Introduction
As we stand on the brink of a technological revolution, AI is reshaping the way we interact with the digital world. I'm here to share my insights regarding this evolving landscape, offering observations into how AI, machine learning, and other related technologies can enhance our work in UX design.
Brief Overview of AI and ML
Artificial Intelligence (AI) is the overarching concept of machines carrying out tasks in ways that we consider 'smart'. It's a broad discipline aimed at creating systems that can simulate various aspects of human intelligence. Machine Learning (ML) is a subset of AI, focused on algorithms that enable machines to improve at tasks with experience. Think of AI as the universe of intelligent computation, with ML being a planet within it.
Deep Learning, in turn, is a subset of ML. It involves neural networks with many layers (hence 'deep') that can learn and make intelligent decisions on their own. Deep Learning has been instrumental in achieving significant breakthroughs in areas like image and speech recognition.
Then come Foundation Models, a recent development in AI. These are large-scale models (like GPT-3) that are trained on vast amounts of data and can be adapted to a wide range of tasks without being specifically trained for them. They are called 'foundations' because they provide a base layer of understanding that can be built upon for various applications.
These technologies are interrelated, forming a hierarchy from broad to specific. AI encompasses everything intelligent that a machine might do. Under AI, ML is the method through which machines learn from data. Deep Learning is a further specialization of ML with a focus on complex, layered neural networks. And Foundation Models are the cutting-edge, versatile systems pushing the boundaries of what AI can achieve.
Generative Pretrained Transformer (GPT) and Transformer Architecture
GPT, or Generative Pretrained Transformer, is a type of language processing AI. It's built on the Transformer architecture, which is revolutionary in the way it handles data. Traditional models processed data sequentially, one piece after another. The Transformer, however, allows for parallel processing, looking at entire sequences of data at once. This is done through mechanisms called 'attention' and 'self-attention,' letting the model weigh the importance of different parts of the input data and learn the context more effectively.
GPT takes this architecture and applies it to language, learning to predict the next word in a sentence. It's trained on a vast corpus of text and then fine-tuned for specific tasks. This pretraining is what makes it 'generative' - it can generate text, not just understand or classify it.
Inception of Generative AI
Generative AI came into existence as a natural progression from earlier AI models that were primarily discriminative. While discriminative models could classify and understand data, they couldn't create new data. The inception of Generative AI marked a shift from understanding to creation.
This shift was fueled by advancements in neural networks and an exponential increase in computational power and data availability. Researchers began exploring how neural networks could not only recognize patterns but also use those patterns to generate new, similar data. This exploration led to the development of models like Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), and eventually to large-scale models like GPT.
Integrating Reinforcement Learning with Human Feedback
Adding to this is the concept of Reinforcement Learning with Human Feedback (RLHF). This approach trains AI systems not just on static datasets but on dynamic feedback loops. It’s akin to teaching a pet new tricks: the AI tries different strategies and, based on human feedback, learns which actions are desirable or undesirable. This human-in-the-loop methodology ensures that the AI’s learning trajectory aligns more closely with human values and preferences, making it particularly potent for personalization and adaptive learning scenarios.
Incorporating RLHF into the design of AI systems compels us to consider not just the initial user interaction but the ongoing relationship between the user and the AI as they adapt and learn from each other over time.
Various Approaches to AI
Understanding the various approaches to AI can be daunting, but it's crucial for harnessing its full potential:
- Symbolic AI: Based on the manipulation of symbols and rules, this approach tries to mimic human reasoning.
- Probabilistic AI: Focuses on using probabilities to make predictions and decisions.
- Statistical AI: Uses statistical methods to infer patterns and make predictions.
- Large Language Models (LLMs): These are trained on vast datasets to understand and generate human-like text.
Each approach has its strengths, and knowing which to apply can significantly impact the success of a project.
How Are Models Trained?
AI models are trained using large datasets. They learn by recognising patterns and making associations. For instance, an image recognition model might learn to identify cats by being shown thousands of pictures of cats and not-cats. Over time, it improves its accuracy through a process called backpropagation, where it adjusts its internal parameters to minimize errors.
AI and Design
In the realm of AI, as designers, we're tasked with shaping a human-machine relationship that is continually evolving. Unlike traditional design, where interactions are fixed, AI-based design involves a fluid, ever-changing dynamic. Non-AI systems interact in a consistent, unchanging manner. However, AI systems learn and adapt over time, leading to a constantly developing relationship.
As both the machine and humans learn and adapt, they engage in a reciprocal learning process, forming a dynamic feedback loop. This loop is characterized by an ongoing exchange of information, with both parties growing and adjusting to the interaction. Our role as AI designers is to foster effective communication within this vibrant human-machine relationship.
In this context, data becomes the pivotal element. Previously, data interaction was straightforward — a user command followed by a machine response. With AI, data is the foundation of learned behaviours, informing the machine and shaping its growth. Here, data inputs are about educating the machine with information, not just instructing it. Conversely, machine outputs are no longer fixed; they have generated responses that necessitate explanations of how the data led to a particular conclusion.
As designers, we are charged with orchestrating these interactions throughout the entire data lifecycle. This includes the initial data capture, guiding the machine's learning process, designing clear and understandable data outputs, and ensuring a seamless flow of information back into the system for continued learning. Our responsibility extends beyond creating interfaces; it's about crafting experiences that support this rich, ongoing dialogue between humans and AI, ensuring clarity, transparency, and a mutual growth trajectory.
A comparison can be drawn from the days when selecting a movie at a DVD store involved browsing aisles for a favoured title or relying on a recommendation from a salesperson. In contrast, platforms like Netflix now curate suggestions tailored to our tastes, informed by our viewing history.
Designers find themselves at the intersection of innovation, with three distinct yet interconnected realms to explore:
1. Designing with AI: Crafting Alongside AI
Designers are now partnering with AI in a collaborative dance of creation, where the output is a fusion of human ingenuity and machine efficiency. Imagine the synergy of a designer working with AI to conjure up ground-breaking architectural structures, much like Autodesk's venture in utilizing generative design principles to conceive their Toronto office. Here, AI becomes an ally in the creative process, providing new perspectives and solutions that push the boundaries of traditional design.
Moreover, designers harness AI to streamline repetitive tasks, freeing up creative energy for more complex challenges. Tools like Airbnb's system for transforming sketches into digital wireframes or Netflix's algorithm for adapting graphics across different cultures exemplify this trend. These innovations signify a new era where AI does not replace the designer but rather amplifies their capabilities.
2. Designing for AI: The Human-Centric Design Approach
Designing for AI requires a human-centred lens, focusing on crafting systems that prioritize user needs and experiences. It's about spotting those unique opportunities where AI can not only function but flourish in addressing real-world problems. Here, the designer's role transcends aesthetics, venturing into the realm of functionality and utility, transforming user needs into data-driven AI solutions.
Human-centered design makes AI effective. These are some ways:
1. Beyond Interfaces: Embracing Human-Centered Design in AI Algorithm Development
UX can aid in designing algorithms that mirror the decision-making processes humans employ, by considering their information, goals, and constraints. It can ensure that the decision environment, which encompasses both the algorithm and its human users should be thoughtfully constructed. Users should comprehend their AI tools well enough to use them effectively. Designers also aid in establishing guidelines and business protocols that translate algorithmic predictions into actionable insights, advising when human intervention is appropriate to supplement or override the AI.
2. To Translate User Needs into Data Requirements: A designer can aid in identifying the type of data necessary for training the model, considering various factors like predictive power (A percentage that refers to an ML model’s ability to predict outcomes given a certain input correctly), relevance, fairness, privacy, and security. Ensure the training dataset is comprehensive, reflecting the real-world scenarios the AI will encounter, and free from biases.
3. Knowing the source of the data & Tuning the Model: Evaluating data sourcing and collection methods for their suitability of the project is critical. Once deployed, A designer will assess if the AI meets the target user’s needs as per predefined success metrics. Provide feedback on adjusting the model’s parameters as needed to enhance its performance, focusing on metrics that reflect user experience, such as customer satisfaction or the frequency of users following the AI’s recommendations.
4. Addressing Bias, Fairness, and Transparency: UX designers can help analyze data with an understanding of the domain, Goal definition, possible and required outcomes and the process that generated it. This leads to designers being a crucial part of designing algorithms that are mindful of the environment they will operate in, avoiding controversial predictors. They aid in conducting usability tests or audits to detect and eliminate unintended biases.
5. Managing the Handoff in AI Systems: Designing for smooth transitions between AI and human control in situations demanding common sense or contextual understanding is very important. The accountability for any action in the real world still lies with human users as over-reliance on technology can leave users unprepared for instances where AI fails, necessitating more skilled human intervention. A UX designer can analyse such situations and will aid in designing smooth handoff processes.
6. Designing Reward Functions and leveraging User Feedback for Model Improvement: Designing the AI’s reward function is critical as it influences the user experience significantly. User feedback is essential in refining AI models and enhancing user experience. Designers analyse and optimize the reward data to enhance the model for long-term user benefits and anticipate the downstream effects of your product. This also allows users to contribute to the personalization of their experiences, thereby increasing their trust in the system.
7. Anticipating Errors and Designing Response Pathways: A Human-centred design prepares your AI system to facilitate user responses to inevitable errors, turning them into opportunities for learning and improvement.
8. Educating Users and Setting Realistic Expectations: Designers help communicate the capabilities and limitations of your AI product to customers. Help users develop accurate mental models and understand how their interactions train the system. It’s essential to balance user trust, avoiding both undue scepticism and over-reliance on AI.
9. Guiding User Trust in AI Systems: Users need to adjust their trust in AI systems appropriately, rather than relying on them implicitly in every situation. The phenomenon of 'algorithm aversion' is well-documented, where users may be sceptical of software systems. Conversely, there are instances where users place excessive trust in AI capabilities, expecting more than what the system can deliver. Designers help users develop a balanced level of trust, aligning with the system's actual capabilities and limitations by taking a human-centred approach.
For instance, openly acknowledging the potential inaccuracies in AI predictions can temporarily reduce trust in those specific outcomes. However, this honesty can foster a more sustainable trust in the long term. Users become more judicious in their reliance on the system, reducing the likelihood of disappointment due to misplaced expectations."
3. Designing of AI: The User Experience Frontier
When it comes to the design of AI, it's about envisioning and sculpting the interactions between AI systems and their human users. It's a space where new forms of engagement, like voice-activated assistants or image recognition software, become gateways to enhanced user experiences. The key challenge here is transparency: designing interfaces that not only serve but also educate. Users should be able to grasp, with just the right level of detail, how AI systems make decisions and learn over time.
An example could be the intuitive dashboards in our cars that provide real-time insights into the vehicle's AI, or the smart home devices that learn our preferences and conversationally explain their actions. These are no longer scenarios from a sci-fi novel; they are today's design challenges that call for a blend of technical knowledge, user empathy, and creative foresight.
In essence, the designer's canvas has expanded, not just in size but in dimensionality. As AI continues to intertwine with our daily lives, it invites designers to step into roles that are as diverse as they are dynamic, shaping not only how AI looks but also how it behaves and interacts in the fabric of human experience.
Understanding the Limitations of AI in UX Design
Incorporating AI into UX design comes with distinct limitations that underscore the irreplaceable value of human insight and direction.
1. Hallucinations and Reliability
AI "hallucinations" refer to instances where a model confidently generates an incorrect response. These can be caused by inconsistencies within a large data set or errors in the model's training methodology. In fields where precision is critical, such as financial reporting or legal documentation, these inaccuracies can introduce significant risks. Combatting this requires robust document structuring and advanced prompt design techniques to direct AI towards more dependable outcomes.
2. Prompt Sensitivity
Large Language Models (LLMs) are highly sensitive to user input. The nuances of how a prompt is phrased can lead to varied and unpredictable responses. This sensitivity necessitates a careful and strategic approach to prompt engineering, ensuring that the AI's responses align with user intentions. The evolving role of "Prompt Engineer" is a testament to the significance of crafting prompts that steer AI toward delivering consistent and accurate results.
3. Context Window
Limits The context window, the amount of information an AI can consider when generating a response, is a notable constraint. As the context window expands, so does the computational complexity. Despite improvements like GPT-4's extended context window, there remains a ceiling to the volume of data an AI can process at a time. This limitation is particularly challenging in tasks that require the review of extensive documents, where the AI must understand and analyze large quantities of text. Designing for AI in UX thus requires a thoughtful balance between the AI's capabilities and the complexity of the tasks it is expected to perform.
Things Designers Should Know Before Designing for AI
Technical Knowledge
- Grasping the basics of data science and AI techniques like NLP and deep learning.
- Basic Understanding of the AI toolchain and DevOps processes for AI development.
- Access to clean data set.
Ethics in AI
- Integrate ethical standards into the design and development process, ensuring accountability, fairness, and transparency.
- Navigate GDPR/compliance mandates and be mindful of the financial and human impacts of design decisions.
Collaboration is Key:
- Work closely with data scientists and engineers to understand the technical aspects and constraints.
- Foster a shared vision across multidisciplinary teams and participate actively in the AI development lifecycle.
- Engage in design thinking activities tailored to AI and address the needs of diverse users.
AI Strategy
- Develop and communicate a user-focused AI strategy, explaining barriers to adoption and the business and user benefits.
- Articulate your company's AI strategy, differentiators, and the journey toward AI success.
Designing AI Interactions
- Translate AI model outputs into understandable insights for users and design interactions that capture human input for machine learning.
- Prototype and test AI solutions frequently, and design for various stages of the AI lifecycle, such as model maintenance and data collection methods.
User Trust and Transparency
- Build trust by making your AI interactions transparent and understandable.
In conclusion, AI is not just a tool but a new frontier in design. By embracing it, we can create more personalized, efficient, and engaging user experiences. Let's embark on this journey together, continuously learning and adapting to ensure technology serves humanity in the most beneficial ways.
Resources to Refer :
- Generative AI at SAP: https://open.sap.com/courses/genai1
- AI Ethics at SAP: https://open.sap.com/courses/aie1-1
- Designing for Generative AI: https://experience.sap.com/internal/fiori-design-web/generative-ai-design/
- Designing for Intelligent Systems: https://experience.sap.com/internal/fiori-design-web/designing-intelligent-systems/
- SAP AI community file: https://www.figma.com/file/SJh2Eb5KrBIbaK3Bo2wqtq/AI-UX-Pattern-Community-Explorations?type=design&n...
- AI Design Community Teams Channel
- AI Design Community Exploration File
- Prompt Pattern Presentation Figjam
- SAP Managed Tags:
