On September 23, I gave a popular science talk at Hsing-Hwa Senior High School in Taiwan, during which I shared some thoughts on how to make responsible and productive use of generative artificial intelligence. As a cognitive neuroscientist studying the human brain’s ability to learn computer programming, I have become increasingly interested in how people interact with AI and how this technology may change the skills we need to develop. You can find the slides HERE.
Earlier this year, my colleagues and I published a commentary in Nature Reviews Psychology, titled Programming education in the AI era. We argued that the growing ability of AI to generate computer code makes computational reasoning and problem-solving skills all the more important. Although our commentary focused on programming education, I believe much of the argument applies to other activities involving generative AI, particularly creative work.
For my talk at Hsing-Hwa, I developed some additional thoughts on this topic, drawing on my own experiences with ChatGPT. I would like to elaborate on them here.
Knowing what we’re working with
Before discussing how to use AI effectively, I think it is important to have a reasonable understanding of how the technology works. We don’t all need to become machine learning experts, but knowing the basic principles helps us develop realistic expectations.
At the core of the training of large language models (LLM) is a relatively simple task: predicting the next token in a sequence, given what comes before it. Through training on enormous amounts of data, these models acquire surprisingly sophisticated capabilities. With additional training and the integration of external tools, modern AI systems can perform tasks that go far beyond ordinary text prediction.
Nevertheless, their outputs are not guaranteed to be factually correct or logically consistent. A generated answer may sound perfectly reasonable even when it contains errors. The same applies to AI-generated images. A model may produce a beautiful illustration while getting some elementary details of anatomy or physics wrong.
I have experienced this firsthand while using ChatGPT to design anime-style characters for an original story set in the World of Warcraft universe. For example, I once asked it to draw an archer, only to find that the character was holding the bow and arrow in a physically impossible manner. In other attempts, the model altered important features of fantasy races, added pieces of equipment that made no sense, or gradually changed the artistic style as I repeatedly requested revisions.
Another amusing observation was that ChatGPT seemed to favor certain hairstyles and facial features. Without sufficiently detailed instructions, characters with quite different backgrounds would somehow end up looking suspiciously alike.
These experiences have taught me to pay attention to the assumptions that generative AI makes when interpreting my requests. It has learned a great deal about the appearance of objects, but its learned visual patterns do not necessarily correspond to a coherent understanding of how those objects work.
AI-generated texts have their own peculiarities. In preparation for my talk, I collected several examples of AI-generated articles that displayed remarkably similar writing styles. Many relied heavily on repetitive sentence structures, dramatic contrasts, exaggerated rhetorical flourishes, and unnecessarily emotional conclusions. Some passages were so predictable that one could almost guess how the next sentence would begin.
I find these patterns quite entertaining, but they also demonstrate why I am reluctant to accept AI-generated writing without further editing. Whenever I use AI to help with writing, I want the final product to reflect my own ideas and writing style.
The 70% problem, now the 80% problem
One of the concepts I introduced during the talk was the “70% problem” of AI-assisted coding, originally discussed by software engineer Addy Osmani in December 2024.
The idea was that AI tools could quickly produce roughly 70% of a working software project, while the remaining 30% could be disproportionately difficult to complete. An AI-generated application might look impressive at first glance, but making it reliable, secure, and maintainable requires considerable additional work. (Interestingly, I later learned that Osmani had already revisited this concept in January 2026, in an article titled The 80% Problem in Agentic Coding. Apparently, my presentation slides were a little behind the times!)
The updated article acknowledges the tremendous progress in AI coding agents, which are now capable of handling much more of the development process. In some contexts, AI can complete 80% or even more of the work. However, Osmani also identifies a growing problem of comprehension debt: as AI produces increasingly large amounts of code, developers may struggle to understand what the code actually does, making verification and future maintenance difficult.
Of course, neither 70% nor 80% represents an experimentally established limit. The amount of work that AI can accomplish depends heavily on the project, the tools available, and how we define completion. A prototype that satisfies 80% of the requirements may still require most of the total development time to become production-ready.
I find this idea particularly relevant to creative applications of AI, although the difficulties involved are somewhat different from those in software engineering.
Suppose we are using generative AI to design a promotional poster. Within a few seconds, we can obtain a reasonably attractive image that would have taken considerable effort to draw manually. However, we may soon discover that the text is incorrect or the composition is awkward. Even more likely, some visual elements, like the anatomy of the cartoon characters, are just completely wrong. We can ask AI to correct these problems, but sometimes the revisions introduce new problems. After several rounds of refinement, we might find ourselves spending more time fixing the generated material than we anticipated.
I have encountered this problem repeatedly while working on my character illustrations. Sometimes, after many iterations of prompting, I simply have to open an image editor and make the corrections myself. At other times, I find it more efficient to start over with a better reference image or a more carefully formulated prompt.
The 80% problem suggests that we should consider the effort required to reach our desired level of quality when deciding how to use AI. If we only need a rough prototype, an imperfect result may be entirely adequate. For a finished product, there may still be considerable work ahead.
Some inequalities of AI usage
Inspired by Osmani’s observations, I proposed a few deliberately simplified mathematical-seeming “inequalities” during my talk. These are, of course, general principles rather than strict mathematical laws, but I think they capture some useful ideas about human-AI collaboration.
Domain expert + AI > Layperson + AI
A professional artist understands artistic concepts that a layperson may not even know how to describe. When using a generative AI tool, the artist can specify the desired composition, lighting, perspective, color scheme, and rendering style with much greater precision. More importantly, the artist knows how to evaluate the result and identify what needs improvement.
This applies to many other professions. An experienced software engineer can recognize problematic code and provide useful debugging instructions. A skilled writer is better equipped to recognize awkward expressions or inconsistencies in an AI-generated article.
I have certainly noticed this advantage in areas where I possess some expertise. My experience with graphic design, for example, helps me communicate specific stylistic requirements to image-generation tools. I can ask for cel-shaded coloring, consistent shadow planes, or a particular treatment of material textures because I understand what these terms mean and how the results should look.
On the other hand, when I use AI to assist with something outside my expertise, I often struggle to judge the quality of its output. I may be able to obtain an answer quickly, but I cannot always verify whether it is a good one.
This is also why I believe students should continue developing knowledge and skills in subjects they are interested in. Expertise gives us greater control over our use of AI, even when we delegate much of the actual production to the machine.
Existing material + AI refinement > Unguided AI generation
For another example, consider those rather horrifying AI-generated restaurant menus that occasionally circulate on the internet. Some feature distorted food, strange combinations of ingredients, or objects that barely resemble anything edible.
I suspect many of these problems could have been avoided by starting with photographs of actual dishes and using AI to improve their presentation. This would give the model a concrete reference to work with, reducing the amount of interpretation required.
Similarly, when designing an illustrated character, I often find it useful to provide reference images and ask AI to modify them. Starting with an existing image allows me to specify which features should be preserved and which should be changed.
Of course, this approach is not foolproof. AI image-editing tools sometimes regenerate large portions of an image even when instructed to make only minor changes. Nevertheless, having an existing design gives us a much better starting point for controlling the output.
Another approach is to generate individual visual elements separately and then assemble them manually. This is particularly useful when a complex composition contains many interacting objects, each of which might otherwise introduce errors.
AI output + Time and effort for refinement > Unedited AI output
This last inequality may sound rather obvious, but it is worth emphasizing.
Generative AI makes it extremely easy to obtain something that looks finished. A polished-looking illustration or a well-formatted article may give us the impression that little further work is necessary. However, appearances can be deceptive.
I have found that taking the time to examine AI-generated material carefully, identify its shortcomings, and make deliberate revisions can substantially improve the final result. Sometimes I ask AI to produce a new version; sometimes I edit the material myself. I also find it useful to compare multiple outputs and combine their best features.
The important thing is to have a clear standard for what we want to achieve.
There is also a practical limit to how much we should rely on repeated prompting. If an image-editing model keeps making the same mistake, asking it to try again for the twentieth time may accomplish very little. A manual correction or a different workflow might be much more efficient.
At the same time, I don’t think every AI-generated product needs extensive refinement. If we are merely exploring an idea, a rough draft may be sufficient. Spending hours polishing something that was never intended to become a finished product would defeat much of the purpose of using AI.
The appropriate amount of effort depends on what we intend to do with the result.
Learning AI without FOMO
One concern I wanted to address in my talk was the fear of missing out (FOMO) that often accompanies the rapid development of AI.
Every few weeks, it seems that another AI model or application is released, promising better performance or an entirely new range of capabilities. People understandably worry about falling behind, particularly when they hear that AI will transform their professions.
I have seen people devote considerable effort to learning the latest AI tools, sometimes without having any particular use for them. I wonder how much of this effort will remain useful a year or two later, given the speed at which these technologies are evolving.
A friend working in the engineering industry once shared with me a motto that I found particularly insightful:
「AI 工具，學得慢就是學得快。」
My translation would be: When it comes to AI tools, late learners learn fast.
The idea is that a tool requiring substantial effort to learn today may become much easier to use in the future. There is little point in mastering every new application simply because it exists.
I think a more productive approach is to identify what we want to accomplish and then learn the tools that can help us achieve it. When we have a specific project in mind, we also have a clearer motivation to learn the relevant techniques.
This doesn’t mean we should ignore new developments, though. I enjoy experimenting with AI tools myself, and I believe that maintaining curiosity about emerging technologies is valuable. However, I would rather spend my time learning something that can contribute to my current projects than chase every new product announcement.
Furthermore, using AI effectively can itself become a field of expertise. Techniques such as retrieval-augmented generation (RAG), context engineering, and agentic workflows require technical understanding and computational thinking. There is a considerable difference between casually asking ChatGPT a question and designing a systematic AI-assisted workflow for a complex project.
As these tools continue to improve, I expect the specific techniques we use will change. However, the ability to define a problem clearly and evaluate whether a proposed solution actually works should remain useful.
AI, learning, and responsibility
There is one additional distinction that I find particularly important for students: using AI to complete a task and using AI to learn how to complete that task may require different approaches.
When we are working on a professional project, efficiency can be a major concern. If AI allows us to automate a routine task without compromising quality, that can be a considerable advantage.
In an educational setting, however, the learning process itself is often the purpose of the exercise. A student who asks ChatGPT to solve every mathematics problem may finish the homework much faster, but the student may also miss opportunities to practice the underlying concepts.
AI can be a useful tutor. We can ask it to explain unfamiliar ideas or provide hints when we are stuck. It can also be used to examine our solutions for mistakes. The challenge is to make sure that we are actually developing our own understanding along the way.
This question is especially relevant to programming education. In our recent commentary, my colleagues and I discussed how programming relies on reasoning abilities that go well beyond knowledge of programming language syntax. AI can generate code for us, but we still benefit from understanding how an algorithm works and whether it solves the intended problem.
The same issue arises when using AI for writing or other creative activities. We should be able to recognize when the generated material is poor in quality or simply untrue. The extent to which we can make these judgments depends on our own knowledge and experience.
Ultimately, I believe responsibility for an AI-assisted product rests with the person using the tool. If I publish an article containing factual errors, the fact that some of the text was generated by ChatGPT would not excuse my failure to check it. Likewise, if I use AI to produce an illustration, I should examine the result before presenting it as a finished work.
Final thoughts
I am enthusiastic about the possibilities of generative AI. It has already allowed me to explore creative ideas that I would have struggled to realize on my own, and I regularly use it to assist with programming and research-related tasks. I also find myself spending a surprising amount of time correcting its mistakes. Sometimes, the experience is rather frustrating. Nevertheless, I have learned a great deal from figuring out why a particular approach failed and how to improve it.
And yes, this article is generated by ChatGPT from a draft I fed to it. I stand by this article. After all, I did spend close to two hours to lay out the points I want it to cover, and it took me more than an hour to edit its output, so I’m sure it’s not AI gibberish.
I hope the students who attended my talk will feel encouraged to experiment with AI and discover ways to use it in their own interests. At the same time, I hope they will continue cultivating their own expertise and judgment, and resist the temptation to accept every generated answer at face value.
The tools will undoubtedly improve, and some of the limitations I discussed may eventually disappear. When that happens, I will gladly update my expectations and find new ways to make use of them.
For now, the motto I shared at the end of my presentation still seems appropriate:
AI is a tool, and people are the ones using it.
And perhaps, when learning AI tools, taking our time can indeed be the faster way forward.
