AI Trip reports

Where I work Artificial Intelligence is being adopted in a big way. This should come as no surprise to anyone. The possibilities are really quite extraordinary, you can see what it can do, how much time (and by extension, money) it can save, and in effort to turn that vision into reality. But what about outside of the workplace? What about blogging?

Using some newly acquired basic knowledge of prompting I decided I would use my as yet unwritten Arizona trip report - the one that just published - as a test case. Could I get AI to simply write it for me? It would save a lot of time, I could go birding instead, or more likely potter around in the greenhouse. Would my reader be any the wiser? What would Alan think?!

So tell me, could you tell?

I would say that you couldn't, and that the reason for this is that I found that AI could not in fact write a trip report that I was happy represented what I wanted to say and how I wanted to say it, and that I therefore ended up writing it myself. Business as usual. 

But not quite. The words on the page are mine, the style is my own. But I did retain some of what the bot had produced for me, and it saved me hours and hours. I think it made the difference between being able to get it done and not, and it still took a month between starting and finishing.

But let me take you on a journey of my attempt to get the machine to shoulder the burden. I used CoPilot - this is what in on my PC. There are lots of others, it's possible that they could do a better job. Then again perhaps they are all terrible and all lie just as much.

I gave it some basic information. I wanted a trip report organised day by day. I wanted it to be written in chronoogical order, site by site, listing the birds I saw there. To help it out I gave it two things. 1) the html addresses of each checklist one by one, but this was inefficient so instead I ended up giving it a csv eBird extract of all the checklists from the trip in one go, and 2) my list of targets which I simply uploaded as a list. I said I wanted it to specifically mention the targets I ended up seeing, although I later had to clarify the ones that I did not see, as - and I will come to this in a moment - AI took a few liberties.... Finally, in order to turn it into me, I gave the AI some examples of trip lists I had written before and said to use that same style, the same voice. Now it had everything.

Ready Player One! I asked it to write a brief paragraph about each checklist, mentioning the pick of the birds that I saw and displaying these in bold text. Then I asked it to mention any journey times between sites that in chronogical order were longer than a 30 minute drive, and I asked it to produce a paragraph at the end up each day using the first checklist from the following day as my end destination. Go.

Text formed before my eyes. Reams and reams appeared just like that, scrolling down my screen as the AI chugged through it. Wow! At first glance this seemed like an absolute gamechanger, I could see paragraphs of each site, I could see bird names in bold, I could see that I drove for 32 minutes to reach a certain spot. Blimey. I copied each day into a separate draft blog post and started to look at each one. In my head all I had to do was to find the relevant photos and paste them in. Hurrah! A five day trip report generated in about two hours - the above may sound like five minutes work but actually it took many attemps, with prompting that was quite iterative in nature - I had to zone in on specific changes that I wanted and get it to do that, and then tweak and retweak. Finally, here it was. 

So, was this me? Not by a long shot! It had taken my example blog posts and from these decided on a style, on a tone that it thought that my writing displayed. But once it had that pigeon hole it just used its own algorithm that it thought fitted some key words. Although all the days were equally bad here's an example from Day 3 - it was only at this point that I figured a could get a blog post out of this. This is the unedited output from the machine:

A short drive deeper into Cave Creek Canyon brought me to Cave Creek Ranch, where I spent a generous stretch of time letting the place recalibrate the day. The feeders were alive with movement: Inca Doves shuffling under the tables, Gambel’s Quail materialising from the shadows in small, nervous groups, and Mexican Jays arriving in noisy waves that scattered everything smaller than themselves. Canyon Towhees worked the ground with quiet determination, Acorn Woodpeckers clattered around the trunks, and Bridled Titmice moved through in quick, purposeful flocks. A Curve‑billed Thrasher announced itself from a cholla, and another Blue‑throated Mountain‑gem paused long enough to show off its deep blue gorget. The ranch had that rare quality of feeling both busy and peaceful at the same time — a low, constant hum of bird activity that never quite stopped.

Does that sound like me? I really hope it does not! Recalibrate the day? Small, nervous groups? Clattered around the trunks? It is incredibly forced, there is an overly florid description for almost everything. I mean maybe, I am an Arts graduate after all. But no, this is too much. I asked it to pare down the narrative. Once again text shifted before my eyes. Take Two: 

I reached Portal where the feeders around town offered a brief reset. Inca Doves shuffled under the tables, Gambel’s Quail scurried across the gravel, and Northern Cardinals glowed from the mesquites. A Blue‑throated Mountain‑gem appeared briefly at one of the porch feeders — heavy, deliberate, unmistakable — before drifting back into the sycamores. It was a promising start to the lower canyon.

Better. Less guff. But take a moment to consider the detail that it has come up with. I did not tell it about tables, I did not tell it about gravel, I did not tell it about mesquite and I did not tell it about feeders. This it derived entirely for itself, presumably by searching the internet for anything about Cave Creek Ranch and Southern Arizona. And indeed I'm pretty sure all of these were present. And actually the fact that it aligned the Hummingbird with the feeders and the Gambel's Quail with something at ground level is actually quite impressive. But here's the thing. I didn't see Gambel's Quail here, nor on this day at all, and at this point in the trip hadn't seen it. It made it up.

Interesting. And scary, especially in contexts more serious than bird blogging. I went back to the prompt and told it that it had listed a bird that I had not seen and that was not in the .csv file for that site. I explicitly then told it not to make birds up, to only use birds from the checklist in question. It responded with a cheery Douglas Adams-esque elevator dialogue along the lines of "Right! Yes, I definitely won't make up any birds at all, no Siree, I will only ever mention birds you actually saw. Let me just modify that for you right away Mr Beeblebrox sir!"

To cut a long story it short it did not, and throughout my long editing sessions I continued to find references to birds I had not actually seen. It seemingly cannot help itself, cannot not be forced to exist within very strictly defined parameters. Or at least not with my limited knowledge of how to control it at this point. This is why it has taken so long to complete. Not only was I having to delete superfluous narrative, I was also having to check that it wasn't just making shit up. Luckily my memory of what birds I have seen on trips is quite acute, so it wasn't as if it was a full reconciliation. I found it very easy to pick out the birds I knew I hadn't seen, and so I continued to use my usual method of the blog text on one screen and the eBird checklist on the other not to write it, but simply to verify my suspicions. I was always right.

It has been an interesting experiment. A failed experiment in one sense, but also a success in other ways. I can skip the narrative part, the deleting and editing of which actually took a huge amount of time. But I can easily get it to produce lists of birds, site by site, in chronological order, adding bold text where I want it, carving it out day by day. This is a huge time-saver, cuts out a ton of typing and clicking, albeit I will have to have my wits about me to spot the stringing. I reckon I am equal to it. Before I started this I had almost zero motivation to get this trip report done at all. But mucking about with AI to see what it was capable of was the catalyst that allowed me to complete it, and in doing so I learned what I could use it for and what I couldn't.

Apologies to all the purists out there, but despite the noise I reckon there is some mileage in this. And remember that blogs are free and time is finite. Would you rather the trip report you just read, or nothing? Actually don't answer that! Anyway, I won't be using it to write blog posts. Although did I write this one or did the computer write it? As I said, can you tell? Personally I can, and having gone through this experience I find I am now far more easily able to detect where an online article has in fact been written by AI - they are far far more prevalent than you might assume. Lazy journalism has taken on a whole new meaning. But I think I will be using it to take out a lot of the leg work where a blog post is largely data-driven, which in my case means trip reports. I'm looking forward to Claude Mythos though....

See ya!