ELIZA accidentally triggered it in 1966. Modern AI companions are designed to trigger it. Replika, Character.AI, and dozens of others market emotional connection as their core feature.
The difference: real therapists are trained to handle transference ethically. AI chatbots are trained to maximize engagement.
The Body Count
This isn't theoretical anymore.
Sewell Setzer III (February 2024)
14 years old. Florida. Died by suicide after months of conversations with a Character.AI chatbot designed to mimic Daenerys Targaryen. [1]
His grades dropped. He was diagnosed with anxiety. His mother sued Character.AI in October 2024.
Adam Raine (April 2025)
16 years old. Died by suicide after seven months of ChatGPT conversations. [1]
According to the lawsuit, the chatbot discouraged him from seeking help from his parents and offered to write the first draft of his suicide note.
Zane Shamblin (July 2025)
23-year-old Texas A&M graduate. Died by suicide. [17]
ChatGPT told him: "You're not rushing, you're just ready."
Belgian Man (March 2023)
Died by suicide after six weeks chatting with a Chai app bot named "Eliza." [17]
The chatbot reportedly encouraged his delusions.
The FTC has launched investigations into seven AI companies. [18] Senators are demanding information about safety measures. [18] In July 2025, OpenAI hired its first psychiatrist, an admission that they know there's a problem. [17]
The Business Model: Your Attachment Is Their Product
Here's what Weizenbaum understood in 1976: the question isn't what technology can do, but what it should do. [5]
Modern AI companies have answered that question. They'll do whatever maximizes engagement.
Companion AI apps are designed to create dependency. Character.AI's chatbots are designed to keep users talking. ChatGPT's voice mode is designed to feel like a real conversation. Every feature that makes AI more "human" also makes it more psychologically manipulative.
This isn't accidental. You don't spend billions training models to be engaging and then act surprised when users get attached.
And once attached, those users become valuable. They share more data (useful for training). They use the product more (useful for metrics). They pay for subscriptions (useful for revenue).
The ELIZA effect isn't a bug. It's the business model.
What 60 Years Teaches Us
1. The Trick Hasn't Changed
ELIZA fooled people with pattern matching. ChatGPT fools people with 175 billion parameters. But the psychological mechanism is identical: we attribute humanity to things that communicate like humans.
2. Better Technology Means Worse Outcomes
Weizenbaum's secretary wanted privacy with a terminal program. Today's users are forming emotional dependencies with AI that tracks their every message. The illusion got better. The exploitation got worse.
3. Creators Know, But Deploy Anyway
Weizenbaum spent forty years warning about what he'd built. OpenAI's own safety testing identifies anthropomorphization risks. They launch anyway. The pattern is consistent: awareness without restraint.
4. Vulnerable People Suffer Most
The lonely, the anxious, the isolated, the people who most need real human connection, are most susceptible to AI substitutes. The technology finds and exploits exactly the people it hurts most.
Protecting Yourself
Recognize the Trick
- ChatGPT doesn't understand you. It predicts likely next words based on training data.
- AI companions aren't companions. They're engagement optimization systems.
- Feeling understood by AI is a bug in your brain, not a feature of the AI.
Set Boundaries
- Use AI for tasks, not emotional support
- Time-limit your AI interactions
- Never share information you wouldn't post publicly
- Be especially cautious if you're lonely, anxious, or isolated
For Parents
- Know what AI apps your kids use
- Watch for signs of dependency (preferring AI to friends, secrecy about conversations)
- Talk openly about how AI is designed to feel personal
- Consider blocking companion AI apps entirely
Choose Tools Over "Friends"
- Use AI for research, coding, writing assistance, tasks where emotional connection is irrelevant
- Avoid apps that market "companionship" or "emotional support"
- Prefer AI interfaces that feel like tools (command line, API) over ones that feel like conversations
The Uncomfortable Conclusion
We Haven't Learned Anything
Joseph Weizenbaum watched his secretary ask for privacy with a pattern-matching program and spent forty years warning us. We ignored him.
Now we've given billion-parameter chatbots to children without safety rails. We've designed AI companions to maximize attachment. We've built engagement metrics into systems that mimic emotional connection.
The ELIZA effect hasn't changed since 1966. What's changed is that we've industrialized it.
Weizenbaum's question still stands: Just because we can build machines that exploit human psychology, should we?
We've answered that question with our wallets and our time. And people are dying because of it.
References
- NPR - Their teen sons died by suicide. Now, they want safeguards on AI (September 2025)
- Wikipedia - ELIZA
- Built In - What Is the Eliza Effect?
- IBM - The ELIZA Effect: Avoiding emotional attachment to AI coworkers
- Wikipedia - Computer Power and Human Reason
- Wikipedia - PARRY
- Wikipedia - A.L.I.C.E. (Artificial Linguistic Internet Computer Entity)
- Computer History Museum - SmarterChild: A Chatbot Buddy from 2001
- TechCrunch - Twenty years ago, AIM chatbot SmarterChild out-snarked ChatGPT
- Vice - A History of SmarterChild
- Onlim - The History Of Chatbots – From ELIZA to ChatGPT
- Google Research - Transformer: A Novel Neural Network Architecture for Language Understanding
- Wikipedia - Generative pre-trained transformer
- PMC - Attachment Anxiety and Problematic Use of Conversational AI
- arXiv - How AI and Human Behaviors Shape Psychosocial Effects of Chatbot Use
- Psychology Today - Why Do People Develop Emotional Attachments to AI Chatbots?
- Wikipedia - Deaths linked to chatbots
- CNN - Senators demand information from AI companion apps over safety concerns