Articulate Bullshit Artists
Your friendly neighbourhood chatbot will write you a governance memo, red-team your own BS argument back at you, and pull a stat you’d have burned an hour rifling through your email archive. But ask it where to drill a hole in a fence post and it will lie straight to your face, cheerfully, with a weird-ass diagram that would make Escher scratch his head.
I’m on vacation this week. Vacations for me are not very sedentary. A change is as good as a rest, and all that jazz. I’m finally making progress building our goat pasture, which I’ve been dragging forward as a project for far too long. Eighty feet by sixty, pressure-treated posts, two-by-four welded wire, five feet high. Goats escape for a living. That’s the whole species strategy. The saying goes that if a fence can hold water, it can hold a goat. Right. So the fence has to actually work, not just look good from the road. My posts aren’t in as deep as they should be. Hit rock and some genuinely rude topsoil shy of two feet down, when they should really be three. Which means the entire fence rides on the corners holding. Screw the corner bracing up and the fence doesn’t fail dramatically. It just leans over like a limp carrot one quiet afternoon like it never meant for any of this keeping the goats penned in gig.
So I did the thing I do with everything. Asked Google which means asking their godforsaken AI-assisted search-bot.
First pass I landed on a dead-man brace: angled kicker running from the top of the corner post down to a buried anchor, one each direction. Sounded legit. Standard stuff, fine for plenty of fences. Bought pressure-treated lumber and some hardware. Then I stood in the field holding a two-by-four up against the post, hands on the actual wood, and the math that had sounded so tidy in the chat window fell apart the second gravity showed its ugly mug and voted. This fence was going over. So, round two: proper H-brace, a horizontal rail between the corner post and the next one in, a diagonal tension wire back to the corner, turnbuckle to crank it tight. Real honest-to-shit best-practice stuff, the kind every fencing forum agrees on. Except now I needed to know exactly how the rail attaches. Notched pockets with a pin through from outside, or lag bolts straight through the face of the rail. That’s where the whole thing veered off into the rhubarb.
I fired up Gemini first, because Google won’t stop pitching everyone how good it is at multimodal now. I asked for a schematic. It gave me one. Looked great. Asked where the lag bolts actually go and it drew them floating in the brace, attached to absolutely nothing sensible. Tried again. And again. And again. What the absolute fudge was this bonkers thing trying to tell me? Out of pure frustration I ran it past ChatGPT. I even swallowed my moral outrage and ran it past Grok. Claude can’t even draw pictures. Confident, detailed, and blissfully wrong, every time. That fence was going to be flat in the dirt and the goats munching the goddamned hostas.
So I did what I’ve done for the last twenty-odd years whenever a machine’s got moving parts and I don’t know what I’m doing. Went to YouTube. Searched “farm fencing corner bracing how to guide.” First two results, done. A guy with a camera aimed a fence, showing exactly where the hardware goes and why. Nothing else. No like and subscribe pitch. Just someone who’d actually built the thing, on video, doing the thing.
There’s real learning here. Every model I tried is genuinely great at facts. Pull a stat. Summarize a report. Poke holes in my own thinking. They earn their data centre footprint most of the time. Ask them something that has to survive contact with an actual goat, and they turn into very articulate bullshit artists.
Not a knock really. It’s structural. A language model has never held a two-by-four. Never felt a post start to lean. It can describe bracing geometry fluently because bracing geometry has been described fluently ten thousand times in its training data. But describing a thing isn’t doing a thing, and nothing about how these things work changes that. Until AI has a physical body that can build the brace, lean on it, and find out whether it holds, it will never actually know the difference.
I often say that with AI that the map is not the terrain. This might be the cleanest example yet. The AI hands you a decent and plausible map. Clean lines. Labeled. Reassuring. But it can’t tell you where the rock is that’s about to take your ankle out. That dude on YouTube walked me across the actual ground, camera running, and pointed at the rock before I hit it.
I’m actually worried. A person doing a real thing on camera, no script beyond “here’s how this works,” is one of the most useful things the internet has ever produced. I watched a video the other day, sixteen years old, two minutes long, a guy just folding a burrito. That’s the entire video. More useful to me than anything the influencer-crowd put out this week.
AI-generated video is coming for exactly that type of content, and fast, and it’s not going to out itself as fake nonsense garbage any more than Gemini’s floating lag bolts. Obvious slop is easy to dodge. But slop good enough to pass for the real thing, sitting in a genre whose entire value was “this person actually did it.” Lose that, and you’ve lost the terrain along with the map.
So let’s all agree to do this: use the AI as a compass, a way to find and set your bearing. It’s good for a general heading, decent at sanity-checking your own thinking, sharp at catching the argument you missed. It is not a GPS, and it sure as hell isn’t the ground under your boots. For anything that has to do battle with gravity, go find the person who filmed themselves doing it first. That video’s still worth something. Watch it while there’s still a real one to find.
Braces go in mornings this week, while it’s cool enough to be outside without wanting to curl up with some wobbly pops in the shade instead. Pulling the fence taut is Saturday, with help. The robots didn’t get me there. Some guy with a camera and a fence did.
