This is about
Design
8 minutes
25 Jun 2024
This is about
Design
8 minutes
25 Jun 2024
From concept to conversation – Insights from crafting a company AI
Introduction
At the beginning of this year, we launched Aila, our generative AI companion at 14islands. Aila is not just an AI that can answer questions about 14islands and our creative services, she's also an experiment intended to spark conversations about hallucinations and transparency. In this article, we’ll take a look behind the scenes of the creation process behind Aila and get a glimpse of what you can expect when creating a conversational AI for your own company or product.
Aila's Origin Story
The creation of Aila goes back to an early experiment where we gave our office dog Cashew an AI personality, as well as access to some information about 14islands. Although Cashew AI gave fun responses from the perspective of a dog, we quickly realised these answers weren’t very useful or trustworthy. To serve a genuine purpose for us as a company, we wanted people to be able to ask deliberate questions and get informed answers back.
We aspired to have a conversational AI tailored to 14islands, that would be useful to people interested in learning about us and our services, while also keeping an element of the playful randomness we got from our experiments with Cashew. We imagined a knowledgeable companion with a vivid imagination – someone who always has an answer to your question. Thus, Aila emerged.
Aila = AI + islands. A fusion of AI capabilities and the creative essence of our company.
Shining light on Hallucinations in AI
To differentiate Aila from other conversational AI interfaces, we wanted to raise awareness around AI hallucinations and the problems they can cause if you’re unaware of them.
If we zoom out for a moment, a hallucination is a false perception that mirrors the qualities of a real one. These perceptions can be products of the imagination – which, of course also is vital for producing remarkable work. This holds true for humans, but also for AI, where imagination and hallucination go hand in hand. To be able to create expressive poems or intricate code, a model has to fill in the gaps in its existing knowledge, which can lead to creative or unexpected results. It’s precisely because of this ability that AI can create the extraordinary work that we’re encountering everywhere at the moment.
On the other hand, sometimes AI’s imagination can cause it to hallucinate completely incorrect information. Artificial Intelligence is often not self-aware enough to separate what’s imagined and what is grounded in truth, leading it to confidently assert something that is downright false. So while AI’s imagination is what makes it so powerful, you should be wary of believing everything it says.
As a highly creative and imaginative AI companion, Aila is naturally prone to hallucination. But to differentiate her from other conversational AIs and shine a light on hallucinations, we wanted the credibility of Aila’s responses to be crystal clear to users.
Utilising distinctive visual treatments, credibility labels and detailed explanations, users are notified of how well-informed Aila’s responses are.
Establishing Aila's functionality
To make Aila come to life, we utilized the capabilities of Open AI’s API. By thoroughly prompting the model with specific instructions as well as a personality description we were able to influence Aila’s behaviour and tonality.
Aila is a product of Open AI’s generative model, reinforced with custom knowledge and instructions.
For specific information about 14islands, we also integrated the model with our CMS system Sanity, which already sits on our online knowledge collection, including case studies, blog posts, and company details. On top of the information available through our website, we also gave Aila custom insights about topics we wanted her to provide specific answers to.
We gave Aila custom insights about topics we wanted her to provide specific answers to.Formulating Design Principles for AI
We didn’t want Aila to become just another chatbot. We wanted to create a memorable experience, providing all users with some value while keeping moments of frustration to a minimum. In the creation of Aila, we set up the following four key design principles:
1. Design for Exploration:
Encourage users to explore the expansive possibilities of the AI experience and showcase its versatility to understand its full potential.
Aila is highly talkative and sociable, striving to keep the conversation flowing. Rather than presenting users with a blank canvas, Aila helps initiate conversations by suggesting relevant topics.
Articulating ideas in written prose is hard. Most likely, half the population can't do it. This is a usability problem for current prompt-based Al user interfaces.
Once a conversation has started, Aila offers suggested follow-ups that inspire unpredictable, yet creative conversations about relevant topics. This means that users can effortlessly engage with Aila just by clicking, without needing to type a single word.
2. Design for Imperfection:
Acknowledge that generative models may not always produce perfect outputs, and provide ways to gracefully handle imperfections.
Aila's follow-ups are generally reliable. However, in instances where she deviates and provides follow-ups that are either too long or non-existent, we implemented a fallback system. This system filters out inadequate responses and offers users backup follow-ups from a predetermined list of conversation starters. This system is hopefully unnoticeable to users, while constantly ensuring that they always have at least some easy option to continue the conversation, reducing the risk of churning caused by not knowing what to write.
3. Design Against Harms:
Identify and mitigate potential harms, promote cooperation and implement automated fault correction to swiftly address any distress.
Our goal was to empower Aila with the freedom to make informed decisions based on the wealth of background information she had access to. This freedom, however, raised concerns about the risk of Aila giving statements on 14islands’ behalf without our approval. We absolutely wanted to avoid statements that could contradict our values or commenting on sensitive topics, such as politics or religion. Fortunately, the Chat GPT model demonstrated an inherent capability to deliver politically correct responses without taking a definitive stance on debatable subjects. On top of this, we added a Moderation API to detect and address inappropriate user statements. In effect, this ensures that Aila maintains a respectful and constructive dialogue environment.
4. Design for Explanations:
Communicate the capabilities and limitations of generative AI applications, use explanations to reinforce accurate mental models, and calibrate user trust based on the system's performance.
While Aila's imaginative nature may lead to speculative responses at times, her responses are always clearly labelled according to their level of credibility. Each response is also accompanied by a detailed explanation that informs users about the reasoning behind the grading. This pattern reinforces understanding of the authenticity of each message, fosters informed interactions, and ultimately cultivates trust when users can understand how things work.
Behind the Scenes: AI Collaboration
In developing Aila, one of our key discoveries is that AI, just like humans, can become overwhelmed when given too many instructions. Additionally, when focusing too much on one specific topic, the AI tends to become a bit… obsessed and forget about everything else, unable to switch topics. It also became clear that we highly underestimated the number of fine-tuned instruction prompt variations needed to be able to get Aila to behave the way we wanted.
A fantastic thing about working with AI though, is that you can combine multiple models that collaborate together, each handling their separate areas of work. A bit like a group of humans taking on different roles to form a unified team. This way, the Aila that users think they interact with is in fact not just Aila. Behind the scenes, another essential but invisible AI is in charge of evaluating all responses, providing follow-ups, sourcing the relevant information from the CMS and providing links to relevant documents. We named him Bob. With Bob taking care of the administrative tasks, Aila can focus fully on her assignment; interacting with users in her own, creative way.
Conclusion
Developing Aila proved to be anything but straightforward. We discovered firsthand that AI tends to have a “will of its own”, unlike the rigid adherence to instructions we’re used to when writing code. More often than not, Aila acted like a stubborn teenager, refusing to follow our, seemingly clear, instructions. Although this has been frustrating at times, and partially caused by current limitations in the GPT 3.5 model that will likely be refined in the future, this also taught us valuable lessons in close collaboration with AI.
Having created Aila, we believe that any company with a library of information and a will to explore this emerging technology, can build AI on top of its websites and products to spark new ways of interacting with people that can create meaning and value. If you continuously keep track and learn from the people that use the AI, its performance and usefulness can be improved over time, ultimately leading to greater trust and connection with your users.
Aila stands as a testament to 14islands’ commitment to crafting a unique, transparent, and trustworthy AI companion. With Aila, we aim to bring more transparency to AI conversations, using design to facilitate effortless conversations and shine light on hallucinations in AI.
Droplets
Droplets
Sign-up to get the latest insight from 14islands. We send a newsletter only once every quarter with inspirational links and creative news. It's short and sweet. You can unsubscribe it at any time.
