Blog: Beyond the Chat Box: Rethinking How We Interact with AI

Chat has become the default way of interacting with artificial intelligence. We ask a question, receive an answer, refine the prompt, and continue the conversation. Whether we are searching for information, writing an email, debugging code, analyzing data, or summarizing a paper, the interaction increasingly takes the same form: a text box waiting for us to type/say something. There are good reasons for this. Conversation is familiar. It requires almost no training; natural language lowers the barrier to using sophisticated systems; and a single interface can appear to support tens of different tasks.

But as chat becomes the dominant way we interact with AI, a more fundamental question emerges: should a chatting interface be the only way we engage with AI systems? The question is not whether chatbots are useful. Clearly, they are. The more interesting issue is what happens when we treat conversation as the natural form of interaction for almost every AI system, including tasks where users need to inspect evidence, compare alternatives, understand uncertainty, or make consequential decisions. Recent work by Ghosh and colleagues directly challenges this assumption. They argue that the chatbot is not simply a neutral container around an AI model. It is a design choice that influences how people learn, work, evaluate information, and exercise agency.

When the middle disappears

For instance, consider how we search for information on the internet. Traditional search engines present us with multiple results, and we decide which sources to open, identify where they agree or disagree, assess credibility signals, compare different claims, and eventually form our own conclusion. As shown in Figure 1, however, chat-based search can compress much of this process into a single step. Instead of presenting several pieces of information for users to interpret, a chatbot can synthesize them into one polished response. In traditional information searching, users remain actively involved in the middle of the process, where comparison, interpretation, and evaluation take place. With conversational search, much of this work may occur between the initial prompt and the final response, largely outside the user's view.

Figure 1: The disappearing middle in AI-mediated search.

This does not mean that chatbot-based search is necessarily worse than traditional search. In many situations, having information synthesized into a clear response is useful and efficient. The concern is that, when the intermediate steps are hidden, users may have fewer opportunities to see how the answer was formed. Disagreement between sources may be smoothed over, important qualifications may be omitted, and alternative interpretations may never appear in the final response. As a result, users are presented with a conclusion without always being able to examine the evidence and comparisons behind it. The challenge, therefore, is not simply to improve the quality of AI-generated answers, but to design search interactions in ways that preserve enough visibility for users to understand, compare, and evaluate the information themselves.

This reduced visibility also raises a broader concern about how conversational AI may affect the way users think and learn. Chatbots make complex tasks remarkably easy to begin, since users no longer need to understand even the structure of an unfamiliar topic before asking for help. Yet some of the effort being removed is not simply an inconvenience. Activities such as formulating a clear problem, deciding what evidence is relevant, identifying contradictions, and constructing a conclusion are important cognitive processes through which expertise develops. Repeatedly delegating these steps to AI may therefore reduce opportunities for users to practice them. This is particularly relevant to metacognition, or the ability to monitor and regulate one's own thinking. Tankelevitch et al note that effective use of AI still requires users to clarify their goals, break complex tasks into manageable parts, and recognize when an approach is not working. In other words, using AI effectively still requires people to think about their own thinking, yet current chat-based interfaces do not always encourage that kind of active reflection.

What if AI did not always talk to us?

These concerns suggest that not every AI system needs to rely on a chat interface. Some systems could still use conversation while giving users more control, such as showing alternative interpretations, comparing evidence, or making sources more visible. In other cases, non-conversational designs may work better. Researchers could compare explanations side by side, and students could complete part of a task before receiving help. Such designs may keep users more actively involved in the process rather than hiding everything behind a single response.

Figure 2: Moving beyond chat-only interaction.

These do not require abandoning conversational AI; instead, they shift more control back to the user. As shown in Figure 2, this can involve higher-agency chat systems, non-conversational interfaces, or modular tools with more visible components. The key idea is to preserve opportunities for users to understand, compare, intervene, and make decisions rather than allowing the system to handle the entire process invisibly. This may be especially important in areas such as education, research, and public decision-making, where the quality of the process can matter as much as the convenience of the final result.

Note: Lumo (https://lumo.proton.me/) has been used for polishing the text.

References:

• What if AI systems weren't chatbots?
• The Metacognitive Demands and Opportunities of Generative AI
• Enhancing Critical Thinking in Generative AI Search with Metacognitive Prompts

- Written by Hamayoon Behmanush (https://www.hamayoon.me/)