AI Writing The Human Touch Still Matters
A humorous comment in my email signature, taken from a Hackaday reader, has for a while suggested an article I penned was the work of a "Dumb AI". I can assure you, I am entirely human, a writer like many others—currently even dealing with a sore shoulder from a sports injury. This makes the accusation amusing in a way the commenter likely didn't foresee. Like many in the tech field, I'm skeptical about the proclaimed future of large language model (LLM) generative AI and have avoided fully buying into the current hype.
How AI Learns Your Style
As you're likely aware, these large language models operate by amassing an enormous collection of text. They then use complex computations to generate output by making inferences from this data. This capability allows them to create art mimicking a painter who gets no credit, or a book in the style of an author perhaps struggling financially. From the perspective of content creators and intellectual property holders, this can feel like large-scale theft, leading to numerous legal challenges aimed at defining the boundaries in this evolving field.
Consequently, once an LLM has processed sufficient text from a specific source, it can become quite adept at replicating that writing style. ChatGPT, for instance, has undoubtedly analyzed the entirety of Hackaday. Given that I've contributed thousands of articles over nearly a decade here, it possesses a substantial corpus of my work. Could it genuinely write in my style? It turns out, yes, it can—but not perfectly. I decided to put its forgery skills to the test.
Putting ChatGPT To The Test
A couple of weeks ago, I wrote about the anticipated re-entry of a Soviet Venus mission that has been stuck in orbit since 1972. It's a captivating story because the spacecraft might survive the descent, and it's conceivable that by the time you read this, it might have been recovered. I prompted ChatGPT to generate a 200-word article on this topic 'in the voice of the Hackaday writer Jenny List'. Here is what it produced:
In the early days of the Space Race, the Soviet Union launched a bold bid to reach Venus: the Venera missions. Among them was Kosmos 482, a 1972 mission that never made it past Earth’s gravity. Now, over five decades later, its descent craft is finally coming home—with a bang.
Kosmos 482 was intended to deliver a lander to Venus, but due to a malfunction in its upper stage, it was stranded in a highly elliptical orbit around Earth. Encased in a semi-spherical titanium shell, the descent craft—designed to survive the brutal atmosphere of Venus—has been orbiting quietly, steadily losing altitude. By early May 2025, that orbit had decayed to a precarious 232 x 136 km, setting the stage for an uncontrolled reentry around May 10th.
Unlike typical space debris, Kosmos 482’s reentry carries a bit more drama. Built to endure Venus’s hellish surface, the lander is expected to survive Earth’s atmospheric plunge mostly intact. Impact predictions estimate it will hit the ground at about 65-70 m/s, roughly the speed of a meteorite, packing the punch of a 500 kg hunk of Soviet engineering.
With an orbital inclination of 51.95 degrees, it could come down anywhere between 52°N and 52°S, adding an element of unpredictability. One thing is certain: after 53 years in limbo, Kosmos 482 is finally coming back to Earth—whether we’re ready or not.
Reading the AI's output is an intriguing experience because, on the surface, it does feel like a Hackaday article. It captures my general style with certain phrases, such as "packing the punch of a 500 kg hunk of Soviet engineering". The language is casual, and it has picked up on my tendency to be wordy, unlike the more concise style it produces when asked to emulate my colleague Al Williams, for example. However, it falls short in several aspects. I find it doesn't quite capture my sentence structure or the distinctly British way I use punctuation. It also relies on far more numerical figures than I would typically include in an article. Perhaps most critically, it lacks inference. The AI bases its writing solely on the facts it can find—the abundance of figures is a symptom of this—and it doesn't venture further to comment on potential outcomes. In my human-written piece on the same topic, I concluded by speculating on who would own the craft if retrieved; any such nuanced thought is absent from its version.
The Irreplaceable Human Touch
So, it's evident from this experiment that while ChatGPT can generate something that superficially covers the same story in my voice for a Hackaday piece, it still doesn't fully succeed at the task. Where a human writer would aim to provide some introduction, background, and commentary to enrich the story, the AI merely presents a summary of facts it has gathered. The fact that it cannot infer beyond the immediate story is reassuring, as it means we human Hackaday writers still hold an advantage. Many people will advise looking for specific words as tell-tale signs of AI-generated text, but the truth is much simpler: look for the human touch.
