In her Building Expertise into Generated Conversations talk at Convey UX Susan Hura talked about the impact of Large Language Models on the role and process of conversational designers. Here are my notes from her talk:
- It is a weird time to be a conversation designer. Two years ago Siri and Google Assistant raised people's consciousness. People started to become willing to engage. But ChatGPT really changed the situation as people realized its capabilities.
- So how are large language models and generative AI are impacting this domain of design?
- Speech is getting things in my head into your head. We used to think this is a metaphor but when you measure the brain waves between speakers, they match across conversation. So conversation is really critical.
- AI has been used in various parts of conversation for years. NLG: natural language generation, TTS: text to speech, ASR: automatic speech recognition (sounds to words), NLU: natural language understanding.
- But we used to have to build custom language models to get to something that looks like understanding. It was a challenging issue. Large language models only work with few shots. You can train a model now with just dozens of examples.
- Conversation is an over learned behavior. We learn it very early in life (about a year old) and automatically. No one needs to teach us how to learn to speak.
- We are designing for one of the most human of behaviors. Their are all kinds of social, relationship, and emotional triggers in conversations.
- The most important elements of conversation are not tied to language, it’s a back and forth. You are in relationship in a conversation. When you get it right, it establishes trust.
- We don’t want people to have to think about how to talk to a computer, instead computers need to play by the rules of conversation.
LLMs for Conversational Design
- What we can do today is play to the strengths of generative AI models. They're great at synthesizing huge amounts of information so use them to help analyze large chunks of data.
- For example, one of the things that's important in conversation design is establishing a conversational style guide. Mature organizations have UX writing guidelines and a voice and tone guide.
- For smaller organizations, you can use LLMs to draft these resources from an organization's Website by analyzing how they talk to customers.
- You can also use Generative AI to analyze unstructured user research results especially open-ended questions that take a long time to categorize.
- Generative AI can also give us raw materials at design time. For example, the first deliverable for conversation design is often a sample dialogue. But once you get more than a small handful of them, it becomes a maintenance nightmare. So use a large language model to generate all the needed variants.
- These are really quite low risk because they are used at design time. They just help us do our jobs better and faster.
- Using Generative AI at run time is also possible. For instance to allow flexibility in entity collection. That is to allow the user to give you any or none or all pieces of information in whatever way they choose.
- This used to require a lot of logic in code but now it can be a single statement.
- Generative AI can handle procedural elements of conversation. Could you hang on a second? Could you repeat that? You can build this one-off, but with the right model, you don't have to.
- Generative AI can also provide a real-time complete FAQ from all the information you want you to draw answers from. With this, you can actually answer questions.
- Very nonlinear, totally unscripted conversations are only possible with a fully autonomous AI agent.
- Instead of writing code, just lay out the rules of the road and hand over the entire interactions: what gets said, the order in which it gets said, everything.
- This is necessary for some use cases that don't seem all that complicated like shopping for a new TV.
- You need an agent because these conversations are unscripted, nonlinear, and the user changes their minds: they back up, they go forward, they skip steps.
- Prompt engineering needs to tell the bot, who are you? What is your role in the conversation that you're about to have with an end user? etc.
- Evaluate anything that's generated. There aren't established guardrails yet but things like RAG and large context windows can help.
- Know when Generative AI is not your right solution like hen you need security, privacy, and compliance.
- Be aware of latency. Maybe you get away with it taking 3 or 5 seconds to get back to you. But even that isn't ideal for natural conversations.
- When large language models fail, they fail in ways that do not make intuitive sense to us as humans. Because LLMs don't really understand the world.<./li>
- There's a difference between knowing the association between words and images and having that real-world grounding of what they mean.
- This has implications on how we set up our users to talk to these things. Are we setting up these AI agents as artificial humans?
- The problem with putting a counterfeit version of a human in front of people is that we can't stop imagining the mind behind the conversation.