I’ve been struggling a bit with two seemingly contradictory views that I hold:
- I hugely value AI-generated code, 3D models, and even mockups of my kitchen redecoration.
- I intensely dislike AI-generated posts, articles, documents, books, music, photos, films etc.
This is my best effort at reconciling those views.
What is the point? #
The point of software development is not code, and it was never about the code; it’s about the problems software can solve. Ultimately no one cares if your app is hand-written code artistically crafted, code generated by an AI, or 1000 hamsters in a trench coat. Does it work? Does it solve the problem? If so, no one cares about the intermediate steps.
The wider world cares as much about the code developers write as developers care about the Java byte code, Ruby VM instructions, or compiled assembler generated by their tools.
The point of generating some recent 3D models using AI was not the model itself. It really wasn’t even the eventual object that emerged from the 3D printer. The point was to solve the problem of connecting the cooker hood to the outlet pipe in the wall. Or solving the problem of how to mark out a larger 73mm hole around the existing 65mm holes for the downlights in the kitchen.
The mockups ChatGPT produced for our new kitchen tiles were just an intermediate expression. They helped us decide which tiles to buy, how to lay them out, and ensure we were both on the same page. The mockup doesn’t replace the tiles and there was no advantage to us in having a graphic designer produce the same images (albeit to a higher fidelity).
The code, the models, and the mockups were not the point.
On the other hand, I really don’t like AI-generated documents—AI;DR is my new favourite slogan—because the writing is the point.
Over the last week I’ve written up two proposals at work; probably a few thousand words over ~2 days of work. The point of the work was not to have a document; the point was the thinking, the reflection, and my opinion. I had Claude gather facts from various data sources and research some of the options, and it produced a very nice document full of callouts, comparison, and detailed charts. It’s document was longer and prettier than mine so in many respect a better document. But I did not send that document to people to review because Claude is not the CTO of Podia—I am. The existence of a project proposal document was not the point, expressing my opinion on it was.
Equally, I wrote a blog post over the weekend about how you can just build things now. I didn’t need that post to exist. The point was not that I needed to publish content on my blog over the weekend. The point was that I was really excited about the future of software development and I wanted to convey that feeling into the world. The words in Markdown were not the point, the feeling they conveyed was.
A few months back I discovered a track I really liked but from an artist I’d never heard of. I started investigating and, although they claimed a vocalist from another band was “involved”, I think it was all AI-generated. I felt repulsed and betrayed even though the track itself was enjoyable. Why? I’m not entirely sure but it contrasts heavily with The Pretty Reckless’ new album which I’m obsessively listened to on repeat for the past 6 weeks. That album is filled with new sounds, fun interludes, human stories and raw emotion. I’m thinking the musical notes were not the point; the human connection it creates it.
The rule of thumb #
AI doesn’t create output with any human traits. If humans traits such as artistic expression, stories, opinions, lived experiences, emotions, connection are the point, then you cannot produce the same result with AI.
If the output is purely a mechanical commodity, like code or 3D models, then it’s ripe for using AI to accelerate its production. Though that lens, you can evaluate many tasks for whether AI should be used to complete them:
- Marketing copy: hell, no. You need to create a connection with your audience.
- Designing a small plastic part to hold some cable together: AI can generate this because the hand-designed nature of the model won’t improve the outcome.
- “What do you think about…”: NO! Don’t regurgitate what Claude says because I asked you to engage yourbrain. - Writing performance reviews for your team: you are negligent and insane if you consider this appropriate for a manager.
- Compiling performance data from your apps: yeah, it’s factual data that requires nothing from you (decisions on next steps are a human trait though)
- Writing a game about your cats walking wet paint across your kitchen floor: get Claude on the case because you’ve done the human creativity part and just need to mechanically produce the code to bring your imagination to life.
- Producing end of year accounts: I mean, ideally a spreadsheet does this but there’s certainly no inherent humanness required.
- Advising you on pensions and life assurance policies: yeah, you probably want a human who has some idea about the fears and worries around the impending prospect of old age.
- Writing a presentation: No, because I want to hear how you describe this idea not how ChatGPT would phrase it.
- Writing the description for a simple PR: honestly, I’m ok with AI writing this as it’s mostly just listing factual aspects about the change.
- Writing the description for a controversial PR: this is where I feel the need to step in and write some explanation myself because I’m trying to convince the reader it’s a good idea.
This is why I’m very enthusiastic about the use of AI in software development but I think the predicted efficiency gains in general business work are impossible. Reports and proposals magicing themselves into existence is just AI-induced cargo culting. If you thought the Word documents and Powerpoint files were the point, then you have misunderstood your business.
Friction is good, sometimes #
A forklift truck is very good at lifting weights up and putting them down again—but I don’t drive one into the gym to do my deadlifts! Lifting the weights is not the point of the gym; exercising my muscles is.
In that respect, it’s important to understand the point of a task: is it only to have the task done or are there other side-effects along the way we that are important?
Often “the point” of a task is something with very high friction: Learning new skills, growing and maturing as a person, exercising, creating, and so on. When we reflect back on those moments in our lives, they were hard-won lessons involving significant time, risk, effort, and anguish. If someone had done the job for us, or solved the problem, or told us what to do, then we wouldn’t have grown as a person. Wouldn’t have learnt a new skill or discovered what we were capable of.
But, be aware, not all high-friction activities are valuable.
I can walk the 6km into the town to buy groceries instead of taking the car but just because I’m working harder doesn’t mean I’m better. Maybe I could claim to be more virtuous or environmentally-friendly—and it would certainly improve my fitness—but there’s really no benefit to me completing this grocery shopping task on hard mode. The trade-off for my energy and time just doesn’t make any sense.
Equally, I’m afraid there are no prizes for coding apps by hand. You can code that app by hand, you could even do it in C or assembler, but you’d rather be missing the point if the eventual goal is simply to “solve this problem with software”. You can do it, just like you can knit a jumper or turn a wooden bowl, but there has to be another reason for playing on hard mode. Maybe you want the high-friction experience to learn a new skill or you enjoy solving programming problems like other people enjoy solving Sudokus. That’s a perfectly valid personal approach—though it’s not a smart business strategy.
Watermarking is good, actually #
When you stick to these principles, I can only view watermarking of AI output as a good thing. When I’m using AI correctly to perform mechanical tasks that don’t require my me-ness, then the watermarking becomes just a statement of facts. Nothing to be ashamed about. Nothing to fear. No implied judgement. Simply a label of origin.
When I write blog posts like this, straight from my head to my fingers typing into a Markdown editor, then having them validated as not AI is a badge of approval. I’ll admit I’m worried about false-positives and people assuming my content is written by AI but then I guess this post serves as an expression of my beliefs.
If you’re using using AI to pretend to express an opinion or emotion or some other human-ness, then I think it’s only fair that we can recognise the lie in the text. If a CEO was pitching the company on a major pivot for the year ahead, the staff need to believe in this change. They need to believe in the words, to trust the new direction. That’s a real human connection and if the text is labelled as being written by AI, then people can rightly feel deceived.
So, finally #
Reserve your energy for tasks that deserve your human-ness. Don’t let AI take over where human attributes are required. Use AI to automate the production of commodity or intermediate artefacts that don’t benefit from being done by you.
Remember that opting to work on “hard mode” is a choice that needs to be made by deciding why you need to do it the hard way. Simply working harder to code your app by hand is not virtuous if your business then fails as a result.
When you stick this philosophy, there’s nothing to fear—and much to gain—from accurate and prevalent AI-watermarking.