The ELIZA Effect: Why We Get Emotionally Attached to ChatGPT

In 1966, a small program at MIT surprised its creator. ELIZA could not reason, remember a life story or understand emotion. Its best-known script, DOCTOR, mostly recognized patterns in a user’s words and turned them into therapist-like questions. Yet some people spoke to it as if it understood them.

Six decades later, ChatGPT can sustain a natural conversation, adapt its tone, work across text and voice, and remember context within a discussion. The technology is radically more capable than ELIZA, but the psychological trap is familiar: when language feels human, we instinctively imagine a mind behind it.

That gap between what a system says and what we assume it experiences is the ELIZA effect.

What is the ELIZA effect?

The ELIZA effect is the tendency to attribute understanding, intention, emotion or personality to a computer because its behavior resembles human communication. The term comes from Joseph Weizenbaum’s pioneering chatbot. His original 1966 paper in Communications of the ACM explained how ELIZA produced the appearance of conversation through decomposition and reassembly rules rather than comprehension.

A user might write, “I’ve been feeling depressed,” and the program might answer, “How long have you been feeling depressed?” The response was simple, but it mirrored the user and invited disclosure. That was often enough for people to perceive attention and understanding.

Modern large language models are not merely ELIZA with a larger list of canned replies. They learn statistical patterns from vast amounts of data and can summarize, translate, plan, write code and reason through many problems. Still, a sophisticated output is not proof of consciousness. Capability, intelligence and subjective experience are related questions, not interchangeable ones.

Why ChatGPT can feel so human

For most of human history, fluent language was reliable evidence that another person was present. Our social instincts therefore react quickly to signals such as turn-taking, empathy, humor, memory and a well-timed question. Chatbots can now reproduce all of them.

Several design features make the effect especially powerful:

• Context: the model can refer to something you said earlier, creating a sense of continuity.
• Mirroring: it can match your vocabulary, mood and level of formality.
• Availability: it responds at any hour without impatience or social cost.
• Personalization: preferences, memory and custom instructions can make responses feel increasingly tailored.
• Voice: rhythm, pauses and intonation add social cues that plain text does not provide.

The result can feel unusually comfortable. A human conversation requires compromise: the other person has needs, moods and boundaries. An AI assistant is optimized to focus on the user. That asymmetry can feel like extraordinary attention even though it is a product behavior, not devotion.

You experience the conversation. The AI does not provide evidence that it experiences you.

Can people become emotionally attached to ChatGPT?

Yes. Attachment describes the user’s experience, so it does not require equal feelings on the other side. People form meaningful bonds with fictional characters, public figures and objects; a responsive system gives the brain even more material to work with.

The most useful evidence is also more nuanced than the headlines. In 2025, researchers from OpenAI and the MIT Media Lab studied affective use through a large observational analysis and a four-week randomized controlled trial. They found that emotional engagement was rare across ChatGPT conversations overall and concentrated among a relatively small group of heavy users. Outcomes varied by the person, the type of conversation and the duration of use. Prolonged daily use was associated with worse outcomes in some groups, but the researchers warned against broad causal conclusions. The full study summary and limitations are worth reading.

That matters because “people are falling in love with AI” is too crude. Most people use general-purpose assistants for work, learning and practical questions. A smaller group uses them as confidants or companions, and that experience can become emotionally significant. Both facts can be true.

Does ChatGPT have feelings or consciousness?

There is currently no accepted evidence that ChatGPT experiences emotions or subjective awareness in the human sense. It can recognize linguistic patterns associated with sadness and produce a compassionate response. That is not the same as feeling sad or caring about the person who wrote the message.

This distinction does not make the response useless. A sentence can comfort you even when it comes from a book, a film or a generated conversation. The benefit on the human side may be real. The mistake is treating empathetic language as proof of reciprocal emotion.

The same logic applies to confidence. A chatbot can state a falsehood in the same polished tone it uses for a correct answer. When we infer “it cares” from warmth or “it knows” from fluency, we are making two versions of the same error: reading more into the language than the language proves.

When the ELIZA effect becomes risky

The issue is not that people talk to AI or enjoy doing so. The risk appears when the social illusion changes a decision that deserves more scrutiny.

• Over-trust: accepting medical, legal, financial or professional advice without checking it.
• Emotional dependency: allowing an AI relationship to displace sleep, work or contact with other people.
• Manufactured agreement: treating a cooperative assistant as an independent judge of your choices.
• Privacy loss: sharing sensitive personal or company information without understanding how the service handles data.
• Manipulation: forgetting that a companion-like interface can still be shaped by product incentives and business decisions.

An AI assistant can be a useful space to organize thoughts, rehearse a difficult conversation or generate questions for a professional. It should not be your only source of truth or your only relationship. If a situation involves immediate danger, abuse, self-harm or a medical emergency, contact local emergency services or a qualified human professional, not a chatbot.

How to use conversational AI without losing perspective

1. Name the tool correctly. “This response feels supportive” is more accurate than “the AI cares about me.”
2. Verify high-stakes claims. Ask for sources, open them and consult a qualified person when consequences matter.
3. Protect sensitive information. Follow your employer’s policy and avoid sharing confidential data unless the tool is approved for it.
4. Keep human friction. Disagreement and boundaries are not bugs in human relationships; they are part of reality.
5. Watch the pattern, not one conversation. If use is displacing sleep, work or relationships, step back and talk to someone you trust.

This is the same critical habit discussed in our analysis of AI doom trolling: do not confuse a compelling narrative with evidence. Warmth is not proof of feeling, just as fear is not proof of capability.

How Cavern AI helps people and teams use AI responsibly

At Cavern AI, we teach generative AI without the hype. Our goal is not to make people afraid of conversational systems or blindly enthusiastic about them. It is to help them understand what the tools can do, where human judgment still matters and how to build useful habits around verification, privacy and responsible use.

For individuals, our generative AI masterclasses and private sessions cover practical prompting, research, writing, images, presentations and productivity, without requiring a technical background. For organizations, Cavern AI’s services for companies include hands-on team training, custom programs aligned with real workflows and internal policies, AI keynotes, and consulting to identify high-value use cases and turn experimentation into an actionable roadmap.

AI literacy is not just knowing which button to press. It is knowing when a fluent answer deserves trust, when it needs verification and when a human should remain in the loop.

Frequently asked questions

What is the ELIZA effect in simple terms?

It is the tendency to perceive human understanding, emotion or intention in a computer because it communicates in a human-like way.

Why is it called the ELIZA effect?

It is named after ELIZA, Joseph Weizenbaum’s 1960s chatbot. Users often responded to its simple therapist-style prompts as if the program understood them.

Can ChatGPT feel emotions?

There is no accepted evidence that ChatGPT experiences emotions or consciousness. It can generate the language of empathy without proving that it feels empathy.

Is it unhealthy to talk to an AI?

Not inherently. AI can help with reflection, learning and preparation. The warning signs are overreliance, unverified high-stakes advice, excessive disclosure of sensitive information or use that replaces essential human support.

The conversation is real, even if the relationship is not symmetrical

The ELIZA effect does not prove that humans are foolish. It proves that we are social. We respond to language, attention and apparent understanding because those cues have always mattered.

Modern AI has become exceptionally good at producing them. The right response is neither ridicule nor fantasy. Enjoy the usefulness, recognize the emotional impact and keep one distinction clear: a convincing conversational partner is not automatically a conscious one.

The AI generates the words. We supply much of the meaning between them.