Sebastian Woźniak — Nucleus AI Platform Designer at UBS
Wearing blinders
September 4, 2023
15:45
Future of design
🇵🇱 Polish
Future of design
Human-Centered AI: The Role of Design in Building Better AI Products
About
Implementing an AI-based solution is primarily a technological challenge. Meanwhile, concepts for new products or functionalities are often born in the headsof designers. Is it possible to provide designers with a workshop that allows them totake a realistic look at the possibilities of Machine Learning? Or should we instead provide developers and architects with a design workshop, so that they can be theones to look for applications of AI and recognize the risks of its use? Let’s discuss Human-Centered AI and the role of collaboration between designers anddevelopers in building the next wave of better Artificial Intelligence products.
Watch the full talk
Watch this WaysConf session, then continue with related talks or explore the current programme.
From the recording
Talk in brief
In this WaysConf 2023 presentation, Mateusz Cygan explores the role of product design in developing human-centered artificial intelligence applications. Cygan explains the fundamental difference between traditional deterministic programming and machine learning, emphasizing that AI systems discover data regularities rather than following explicit step-by-step instructions. To avoid over-engineering, product teams must evaluate whether a problem requires predictive machine learning or simple computational logic. Using metaphors and industry examples, Cygan demonstrates how designers act as an essential bridge between engineering capabilities and human needs. Key design responsibilities include establishing explainability in interfaces, anticipating data bias, building user trust, and augmenting human capabilities instead of replacing them completely. During the Q&A session, Cygan discusses practical methodologies for AI design, feedback loops, and fine-tuning specialized domain models to prevent performance degradation.
Key takeaways
- 01
Distinguish Direct Rules from Learned Patterns
Products should only implement machine learning when deterministic rules or explicit calculations cannot solve the problem. Direct logic works best for precise calculations, whereas machine learning excels at complex pattern recognition like computer vision.
Watch from 18:15 - 02
Prioritize Human-Centered Value over Technical Capability
Product development should begin with user preferences and constraints rather than raw model capabilities. Combining technical data science with human-centered design prevents over-engineered solutions that fail to address real user needs.
Watch from 16:04 - 03
Design for Explainability to Maintain User Trust
AI interfaces must present contextual data alongside recommendations so users understand how outputs were generated. Providing alternatives and contributing factors allows users to remain in control rather than relying blindly on opaque system predictions.
Watch from 22:22 - 04
Augment Human Workflows Instead of Automating Them Entirely
Product teams should apply AI to eliminate tedious, repetitive tasks while keeping humans in charge of creative decisions. Designing collaborative human-AI setups preserves user agency and improves overall task performance.
Watch from 26:00 - 05
Fine-Tune Specialized Models for Safe Enterprise Deployment
Rather than relying on open, unconstrained general models that risk user-induced degradation, organizations should fine-tune pre-trained models on curated domain data. Constraining AI outputs to specific business tasks protects performance and reduces hallucination risks.
Watch from 43:47
Read edited transcript highlights
These concise notes were edited from automatic captions and checked against the talk structure. They are not a verbatim transcript.
Contrasting Traditional Code with Machine Learning
In standard software development, engineers create explicit recipes containing conditional logic to instruct computers how to process data step by step. This approach powers deterministic features like liking a post on social media. In contrast, complex tasks such as music recommendation or movement control rely on machine learning. Instead of executing direct human instructions, machine learning models analyze massive datasets to discover underlying numerical regularities and relationships. The model generates an algorithm independently, predicting accurate outputs based on given inputs even if creators cannot fully articulate every internal decision factor.
Watch from 3:36The Party Hosting Metaphor for Product Design
Human-centered AI prioritizes end-user value over technical novelty. Building an AI feature can be compared to throwing a party. A data scientist hosting a party often looks through the available ingredients in the refrigerator and tries to cook the most complex dish possible using everything on hand. Introducing a product designer transforms this dynamic by focusing first on the guests' dietary restrictions and personal preferences. The ideal development process pairs data engineering capability with editorial design judgment, ensuring that advanced algorithms directly serve human needs rather than showcasing raw technological power.
Watch from 16:04Filtering Problem Scenarios for AI Application
Product teams frequently request artificial intelligence without evaluating whether simpler methods exist. To decide if AI is necessary, teams should ask whether a direct algorithmic solution already exists. For example, selecting the cheapest grocery store for a shopping list relies on explicit calculation rather than predictive machine learning. Conversely, distinguishing a cat from a dog in a photo is trivial for humans but nearly impossible to define using rigid conditional code. Tasks lacking explicit formulas that possess ample training data represent ideal candidates for machine learning solutions.
Watch from 18:15Enhancing AI Explainability Through Interface Context
Because machine learning algorithms often operate as opaque systems, designers must introduce explainability into user interfaces. Interfaces should not merely display a single prediction; they must supply background context explaining how that answer was formed. Google Maps demonstrates this by recommending a primary route while simultaneously displaying alternative paths, estimated travel durations, and predicted traffic congestion points. Providing transparent contextual data gives users agency over machine predictions, building appropriate levels of trust and preventing blind reliance on potentially flawed system outputs.
Watch from 22:22Translating Engineering Utility into Human Behavior
Design serves as the critical communication link that transforms technical utility into meaningful human behavior. In large language model development, techniques like Reinforcement Learning from Human Feedback (RLHF) represent a form of design by aligning raw pre-trained data with human expectations. Incorporating slight human supervision dramatically improves usability, making interfaces feel intuitive, responsive, and aligned with user intent. Product teams must establish visual tools, feedback loops, and structured workflows that allow designers to safely shape AI capabilities into helpful human experiences.
Watch from 30:45Managing Feedback Loops and Domain Fine-Tuning
Deploying fully open AI models that learn continuously from unfiltered public interaction carries significant risk, as historical experiments have shown models absorbing harmful behaviors. To mitigate this, organizations should deploy specialized models fine-tuned on internal datasets rather than unconstrained general systems. By grounding core pre-trained models with domain-specific knowledge and explicit guardrails, companies can deliver focused capabilities—such as dedicated customer support bots—without exposing system performance to public manipulation or hallucinated errors.
Watch from 43:09
Explore WaysConf 2026
