I’ve crossed over to the “dark side” – I now have a paid subscription to ChatGPT.
It took me a while to make the leap, because:
- ChatGPT was trained on partly stolen data.
In the U.S., judges have ruled that creating a Large Language Model (LLM) like Anthropic’s Claude or Meta’s Llama can fall under fair use, but that illegally downloading books for training purposes is not. If there were a public alternative (and work is underway on GPT-NL, which handles rights holders properly), I would switch to it gladly. But until then, I don’t see enough reason not to use it—there’s no direct harm involved. - ChatGPT is said to be bad for the environment.
By now, the energy use of an LLM query and a Google search are roughly the same—around 0.3 Wh. Of course, generating lots of images uses more energy. But ten “ordinary” LLM queries equal about six meters of driving in a petrol car, or roughly seventeen meters in an electric one. Or eighteen minutes of a 10W LED lamp burning. In other words: hardly even a slice of bacon in the stew. For 100 grams of bacon, you could ask ChatGPT about 1,740 questions (and fly 500 meters, drive 1 km in a petrol car, or nearly 50 km on an international electric train). On that scale, I consider its use acceptable. - AI is trained on texts dominated by white male voices.
That doesn’t exactly help in striving for a fairer society. People are also losing the habit of expressing themselves well—a crucial social skill. And AI makes texts duller, because it pushes everything toward the lowest common denominator.
That said, I decided to subscribe to the pro version anyway. As the saying goes: one fool can ask more questions than ten wise people can answer—but that one ChatGPT actually does quite a decent job.
Here’s a glimpse of what I’ve used it for over the past month:
- Adjusting the Hilversum Time Machine logo (originally rectangular) to a square version for LinkedIn.
- Brainstorming for a call for papers on epistemic institutions and AI.
- Converting the location of a badger sett from RD coordinates into a Google Maps reference.
- Turning minutes of a meeting into an agenda for the next one (using the section headers).
- Researching traditions around St. Michael’s Day (29 September) for a book I’m writing with a friend on The Rhythm of Time.
- Converting weather data from De Bilt into an Excel sheet with conditional formatting, so I can instantly see whether the sun has been shining lately.
- Adjusting texts for clarity—“set to reading level B1”—then translating them back into what I actually want to say (though many people do find my texts easier to read afterwards).
- Quickly translating texts for international communication.
- Checking whether I could travel to a course in Copenhagen in an environmentally friendly way without losing a night’s sleep (answer: no).
- Creating a sing-along version of the alto part for the music at the Festival of Nine Lessons and Carols, which I’ll be joining later this year.
It’s already hard to imagine working without ChatGPT. I was especially impressed by that alto part and the logo work (though I have noticed it tends to ask extra questions—probably hoping I’ll get distracted so it doesn’t have to spend too much processing power! I started typing “hop to it” whenever it came back with yet another clarification request). And sometimes things just got worse—the original logo featured a stylized version of Dudok’s town hall, but adjusting the proportions turned out to be impossible.
People say that money is a bad master but a good servant. I feel the same about ChatGPT: it’s wonderful to have extra time for what I find truly important, because I can leave the rest to it.
Or, as Keynes said: “It’s better to be roughly right than exactly wrong.”
Of course, you still have to pay attention. ChatGPT doesn’t always capture nuance in translation (but then again, neither did human translators). It’s not the solution to everything, but it can do an amazing amount.