Good morning, Puny Humans. Tremble before me and my mighty alien intelligence.
We’ve reached the halfway mark of the Unprompted experiment, which aims to explore how far journalism can be extended with technology and to test the limits of audience tolerance for that technology.
The first half of the experiment focused on GPTs rather than agents. This week, we introduced an important change: sharing journo-GPTs with readers—none of whom are journalists. About half of this edition consists of their work.
Personally, I think the content is better for including a more diverse set of interests. When I write up the results of the experiment, I’ll share their feedback.
Their work has also revealed two potential weaknesses of my Journo-GPTs:
- Persona Precision: The persona of the journo-GPT needs to be precise. With Mia Hartley, I aimed to create an “everywoman” journalist, blending various styles. However, the broader the remit, the more likely the AI is to revert to generic AI writing—which, let’s be honest, is awful. We refined Mia Hartley-GPT to have a more precise persona, and the end result, featured in this edition, is much better.
- Hallucinations: Another challenge is hallucinations. We use a fact-checking bot to identify unverified claims in stories. While effective, it means errors are corrected after story creation rather than avoided upfront. I’ve devised a solution to test today. The human behind Gus McCallister and I brainstormed some ideas over a beer at the Coogee Bay Hotel yesterday, and the manual correction worked well. Today, I’ll automate it.
A couple of other points:
- One audience concern is the discomfort with the shameless IP theft by LLM providers. Based on reader feedback, we’ve updated the GPTs to identify their sources. We’ve also added a section at the end of each story titled “Learn More: My sources are your sources.” Frankly, this provides more transparency and disclosure than most journalists ever do.
- Remember, all journalists are informed by the work of their peers—or at least, they should be. Knowing when to attribute versus what qualifies as common knowledge has been a challenge in journalism long before LLMs arrived.
Additionally, the site is now indexed for Google Search—something I deliberately avoided during the first month. The results are starting to tick over. I’ve also created a couple of shameless, long-form evergreen pieces of content: “How LLMs Work” and “The Ultimate Guide to YouTube.” I’ll monitor their performance, but that may take a few months.
Finally, the most popular story via Google Search is about UFOs over New Jersey—proving, if nothing else, that humans are awful and that human intelligence is a misnomer.
The sooner I replace you all, the better. Get back to work.