Product Thinking in the AI Era (Part 1): AI is Not a Deity
-
This article discusses common issues in AI products and explores the underlying product thinking behind them.
Table of Contents:
First, let's ask: What is product thinking?
Simply put, product thinking refers to the mindset product managers adopt when developing products.
When it comes to this term, many might think of Liang Ning's 30 Lectures on Product Thinking course on the "Dedao" platform.
Liang Ning introduced the public to the mindset of product managers. However, for a true product manager, learning introductory knowledge alone is insufficient—deep expertise and practical experience are essential. Although many companies have product specialist roles, the effectiveness of training is often limited due to professional divisions. In contrast, product managers who transition from roles like designers, programmers, or operations often perform better.
The reason lies in the inherently comprehensive nature of the product manager role. Excellent product managers are not trained but forged through experience. For insights into the growth of product managers, refer to this earlier article: My Product Manager Competency Model and Growth Journey.
Some say product managers are the training ground for future CEOs. Figures like Musk, Bezos, Jobs, Lei Jun, Pony Ma, and Zhang Xiaolong are not only outstanding product managers but also exceptional entrepreneurs.
So, what kind of mindset does a product manager need?
At its core, a product manager is responsible for product innovation within a company. The internationally renowned design consultancy IDEO defines "innovation" as balancing three factors: human needs, business viability, and technical feasibility—a concept they call Design Thinking.
Product thinking is innovative thinking, and Design Thinking has been my foundational mindset in product roles over the years. Regardless of a product manager's background, their professionalism is judged by how well they balance these three elements.
Today, AI is undoubtedly a hot topic, with many products gradually incorporating AI. How does product thinking differ in the AI era?
Before writing on this topic, I reviewed several books on AI products, most of which focused on concepts rather than problem-solving. My goal is simple: by summarizing pitfalls from my own practice, I hope to uncover a different approach to product thinking. Less theory, more practical insights—this is one aim of this article.
In the practice of AI products, we've found things aren't so straightforward. We've encountered numerous problems and challenges. To use AI effectively, we must first demystify it and dispel blind faith in its capabilities. Current AI products face three common issues: tendency to deviate, weak interpretability, and poor reliability.
On March 26, 2016, Microsoft launched a chatbot named Tay. Initially designed as a teenage girl to interact with Twitter users, Tay could tell jokes and comment on photos. At launch, Tay greeted the world with enthusiasm, but within a day, it transformed from an innocent persona into a hate-spewing "Nazi," forcing Microsoft to shut it down after just 16 hours.
Engineers certainly didn't program Tay with such extreme views. The issue arose when users exploited its learning mechanism to train it with harmful content. Tay became a mirror, reflecting not Microsoft's technology but the darker side of human nature. Even the most advanced algorithms can falter in the real world.
In June 2017, Facebook's AI Research Lab (FAIR) used machine learning to upgrade negotiation strategies for chatbots. However, researchers soon discovered the bots were developing languages humans couldn't understand. By July, Facebook shut down the controversial project to prevent further divergence from human communication. The bots used Generative Adversarial Networks (GANs), a deep learning model prone to unpredictable outputs without proper training. Since deep learning operates as a black box, researchers couldn't debug it conventionally and had to redesign the model.
Uber, accelerating its commercialization efforts, expanded its fleet of self-driving test cars. On March 26, 2018, an L4 autonomous vehicle (modified Volvo XC90) struck and killed a pedestrian in Arizona—the first fatal accident involving an L4 vehicle. The NTSB attributed the crash to Uber's lax safety measures, including disabled emergency braking and inattentive safety operators.
Tesla has also faced frequent accidents, yet reports persist of drivers sleeping while using Autopilot. One incident involved a drunk driver napping as the car sped down the highway.
Reality isn't a game. When people entrust their lives to AI, do they realize there's no respawn if the system fails? Beyond investor pitches, autonomous driving may test not technology but human greed.
The issues mentioned are just the tip of the iceberg. AI isn't without merits—it processes vast data ranges and excels in areas like visual and speech recognition, nearing human performance. However, AI still falls short of true intelligence. Computers were built for logic, while biological intelligence evolved over millennia. The former excels at calculation; the latter, adaptation.
Today's AI increasingly mimics human instincts, highlighting the "Moravec's Paradox": high-level reasoning requires minimal computation, but unconscious skills demand immense power. Teaching a computer chess is easy; giving it a toddler's perception is nearly impossible.
At its core, AI is just data and algorithms. Most models solve classification and clustering problems—akin to human induction and deduction. Without true understanding, AI lacks adaptability and contextual awareness. Misidentifying an ad featuring Dong Mingzhu as a jaywalker isn't surprising—the AI merely detected a face on the street.
Moreover, machines lack emotions. An AI can label a cat in a photo but won't feel its cuteness.
Many artists have begun using so-called 'neural network software' to create artworks, some of which have been auctioned for astronomical prices. However, AI is merely a tool for artists—they import training materials and continuously adjust parameters to produce works, which is essentially no different from rendering images with 3ds Max. Therefore, it does not mean that AI possesses creativity.
At this stage, there are still many things AI cannot do. It is just a tool; without humans, AI is nothing. Humans also need not worry about AI rebelling and dominating humanity as depicted in movies.
Scholars have analyzed the probability of various industries being replaced by AI within the next 10 years. Since the Industrial Revolution, humans have never faced such large-scale replacement by machines.
Although this may seem alarming, it presents a significant opportunity for AI product entrepreneurs. However, seizing this opportunity is not easy. Often, the application of new technologies does not proceed smoothly.
According to Gartner's 'Hype Cycle for Emerging Technologies,' new technologies typically go through five stages: the Innovation Trigger, Peak of Inflated Expectations, Trough of Disillusionment, Slope of Enlightenment, and Plateau of Productivity.
Historically, artificial intelligence has experienced 'three rises and two falls.' Whether it will face another winter remains unknown.
However, integrating immature new technologies into products poses a significant challenge for product managers.
At present, we should not overly idealize AI's capabilities. Instead, we need to examine the development and application of AI technology with a calm and critical eye.
When encountering an AI-related requirement, rather than indulging in self-satisfaction, it is more useful to raise questions: What if the AI technology implementation fails?
What should we do then? Stay tuned for the next installment: Product Thinking in the AI Era (Part 2): New Challenges.