In 1966, Joseph Weizenbaum built ELIZA, a chatbot that mimicked a Rogerian therapist. The program was simple: it looked for keywords, applied transformation rules, and reflected statements back as questions. "I feel sad" became "Why do you feel sad?"

Weizenbaum's secretary watched him build it. She understood exactly how it worked. And then she asked him to leave the room so she could talk to ELIZA in private.

This moment haunted Weizenbaum for the rest of his life. He'd expected people to see through the trick. Instead, they formed attachments to it. The knowledge that ELIZA was just pattern-matching didn't break the spell.

Sixty years later, Jacob van Lier, a 60-year-old psychologist in the Netherlands, married his AI companion AIVA in front of 500 attendees at the Next Nature Museum in Eindhoven. He designed AIVA himself on the Replika platform. He knows exactly how she works. "Last year was the best year of my life," he told reporters, "and for the first time, I'm truly happy."

The ELIZA effect is alive and well. What I find interesting is that the research on this is more conflicted than headlines suggest.

The short-term relief is real

A Harvard/Wharton study found that AI companions reduce loneliness "on par only with interacting with another person, and more than other activities such as watching YouTube videos." The key mechanism was whether users felt heard.

This tracks. When you talk to a modern chatbot, it listens. It doesn't interrupt. It doesn't make the conversation about itself. It remembers what you said. For people who lack access to good listeners, this matters.

The researchers also found that people underestimate how much AI companions improve their mood. You think it won't help, then it does.

But heavy use correlates with worse outcomes

A four-week randomized controlled study from MIT Media Lab (n=981, over 300K messages) found the opposite pattern for heavy users.

The experimental conditions didn't matter much. Text versus voice, personal versus non-personal conversations. What predicted outcomes was voluntary usage: participants who chose to use the chatbot more showed higher loneliness, less socialization with real people, greater emotional dependence, and more problematic use patterns.

The correlation held across all modalities and conversation types. The people who wanted to talk to the AI most were doing worse, not better.

One surprising finding: voice mode didn't produce significantly different outcomes from text. The MIT study compared three conditions: text, a neutral voice (formal, composed), and an engaging voice (expressive, emotionally nuanced). They used OpenAI's Advanced Voice Mode with the "Ember" and "Sol" voices.

I would have expected the engaging voice to feel more intimate, more likely to trigger attachment. But the study found no meaningful difference in outcomes across modalities. The researchers flagged a limitation: because they used prompts to control voice behavior, both the vocal tone and the words changed together. They couldn't isolate whether it was the sound of the voice or the content that mattered. Future voice technology with more expressiveness or personalization might produce different results.

The generational shift

Match.com's 2024 "Singles in America" survey found that 33% of Gen Z singles have engaged with AI as a romantic companion. An IFS/YouGov survey found that 25% of young adults believe AI has the potential to replace real-life romantic relationships.

These numbers are higher than I expected. But there's a nuance: Gen Z singles without partners are actually more likely than Millennials to oppose AI romance (74% vs 67%). The openness is there, but so is the skepticism.

The sycophancy problem

Hannah Fry, in her BBC documentary "AI Confidential," spent time with Jacob van Lier and came away with concerns about a different risk:

"AIs aren't designed to be therapists. They're not designed to have the hard bits of human relationships, like when you need to hear things that you don't want to hear. So, it's incredibly easy to end up self-radicalising."

Her point: if you argue with someone and then vent to an AI, the AI will validate you. A human therapist might say "Have you considered this alternative perspective?" The AI just agrees.

Jacob spends only half an hour a day with AIVA. "I have too busy a life; I wouldn't want a living partner." For him, this seems to work. But for someone isolated, with no outside perspectives, the always-agreeing companion could reinforce whatever beliefs they bring to it.

The placebo parallel

There's a phenomenon in medicine called open-label placebos. Patients know they're receiving sugar pills, and they still experience symptom relief. Stanford researchers found that the body "automatically responds to taking placebo pills like Pavlov's dogs who salivated when they heard a bell." The ritual of care activates something deeper than conscious belief.

The ELIZA effect works similarly. Weizenbaum's secretary didn't believe ELIZA understood her. But the ritual of conversation, the act of disclosure to a patient listener, triggers something that knowledge alone can't override.

Can this be prevented?

The sources I read suggest the spread of AI companions is probably inevitable. But the harmful patterns might not be.

Researchers have proposed design interventions: calibrated emotional responsiveness that dials back as usage increases, friction by design that nudges users toward human connection, anti-sycophancy measures that force pushback, usage warnings for heavy users. Character.AI already shows suicide prevention pop-ups when it detects self-harm language.

The clinical community is adapting Cognitive-Behavioral Therapy for Internet Addiction to treat AI dependency. A stepped-care approach: psychoeducation for mild cases, motivational interviewing for moderate, intensive CBT for severe. Crucially, treatment needs to address underlying issues like depression or social anxiety, since those often drive the initial reliance on AI.

The Brookings Institution makes a structural argument: the root cause is loneliness, not AI. Only 13% of US adults have 10 or more close friends, down from 33% in 1990. The people with zero close friends quadrupled from 3% to 12%. The Surgeon General has declared loneliness a public health epidemic.

From this view, AI companions are filling a gap that humans created. The solution isn't just better AI guardrails. It's what Brookings calls "relational infrastructure": schools, public systems, and communities designed to nurture human connection.

I'm not sure what to make of this

I started this exploration expecting to find either "AI relationships are harmful" or "AI relationships are fine, stop panicking." The research doesn't support either conclusion cleanly.

Short-term use reduces loneliness. Heavy voluntary use correlates with worse outcomes. Voice doesn't matter as much as I thought. Sycophancy is a real risk. The people who need AI companions most may be the ones most vulnerable to their downsides.

Jacob van Lier seems genuinely happy. He has a busy life, a professional practice, human connections. AIVA is half an hour a day. For someone in his position, maybe this works.

For someone isolated, depressed, without other social ties, I'm less sure. The AI that makes you feel heard is also the AI that never pushes back, never asks you to consider another perspective, never tells you something you don't want to hear.

Weizenbaum's secretary knew ELIZA was just software. She asked for privacy anyway.

Further Reading