.jpg)
In just a few years, AI interfaces have come a long way, transitioning from the cryptic command lines of the past to the intuitive, visual interfaces we see today. This evolution hasn’t just been about making AI more accessible; it’s also been crucial in enhancing the usability, explainability, and effectiveness of AI applications across various industries. This, in turn, came with mainstream adoption and led to some of the most substantial business value creation since the creation of the smartphone. Understanding this journey is key to appreciating where we are now, where AI interfaces are headed in the future - and why we must ditch chatbot-like interfaces.
Interaction with AI used to take place in code and was strictly the domain of technical specialists. The technology had existed for decades in various incarnations, but - until very recently - it had been inaccessible to the average person or obscured and unrecognisable as AI (for example, in the ranking of search results on streaming platforms).
With the likes of ChatGPT and Midjourney, this changed to a degree. You no longer had to be a trained machine learning engineer to use AI. Still, to this day, specialist knowledge on how to compose a prompt has been vital for the most predictable results, initially prompting a wave of a new type of job profile being hired for: the prompt engineer.
In the early days of Google, when power users could employ special formulations—like using quotation marks for exact phrases or the minus sign to exclude terms—to control search results more precisely than others. This approach was very effective but only if you knew the specific commands, similar to how today’s users interact with tools like Midjourney.
In Midjourney, one of the most popular image-generation tools, users can control some aspects of the result by adding special suffixes or numerical inputs, echoing the early command-line approach in a more modern context.
Another significant limitation of early AI models was their small context and token limits. This meant that users could only input and receive relatively small pieces of information at a time. While this made it easier to iterate on shorter outputs, it also limited the complexity of tasks that AI could handle. Users often had to break down larger tasks into smaller segments, which made the process more cumbersome and less efficient.
As AI models evolved and these limits expanded, the ability to handle larger inputs and outputs increased dramatically. However, this also introduced new challenges—particularly the need for more sophisticated ways to manage and refine the outputs. This is where the evolution of AI interfaces becomes particularly significant.
As AI technology progressed, so too did the ways we interacted with it. The development of natural language processing (NLP) allowed AI to understand and respond to human language, paving the way for conversational interfaces. Chatbots and voice assistants were among the first widely adopted applications of this technology, marking a significant shift from the text-based commands of earlier systems.
In the mid-2010s, traditional decision-tree-based chatbots became particularly popular. These chatbots operated on predefined scripts, with hard-coded terms triggering specific, pre-written responses. They were often used as an additional meants of navigating content or triage customer service requests. However, they were also notorious for being exceedingly unhelpful in most situations. These bots could only handle limited interactions, often frustrating users when they couldn’t go beyond their programmed responses. The rise of these chatbots highlighted the need for more sophisticated, adaptable AI that could genuinely understand and respond to user needs.
Voice assistants like Apple’s Siri and Amazon’s Alexa took this a step further by enabling voice-based interactions. These interfaces made AI more accessible to the general public, allowing users to perform tasks like setting reminders, searching the web, or controlling smart home devices through simple voice commands. But then again, I yet have to meet one person who has an entirely cordial relationship with their smart assistant.
While the release of ChatGPT and similar conversational AI tools increased accessibility of the underlying technology, Large Language Models, they are still failing on numerous fronts: In a brief study on the usability of ChatGPT, the Nielsen Norman Group identified several heuristics which are the direct result of a linear, chat-based interaction, and which are in dire need of addressing for better usability for a broader set of tasks.
While conversational AI made interacting with technology more intuitive, it is the shift to visual interfaces that truly democratises AI for a broader audience. Visual interfaces allow users to interact with AI systems through graphical elements such as icons, buttons, and data visualisations. This transition is crucial for reducing the cognitive load (the amount of information a user has to hold in their heads to complete a task) and making complex AI systems more approachable. To create such a system, very likely requires refocusing it around specific tasks and bidding farewell to a generic tool.
As AI becomes more integrated into various industries, it’s clear that a one-size-fits-all approach to interface design doesn’t work. Each industry has unique needs, challenges, and regulatory requirements to consider when developing AI interfaces. Tailoring these interfaces to specific verticals is crucial for ensuring they deliver maximum value and usability.
During a conversation with a leader at a well-known AI company, we discussed the challenge of outperforming generic AI competitors. We concluded that it would be far more effective to create capabilities tailored to specific verticals like finance or the legal industry. This not only provided a competitive edge but also streamlined product development. By focusing on specific user needs within a well-defined context, it became much easier to determine what to do next and how to define success.
As I wrote, only vertical specialisation allows for creating of an end-to-end solution to a valuable problem. Even big players like OpenAI aren’t immune to this, which I suspect is why they partnered with Microsoft to create the Microsoft Co-Pilot, which adapts to whichever application you are presently using.
One of the significant limitations of many current AI interfaces, particularly text-based ones, is the lack of built-in mechanisms for iteration. Typically, when you interact with a text-based AI model, you provide a prompt, receive a result, and then decide whether that result meets your needs. If it doesn’t, the usual course of action is to rewrite the prompt and submit it again, hoping for a better outcome. This process is often inefficient and can lead to frustration, especially when dealing with complex tasks, or in situations where exact requirements are to be met.
In contrast, the ability to iterate on specific parts of an AI-generated output is crucial for refining and improving results. This is particularly important in fields like image generation, data analysis, code generation, and content creation, where the first output is rarely perfect.
Consider the example of AI-powered image generation tools, such as those used in creative design. An artist or designer might use an AI model to generate a concept image. However, the initial output might not fully meet their vision. Instead of starting from scratch with a new prompt, an effective AI interface should allow the user to make targeted adjustments. For instance, the user could select a specific part of the image—say, the background—and refine its color, texture, or composition without altering the rest of the image. This iterative process allows users to progressively build on the AI’s output, leading to a final result that closely aligns with their creative vision. Tools like Adobe’s Photoshop with AI-powered features and platforms like Runway ML are starting to incorporate these capabilities, enabling more granular control over AI-generated content. As we move further into not just image but video generation, costs and wait times become greater, so a way of working that doesn’t lead to 90% or more waste is crucial.
Iteration is equally critical in data analysis. When an AI model analyses a dataset and flags certain patterns or anomalies, the ability to iteratively refine the analysis can dramatically improve outcomes. For instance, a financial analyst might want to drill down into a specific anomaly detected by the AI, exploring it in different contexts or comparing it across multiple datasets. A well-designed AI interface would allow the analyst to adjust parameters or focus on specific data segments without having to start the analysis from scratch. This iterative process not only saves time but also enhances the accuracy and depth of the insights generated. The iteration process could also be captured, offering better explainability of the result as a by-product.
In code generation, iteration enables developers to refine the outputs of AI tools like GitHub Copilot. For example, after generating a block of code, a developer might need to tweak a specific function or optimise a certain section for performance. Instead of rewriting the entire prompt, the developer should be able to focus on iterating just that portion of the code, building on what the AI has already generated. This approach streamlines the development process and helps ensure that the final code is both functional and efficient.
Similarly, in text generation, whether for marketing copy, reports, or other content, the ability to iterate on specific sections is crucial. A writer might be satisfied with the overall structure generated by the AI but may need to adjust specific phrases for tone or clarity. A robust AI interface would allow the writer to focus on iterating those parts without affecting the rest of the text, enabling a more refined and polished final product.
Explainability is one of the most critical aspects of AI, particularly as these systems are increasingly used to make important decisions in fields like finance, healthcare, insurance, aerospace, and even defence. Visual interfaces play a vital role in enhancing the explainability of AI systems by providing users with a clear understanding of how and why certain decisions are made.
Consider a scenario in the aerospace industry, where engineers are responsible for monitoring and analysing signal data from satellites. An AI system might flag certain signals as abnormal, but without context, it’s difficult for engineers to understand the significance of these alerts. A visual interface that presents the flagged signal within the context of a timeline graph and potential causal relationships allows engineers to see the abnormality in relation to normal signal patterns. This contextual information is crucial for diagnosing the issue and determining the appropriate course of action.
Visual interfaces can also help bridge the gap between AI and its users by making complex processes more transparent. For example, in machine learning, models often operate as “black boxes,” where it’s unclear how they arrive at certain conclusions. By using visual tools like decision trees, heatmaps, or feature importance charts, developers and users can gain insight into the decision-making process of the AI, leading to greater trust and more informed decision-making. A good example of this is Ground News, which presents readers of a news article with a breakdown of its political leaning relative to other publications.
The evolution of AI interfaces from text-based command lines to sophisticated visual UIs has been transformative, making AI more accessible and effective across various industries. However, to truly unlock AI’s potential, it’s essential to tailor these interfaces to the specific needs of each vertical, ensuring they are not only intuitive but also explainable. Visual interfaces, in particular, play a crucial role in enhancing explainability by providing context and making complex processes more transparent.
Companies that prioritise user-centred design in their AI interfaces will lead the way in innovation and adoption, setting the standard for how AI is integrated into our daily lives. By focusing on their industry’s unique needs and continuously adapting their interfaces to meet these needs, companies can ensure they remain competitive in an increasingly AI-driven world.
Ready to take your AI product to the next level?
Contact us for a commitment free chat about how you can move beyond the prompt for a more successful AI product.
