You might encounter conversational AI when asking an assistant to explain a document, practising a language with a tutor, speaking to a customer-service system or interacting with an AI-powered character in a game.
Some conversations serve a specific purpose: finding information, resolving a problem or completing a task. Others support learning, creative exploration, entertainment or an ongoing sense of companionship.
AI companion apps and products place particular emphasis on social interaction. They may offer friendship, role-play, romantic conversation or a recurring character that users return to over time.
A useful distinction
Is conversational AI the same as a chatbot?
The terms overlap, but describe different things. A chatbot is a system people interact with through chat. Conversational AI describes the AI capabilities that can support that interaction.
Some chatbots follow predefined rules and scripted paths. Others use AI models to interpret messages or generate replies. Products can combine both approaches—for example, generated conversation alongside a fixed process for account verification.
Conversational AI also includes spoken interactions, so it extends beyond text-based chat.
Another useful distinction
How is it different from generative AI?
Conversational AI concerns interaction through dialogue. Generative AI concerns the creation of content, such as text, images or audio.
Many modern conversational systems use generative AI to produce their replies. But generative AI can also operate without conversation—for example, generating an image from a single instruction.
Conversational systems do not all rely on generative models.
02 / Under the surface
How does conversational AI work?
Different systems use different approaches. Many contemporary products use a large language model, or LLM: a model trained on large amounts of data to generate language in response to the context it receives.
That context might include your latest message, earlier parts of the conversation, instructions about the assistant’s role, or information retrieved from documents and other services.
A simplified view of an LLM-based conversation
01Your messageText or speech
→
The surrounding system
InstructionsConversation historyRetrieved information
Language model
Tools · Permissions · Memory · Safeguards
→
02A responseSometimes an action, too
The model is one component. The product shapes the experience.
The surrounding system determines what information the model receives, what tools it can use, what is stored, and what happens when something goes wrong. Two products using the same underlying model can behave differently because those choices differ.
A system using information from your conversation does not necessarily mean the model is learning from you or being retrained. Conversation history, stored memory and model training are separate processes.
03 / The CAS perspective
Why does conversational AI need more than engineering?
A technically fluent response can still be inaccurate, inappropriate or poorly suited to the person receiving it. Evaluating a conversational system means considering its purpose, its users and the consequences of its behaviour.
Different disciplines bring different questions.
01
Engineering & system design
What can the system access and do? How reliable is it?
02
Psychology & interaction design
What expectations does its language, voice or persona encourage?
03
Safety & ethics
Who could be harmed, and which product choices might increase that risk?
04
Governance & law
Who is accountable, and what obligations apply?
05
Research & evidence
What evidence supports claims about usefulness, safety or benefit?
06
Model welfare
An emerging and contested field asking whether AI models could have subjective experiences, and what protections—including the ability to decline potentially harmful interactions—might be appropriate if they do.
These questions belong throughout the life of a product, from its initial design to updates, incidents and eventual closure.
04 / Our particular curiosity
What changes when conversations continue over time?
One replyWeeksMonths
An isolated exchange gives us a limited view of a conversational system.
An invented example
Imagine a language tutor that remembers a learner’s goals. That continuity could make later lessons more useful. But if it stores an incorrect assumption about the learner, repeatedly ignores corrections or retains information the learner expected it to forget, the same feature introduces problems.
In companion systems, repeated interaction also raises questions about persona consistency, attachment, boundaries and what happens when the service changes or ends.
At CAS, we ask how earlier interactions shape later behaviour, what systems should remember, and how to evaluate experiences that cannot be understood from a single reply.