2023 Ways Conference Talk
On the 3rd of September, I had the great pleasure of speaking at WaysConference 2023. In my presentation titled "Human-Centered AI - The Role of Design in Building Better AI Products", I shared a designer's perspective on how artificial intelligence systems are different from traditional programmed products and why design will play a super important role in building AI products in the future.
Thank you to all the attendees and Ways Conference organizers. I'm proud that such a great conference takes place in our Krakow.
AI Buzzwords Dictionary
Please note that the definitions below aim to build a general understanding of technology for designers and product managers based on my experiences. Treat them as an inspiration and entry point for further research.
#1 AI Summer
AI has a long history, as long as the history of computers and the digital revolution in general. Periods when AI popularity grows are called AI Summers, and often in history, were followed by AI Winters.
It's good to remember that in the past, expectations regarding AI development were overhyped.
Great articles and timelines about AI History: https://www.historyofdatascience.com/xperiences/
#2 Machine learning
The way in which computers 'discover' their 'own' algorithms without needing to be explicitly told what to do by any human-developed algorithms. It's a base technology that makes AI-based products different from those that we can call "traditional".
#3 Neural Networks
They are a subset of machine learning and types of algorithms that are similar in structure to the human brain and nervous system. It's worth remembering that neurons in the human brain are incredibly complex, and how they work hasn't been fully discovered yet. Consequently, this similarity works only on a very general level. Although the way in which neural networks "learn" or are "trained" often has a lot in common with how we humans learn.
#4 Deep learning
Deep learning is a subset of machine learning where artificial neural networks learn from large amounts of very complex data. In other words, deep learning is a name for complex neural networks with many levels that can analyze complex, even unstructured datasets. The discipline has progressed quickly in recent years thanks to increased computing power. Most of today's AI products or features are based on deep learning.
#5 Natural Language Processing (NLP)
A discipline that focuses on enabling computers to understand, interpret and manipulate human language, mostly in the form of text. Today the main difference between how we interact with computers and other humans is related to language. We build a simpified way of communication with computers (mostly GUI), but our firs-choice form of communication with other humans is based on natural language.
#6 Large Language Models (LLMs)
Huge deep learning models trained on huge base of texts (corpus). Instead of being trained for any specific purpose (narrow AI), they are built to understand humanity's knowledge and language in general. Chatbots that use those models can generate text that is difficult to distinguish from text written by humans. The most popular LLM is GPT from OpenAI, but there are many others (worth checking!) with various scales and levels of openness. The potential to use them in products is huge, but it's not fully discovered yet.
#7 Generative AI
AI models that can generate content that we could call "creative" from a human perspective. Those can be text (articles, poems, lyrics, e-mails) but also images (like from DALL-E or Midjourney) or songs, movies, sounds, etc. The availability and development of generative AI are responsible for today's boom in the popularity of AI.
#8 AI as a Black Box
Some complex deep learning models can generate impressive outputs, but their structure doesn't allow us to fully discover what factors and patterns they discovered in reference data to produce it. This inability to understand how an AI system makes decisions or predictions can be problematic when it comes to issues such as bias or ethical concerns.
#9 Explainability
It refers to designing and improving AI models to provide output, allowing humans to understand what factors influenced it and control how it will be used. It's a creative challenge to figure out what data should model output next to the main answer and how to present that to the user so he will have the complete picture and a good understanding of AI's finding process.
#10 AI Bias
The tendency of an AI system to make decisions or predictions that reflect the biases of its creators or the data it was trained on. As AI models are trained on very complex data sets, it requires good foreseeing and creativity to determine what potential biases can affect the model's output. Also, it's a significant challenge to build good mechanisms to prevent and report them on various levels - from data labelling to UX.
#11 Augmentation instead of Automation
It's an approach of using AI to augment human capabilities rather than replace them completely. Whenever we are designing a product that aims to automate some process, we should deeply study what parts of this process require human control and vision or are simply satisfying for humans. Instead of automating the whole process, we can use AI to automate only those boring or repetitive parts and keep humans in control of the new AI-boosted process.
#12 Human+AI Teams
Teams that combine human intelligence with machine intelligence to achieve better outcomes than either could achieve alone. In this approach we should treat AI as a servant team member which can support humans in some tasks.
#13 AI Hallucinations
Those are situations where an AI system generates output that is completely untrue and incorrect and presents it very convincingly. AI systems output predictions instead of direct calculations, so we can never fully trust them. From a design perspective, we must build good communication in our products to provide users with a good understanding and manage their expectations.
#14 AI Feedback
AI Feedback refers to using feedback loops to improve the performance of an AI system over time and let users control and understand the product. The nature of AI products allows us to figure out some ways of looping feedback into the model without user interaction (for example, we can check if the user is listening to the song system recommended him to check if the recommendation was valuable). Still, we should often build a way for users to control and customize their expectations for AI models and report unwanted outputs. It's a vital role of UX design to create new patterns and consistent ways for users to communicate with AI.
Slides to download
Below, you can download a PDF with the most important slides from my presentation. Please provide attribution and source, and let me know if you are using it.
Recommendations and sources
- The People + AI Guidebook by Google - https://pair.withgoogle.com/guidebook/
- Book: Big Data, Big Design: Why Designers Should Care about Artificial Intelligence, by Helen Armstrong - https://helenarmstrong.info/
- DARPA Robots Falling Video - https://spectrum.ieee.org/darpa-robotics-challenge-robots-falling
- Boston Dynamics Robots Parkour Video - https://www.youtube.com/@BostonDynamics
- AI Hallucinations article and example - https://blog.finxter.com/hallucinations-in-ai-with-chatgpt-examples/
- Neri Oxman - Krebs Cycle of Creativity - https://spectrum.mit.edu/winter-2017/neri-oxmans-krebs-cycle-of-creativity/
- Lex Fridman Podcast with Sam Altman - https://lexfridman.com/sam-altman/
- Illustrations by super-talented Zuzanna Łomzik - https://www.behance.net/zuzannalomzik/appreciated
