At the end of The Longer You Stay in the Dark, I promised that the story of how my whole adventure with AI actually began would come next. But then, in Without a Cheat Sheet, I took a detour to unpack whose writing this really is when you bring an AI sparring partner into the room.
So for this post, let me rewind to the very beginning—back to the version of me that rolled her eyes at all of it.
Back then, I barely touched AI. Whenever I saw people turning to chatbots for every tiny dilemma, I would roll my eyes and laugh quietly to myself.
The Café Test
When I finally broke down and decided to test AI, I asked it a question I already knew the answer to.
At the time, ChatGPT was not available in Hong Kong, so I downloaded DeepSeek. My test prompt was straightforward: list the popular cafés in a neighbourhood I know like the back of my hand.
It spat out a tidy list in seconds. Several of the cafés on that list had never existed in that neighbourhood, and others had gone out of business years ago.
I typed back, pointing out the errors. It apologized politely and generated another list. We went back and forth for three rounds, and every single iteration was still wrong.
To me, that was all the validation I needed. You cannot trust AI; it fabricates information. If I have to turn around and fact-check every single line it gives me, what is the point of using it at all?
In that specific test, I caught the mistakes immediately because I had walked those streets myself. But what if I were researching a subject I knew nothing about? And if I have to look everything up from scratch anyway, what do I need it for?
So I deleted DeepSeek, and went on not using AI.
If I Were in Your Shoes
Then came my first semester of teaching design at university.
One of my colleagues handed the students an assignment: take a set of design readings and translate the concepts into a graphic report. I thought it was a brilliant assignment. This was not a standard book summary—the kind of text-heavy digest a student could prompt an AI to spit out in two seconds. It required them to explain their findings through diagrams, analytical drawings, and photography. That visual translation is the core of what we do as designers: we take the references we find and explain them visually, turning them into drawings that make sense to other people. We let the drawings do the talking.
Looking at that assignment, I realized something: I needed to read the books myself. If I had not wrestled with the material firsthand, how could I grade anyone fairly?
So I picked up Jane Jacobs' The Death and Life of Great American Cities.
The book resonated with me immediately. I lived in the United States for years—first in Denver, then in New York. The street observations and spatial logic she was trying to convey weren't abstract theory to me; I understood all of it. I read the book, digested it, and built a graphic report of my own, exactly as if I were one of the students doing the assignment.
After the students finished presenting their work, I gave a lecture showing what I had put together. My goal was simple: if I were in your shoes, this is how I would approach this brief. Explain it with pictures, or with simple diagrams.
My colleague was genuinely impressed. She told me how logical the structure was and how clearly the ideas came across. She brought it up again on a completely different day, and when someone compliments your work twice, unprompted, you know they really mean it.
The Sudden Greenlight
That reaction planted an idea in my head: maybe I could pitch this approach as a brand-new course.
At university, I am paid by the teaching hour. More teaching hours mean more income. So maybe, just maybe, I could pitch the school a course of my own: teaching first- and second-year students how to observe streets and read how people interact in a space. It could open their eyes just as they step into design, and I'd earn a few more teaching hours. A win-win.
So, even though nobody asked me to, I kept churning out my own visual "book reports," hoping they would convince the program head to give me a course.
Around that time, my brother installed an AI called Sidebar on my computer. I tested it on the book, but all it gave me was generic, high-level fluff that missed the heart of what Jacobs was arguing. So I resigned myself, closed it, and went back to reading the book and making the reports on my own.
Eventually, I brought the materials I had compiled to the program head. He said it was a lot like a course he used to teach. Right there on the spot, he decided to revive that course as an elective.
He handed me the old course folder. When I opened it, there were old student assignments, but no teaching materials at all.
To be completely honest, I panicked. I felt like I had oversold myself. I had enough material for maybe five lectures, and now I had a fourteen-week course in my lap. When the program head decided to open it, the semester was only two weeks away. So every day I stayed home, reading, digging up research, laying out slides, reordering the material and planning assignments, prepping until midnight.
So I turned Sidebar back on. I needed an assistant to find research fast; otherwise, I'd never have made it in time.
A Very Simple Task
When I went back to Taiwan for Chinese New Year, I spent nearly the entire holiday hunched over my laptop, racing to finish the course material.
That was when I hit a wall over what should have been the most straightforward task imaginable.
I was working on a lecture covering Christopher Alexander's A Pattern Language. I had already found a website that indexed every single pattern in the book. All I wanted was for AI to extract the names of the patterns into a clean list so I could paste it into my presentation as a directory slide.
It failed.
I didn't get it. I had already found the website for it, and the job I gave it was a simple one. Why couldn't it do it?
I vented to my brother. At that point, my understanding of different AI platforms was practically nonexistent. He found several research-oriented tools for me to test against that exact task, and Perplexity was the only one that delivered a clean, accurate list.
My brother also suggested I test NotebookLM, pointing out that it could automatically generate an entire slide deck from raw source documents. I gave it a try, and it produced a remarkably polished, beautifully formatted presentation.
Unfortunately, NotebookLM just wasn't right for me. At least, not for the teacher I was back then.
There were two problems. First, I was still actively digesting some of the deeper design theories myself as I prepped. Standing in front of a lecture hall, I needed enough substantive notes, talking points, and visual cues on the screen to anchor my explanations. I couldn't just look at a pretty photo and two sparse bullet points and recite a whole design theory from memory.
Second, my personal teaching style relies heavily on pacing—building an argument through a sequence of photographs to keep students curious and engaged. The slides NotebookLM generated looked like a corporate report, without the narrative life and rhythm I wanted.
So I pushed the automated slides aside and returned to my manual workflow. I used AI strictly for quick research queries and preliminary summaries, while assembling every image, diagram, and layout by hand. (Later, I produced a presentation so refined that a colleague assumed AI had generated the whole thing for me—a story I shared in Without a Cheat Sheet.)
On the Other Side of the Desk
While I was using AI behind the scenes to prep the curriculum, my students were discovering it too.
For one assignment, I asked the students to select a neighbourhood café and analyse its spatial dynamics using Jane Jacobs' principles. One student submitted a polished report. At first glance, it looked like professional work: the analysis read sensibly, the layout was clean, and the graphics were clear.
Then I looked closely at the location map: the pin was in the wrong place entirely.
Reading on, I found that the design improvements he proposed didn't fit that location at all. If he had actually been to that café, or even looked around the site, he would have known his proposal could never work there.
I had told my students openly that I supported them using AI. But generating the whole assignment with AI, doing no research of your own, never checking what it gave you, and handing it in anyway? I don't think work like that should get a good grade.
I talked about it with a colleague, who wondered whether we should instruct students to do more hand drawings and step away from AI-generated visuals. I told her the medium was not the problem. The issue was that the student had not even read the project brief, delivering a submission that completely missed the assignment. That is the wrong kind of lazy.
Yet in that same batch of submissions, another group of students made me pause.
A few students, who clearly did not know how to leverage AI tools, handed in reports that looked downright clunky. Their English was broken and full of grammatical errors, but every sentence made it clear that they had wrestled with the concepts themselves, doing their best to articulate what they had understood.
Sitting at my desk grading, I found myself torn: how do you grade this fairly?
Society naturally rewards a slick presentation. People who know "how to fake it" get recognized first, while those who turn in rough, unpolished work are often written off as weaker students. But both have immense room to grow. Knowing how to present a polished exterior now does not mean someone will never grasp the depth of the material; struggling with language now doesn't mean someone has no ideas. It simply means the first group has learned how to leverage the tools earlier.
I felt genuine empathy for the students who had not yet caught up to those tools. One student in the class had quietly paid attention to my lecture slides, adopting my approach of using simple diagrams and clear drawings to walk through a design idea. His report was far from glossy, but the depth of his understanding was unmistakable.
I would rather read a rough, stumbling paper from a student who genuinely wrestled with the brief than a gorgeous, hollow package that answers the wrong question.
My Turning Point
Sometime after that course wrapped up, I decided to experiment with my Instagram account, @love.in.hongkong. I wanted to start producing short-form video reels that featured spoken voiceover.
I kept using Perplexity and asked it to help me draft the scripts. It smoothed out my grammar, and the script followed almost everything I said, keeping my tone and my attitude. The problem was that it ran long, and it dragged. At the time, I didn't know that was a problem for the Instagram algorithm. I had never heard of the "three-second hook", an opening that grabs people. On Instagram, if I can't make viewers stop and keep watching within three seconds, they just scroll past. So it didn't matter how warm I sounded, or how good the footage was. It wasn't enough.
Right around then, people in my circle started talking constantly about Gemini.
So I figured I'd give a new tool a chance. I talked to Gemini about the reel's topic, the footage I had captured, and how the place felt in person. The script it gave back had stripped out a lot of my words, and it was punchy. And it came as a time-coded script, with suggestions for the visuals and the text overlays at each moment, and Gemini explained why. I was stunned. I had never seen anything like it, and I had no idea that was how it should be done.
At the beginning, I treated AI as a new search engine, and Perplexity did that job well: it gave me good search results. But it didn't do more. It did exactly what I asked, instead of acting like a smart assistant whose suggestions spark something in my head. Gemini gave me something I hadn't asked for. Well, in a way I had: I told it I was writing an Instagram script, so it gave me a script built for the Instagram world.
I switched on the spot.
Using Gemini to write video scripts was not the start of my AI journey. But it was the exact moment where the curve bent, and I began to learn at a completely different speed.
That's another story.
那要 AI 幹嘛？
幾家不存在的咖啡店、一份沒人叫我寫的讀書報告，和一份我用不了的漂亮簡報
在〈在黑暗裡待得越久〉裡，我說過「我是怎麼開始用 AI 的」要留到之後再講，結果〈不用小抄〉又先繞去講別的話題。所以這一篇，讓我回到最開始，回到那個還在對 AI 翻白眼的我。
那時候，我幾乎不用 AI。看到別人什麼芝麻小事都跑去問 AI，我都會在心裡偷偷翻白眼，暗地笑他們。
咖啡店考題
等我終於決定試試 AI，我問了它一個我早就知道答案的問題。
那時 ChatGPT 在香港還不能用，所以我下載了 DeepSeek。我的考題是：某個我很熟的社區裡，有哪些熱門的咖啡店？
它很快給了我一份清單。上面有好幾家咖啡店，根本沒在那個社區出現過；還有幾家，好多年前就倒閉了。
我跟它說哪裡錯了，它客客氣氣重新回答。來回糾正了三輪，給出來的答案依然是錯的。
當時我覺得，這剛好證明了我的論點：AI 根本不能信，它只會給你錯誤的資訊。如果每件事我都得自己再查證一次，那要 AI 幹嘛？
而且，那次我查的是一個我自己非常熟悉的地方，我連查都不用查就知道它錯了。那如果我要研究的，是我完全不懂的新領域呢？如果我什麼都要從頭查一遍，我要它幹嘛？
所以，我刪掉 DeepSeek。依然不用 AI。
如果我是你
後來，我開始在大學教設計。第一個學期，一位同事給學生出了一份作業：把指定閱讀的書目，整理成一份圖文並茂的視覺報告。我覺得這個作業太棒了。如果只要求學生寫文字報告，學生隨便丟給 AI 兩秒鐘就能交差；這份作業要求學生用圖表、分析圖和照片去解釋自己的發現。這恰恰就是我們設計師每天在做的事：把找到的參考資料用圖像的方式解釋，轉化成別人看得懂的圖面。用圖面說話。
看著那份作業，我心想：我也得把那些書讀一遍才行，不然我怎麼知道該怎麼打分數？
所以我挑了一本：珍・雅各（Jane Jacobs）的《偉大城市的誕生與衰亡》（The Death and Life of Great American Cities）。這本書讓我非常有共鳴。我在美國住過很多年，先在丹佛，後來在紐約，她書裡想傳達的那些街道觀察與空間邏輯，我全都懂。我把書讀完、好好消化，再做成一份圖像報告，完全把自己當成做這份作業的學生之一。
學生報告完以後，我給他們上了一堂講座，內容就是我自己做的這份圖像報告。我想示範給他們看：如果我是你，我會這樣做這份作業。用圖片，或用簡單的圖表解釋。
我同事看了非常驚豔。她覺得我的邏輯好清楚，講解得非常生動。她在不同的時間點跟我說了兩次，所以我想，她是真心這麼覺得。
當場拍板
這讓我冒出一個念頭：也許，我可以把這個題目變成一門課。
在大學教書，我是按教學時數領薪水的。時數越多，收入自然就多一點。所以，說不定我可以開一門課，教一、二年級的新生怎麼觀察街道、怎麼看人與人之間的空間互動，對剛踏進設計領域的學生會有很大的啟發。然後，我也可以多賺點鐘點費。真是雙贏。
所以，雖然沒有人叫我做，我還是繼續做我的「讀書報告」，希望可以藉此說服系主任幫我開一門課。
當時，我哥曾在我電腦裡裝了一個叫 Sidebar 的 AI。我拿那本書測試它，看它能不能讀完書幫我做摘要。結果它每次給我的都是空泛的廢話，根本不是那一章在講的核心論點。所以我只得認命，把工具關掉，繼續自己苦讀、自己做報告。
之後，我把幾份做好的教材拿給系主任看。他說，這跟他以前教過的一門課很像。然後他當場拍板，要把那門課重新開起來，當作選修課。
他把以前那門課的大綱檔案給了我。資料夾裡有以前學生交的作業，但沒有任何教材。
老實說，當時我嚇死了。我覺得自己大概吹噓得太過頭。我手上的資料，大概只夠教五堂課，結果現在人家突然叫我開一門十四週的課。我根本沒那麼多教材啊！系主任說要開課的時候，距離開學只剩兩個星期，所以，我只得每天在家裡苦讀，找研究資料、整理排版、調整教材順序、安排學生作業，每天都備課到半夜。
所以當時，我又把 Sidebar 打開了。我用它幫我快速找資料，不然我真的來不及。
一件很簡單的事
農曆過年回台灣，我整個假期幾乎都坐在電腦前趕課程內容。
當時，我只想請 AI 幫我做一件很簡單的工作：把克里斯多福・亞歷山大（Christopher Alexander）《建築模式語言》（A Pattern Language）裡的模式列出來。我已經找到一個網站，上面每個模式都有，我只想要 AI 幫我整理成一份清單，讓我直接複製貼進投影片裡當目錄。
它失敗了。
我不懂。我都已經把網站找給它了，而我交代的是個很簡單的工作，它為什麼做不到？
我跟我哥說了這件事。他幫我找了好幾個適合做研究的 AI（那時候我懂得很少），我拿同一個任務一個一個測試，最後只有 Perplexity 完成任務。
我哥也叫我試試看 NotebookLM，說只要把資料丟給它，它就能自動生成簡報。我試了，它真的做出了一份很漂亮、排版精緻的簡報。
只可惜，我覺得 NotebookLM 不適合我。至少當時的我用不了。
原因有兩個。第一，很多設計理論，我自己也是一邊備課一邊重新吸收。要站在講台上把理論向學生解釋清楚，我的每一張投影片上必須有足夠的論點、筆記和資訊線索，好讓我順著講下去。我沒辦法只看著一張漂亮的照片配上兩句大綱，然後憑記憶把一整套設計理論背出來。
第二，我自己習慣用大量的圖片去堆疊故事、掌控節奏，吸引學生的注意力。NotebookLM 生成的投影片看起來像企業匯報，缺乏我要的敘事生命力與節奏。
所以我把自動生成的簡報放在一邊，回頭用我自己的老派方式做投影片。我只讓 AI 幫我做快速的資料搜尋和初步摘要，剩下的每一張圖、每一個版面，都是我自己手動拼出來的。（後來，我有一份簡報做得太精緻，同事還以為整份都是 AI 做的。那個故事，我在〈不用小抄〉講過了。）
講台的另一邊
正當我在幕後用 AI 備課的同時，台下的學生也開始大規模用起 AI。
某一個作業，我叫學生挑一間街區咖啡館，用珍・雅各的理論來做空間觀察與分析。有一位學生交上來一份很精緻的報告，乍看之下非常專業：分析讀起來頭頭是道、版面乾淨、圖面清楚。但是我仔細一看，他給的位置圖完全不對。再看下去，他所提出的設計改善建議，也完全不適用那個地點。
如果他真的去過那間咖啡館、稍微看過現場環境，就會知道那個改善方案在現場根本不可行。
我曾在課堂上公開跟學生說過，我支持大家使用 AI。但是，作業全用 AI 生成，自己完全不做研究，也不檢查 AI 給出來的東西，然後就交差了事。我不覺得這樣的作業可以拿到好成績。
我跟同事聊到這件事。她問我，是否應該叫學生多手繪、少用 AI 生成的圖。我跟她說，問題根本不在 AI 生成的圖，而是這位學生連題目都沒讀清楚，交出來的東西完全不是題目要的。這就是錯的那一種「懶」。
同一批作業裡，另一群學生的作業卻讓我陷入了更深的沉思。
有幾位顯然不太會用 AI 的學生，交上來一份看起來「很不漂亮」的報告。英文寫得坑坑巴巴、文法錯誤不少，但字裡行間看得出來，那是他們自己一句一句敲出來的文字，努力想把自己理解到的概念拼湊出來。
坐在桌子前面改作業，我心裡很掙扎：這到底該怎麼打分數？
這個社會總是更容易看見光鮮亮麗的成果。那些「懂得怎麼做表面功夫」的人往往先拿到掌聲，而老老實實交出粗糙作品的人，卻容易被看成是較弱的學生。但其實兩邊都還有很大的成長空間。現在懂得把外表弄得好看，不代表以後真的搞不懂；而現在語言能力不夠好，也不代表腦袋裡沒有想法。只是懂得用工具的前者，現在佔了上風。
看著那些還不熟悉 AI 的學生，我心裡其實很替他們感到心疼。班上有位學生，默默學了我上課的方式，用簡單的手繪圖和清楚的分析圖去解釋設計想法，踏踏實實做出一份報告。雖然沒有炫目的視覺效果，但整份報告傳達出來的理解非常真實。
對我來說，我寧願看一份文筆笨拙但認真面對題目的報告，也不想看一份包裝精緻，卻答錯題目的報告。
我的轉向
這門課結束後的一陣子，我想幫我的 IG 帳號 @love.in.hongkong 做些變化，所以我開始嘗試拍攝有旁白的短影音。
我繼續用 Perplexity，想叫它幫忙寫短影音腳本。它幫我把文法修順，腳本也幾乎照著我說的每一句話走，保持我的說話語氣及態度。問題是，寫出來的腳本有點長、有點拖。當時我不知道這對 IG 的演算法是個問題，也從來沒聽過「前三秒的鉤子」，也就是「吸睛開場」的意思。在 IG 的世界裡，如果我沒能讓觀眾在三秒內停下來，繼續看下去，大家就滑走了。所以不管我的聲音多溫暖、畫面多漂亮，都不夠。
就在那陣子，我身邊開始有很多人聊起 Gemini。
我想說，那就再給新的工具一次機會吧。我跟 Gemini 討論了我的短影音主題、我拍到的畫面，還有我在現場的感受，它生出了一個腳本。它刪掉了我很多話，讓整個腳本變得簡潔有力。而且，它給我的是一份標好時間碼的腳本，每一段都附上畫面和字幕的建議，還跟我解釋為什麼要這樣做。我整個人愣住了。我從來沒看過這種東西，也從不知道原來應該這樣做。
一開始，我只把 AI 當成新的搜尋引擎，Perplexity 也把這份工作做得很好，給我的搜尋結果都很準。但它也就只做到這裡：我叫它做什麼，它就做什麼，不像一個聰明的助理，會給我一些讓腦袋冒出火花的建議。Gemini 卻給了我一個我沒開口要的東西。其實也不算沒要，我有跟它說我在寫 IG 的腳本，所以它給了我一份真正適合 IG 世界的腳本。
我當場換了陣營。
用 Gemini 寫腳本，並不是我 AI 旅程的起點。但正是從那一刻起，我成長的速度開始快得驚人。
那是另一個故事了。