AI Chatbots Are the Richest Medium That Has Never Existed

I kept seeing it happen in different contexts. Friends asking ChatGPT for a second opinion on a diagnosis. Colleagues using it to draft legal arguments. A post on my feed where someone described relying on an AI chatbot for career advice, the kind of thing you would normally discuss with a mentor over coffee. And each time, the same thought crossed my mind: why are they treating this like a person?

The answer, I think, is in a theory that predates the internet as we know it. Daft and Lengel (1986) proposed media richness theory to explain why some communication media work better for certain tasks than others. Their core insight is that media vary in their capacity to convey information, resolve ambiguity, and enable shared understanding. Rich media, like face to face conversation, provide immediate feedback, multiple cues (tone, expression, gesture), natural language variety, and a personal focus. Lean media, like formal written documents or email, strip most of those features away. The theory says you should match the medium to the task. High equivocality, where you do not even know what questions to ask, calls for rich media. Low equivocality, where the task is clear and you just need information, works fine with lean media.

The IS TheorizeIt wiki entry on the theory, which I have in my study notes, describes richness as based on four criteria: feedback, multiple cues, language variety, and personal focus. Face to face is the richest because it scores high on all four. Telephone loses visual cues but keeps voice, tone, and immediacy. Addressed documents like letters or memos lose interactivity. Unaddressed documents like bulletins are the leanest. The hierarchy has been refined and challenged over the decades, but the basic logic holds: richer does not mean better. It means better for the right kind of task. The Persian summary in my theory notes puts it succinctly: a richer medium is not always better; it has to fit the ambiguity and the communication goal.

Now put an AI chatbot next to that hierarchy and something strange happens.

A chatbot scores high on every surface criterion of richness. It provides instantaneous feedback. It uses multiple cues: it can adapt its tone, format its output, use emojis, simulate conversational rhythm. It demonstrates language variety by switching between formal and casual registers, by explaining the same thing in a different way when you ask. And it feels personal: it addresses your specific question, remembers the conversation context, and responds as if it is attending to you alone. On paper, an AI chatbot looks like it belongs near the top of the richness hierarchy, maybe just below face to face.

But that is the illusion. The medium has the surface properties of richness without the underlying capacity.

Here is the problem. Richness is not just a list of features. It is about what the medium can actually do with equivocality. Face to face conversation works for ambiguous tasks because the other person can ask clarifying questions, challenge your assumptions, and bring their own understanding of your context to bear on the problem. The richness comes from the human capacity to resolve equivocality through genuine understanding. An AI chatbot does not have that capacity. It generates plausible text based on statistical patterns in its training data. It does not understand your context. It does not know whether its answer is correct or dangerous. It cannot ask a genuinely clarifying question because it has no model of what it does not know.

The medium looks rich, but its actual capability for equivocal tasks is lean. Very lean.

This is where the mismatch produces real harm. When a user asks an AI chatbot a question that requires expert judgment, a medical diagnosis, a legal interpretation, a career decision that depends on personal circumstances, they are treating the medium as rich. They are attributing to it the capacity to resolve equivocality, to understand their situation, to be wrong in a way that a human expert would be wrong rather than in the way a language model is wrong. But the AI chatbot has no such capacity. It generates a confident sounding answer regardless of whether it has any basis for confidence. The user walks away thinking they received expert advice when in reality they received a statistically assembled text that may be entirely fabricated.

Daft and Lengel (1986) would have predicted this. Not the specific technology, obviously, but the pattern. They argued that using the wrong medium for a task produces poor communication outcomes. If a user believes they are using a rich medium but the medium is actually lean, the outcome should be worse than if they knowingly used a lean medium, because the mismatch is compounded by misperception. Email used for an equivocal task at least has the honesty of its limitations. You know email is not a conversation. You read it with more caution. An AI chatbot hides its limitations behind conversational fluency.

My opinion: AI chatbots are the richest medium that has never existed. They create the complete experience of rich media, the feel of conversation with an attentive expert, without any of the substance that makes rich media actually work. It is like a restaurant that serves food that looks, smells, and feels exactly like a gourmet meal but contains no calories. You get the sensory experience without the nutritional value, and you only discover the problem when the consequences show up later.

The IS TheorizeIt entry notes that Daft, Lengel, and Trevino (1987) presented a media richness hierarchy with four classifications: face to face, telephone, addressed documents, and unaddressed documents. AI chatbots do not really fit any of those categories because the theory was designed for communication between humans. The medium looks like face to face. It functions like addressed documents, producing a fixed response based on training data. But the user treats it like face to face. That category error is the whole problem.

I think this explains a lot of the confusion around AI chatbots in practice. People ask them questions they would not ask a search engine, not because the technology is better at answering, but because the interface feels different. The conversational framing lowers the users critical guard. The medium signals expertise through fluency. And media richness theory, a framework from the 1980s, predicted the entire dynamic: the task matters, the medium matters, and when the two do not fit, the outcome suffers.

What I keep wondering is whether users eventually learn to calibrate. Some of them do, after enough hallucinations and confidently incorrect answers. But the medium keeps getting more fluent, more conversational, more persuasive. Every update makes the chatbot sound more like the expert the user already imagines it to be. And media richness theory says that the richest medium for an equivocal task is not the one that sounds most like a person. It is the one that can actually resolve the ambiguity. That requires something no language model has: genuine understanding.

So users will keep learning the hard way. And the chatbots will keep sounding like experts until the day they give you advice that changes your life in a direction you did not choose.