An owner rang me a few weeks back about her ten week old puppy. She'd been asking an AI chatbot for help every night, following it properly, and getting nowhere. Then she said the line that stuck with me. "I don't want to raise a ChatGPT dog."
I knew what she meant straight away. The advice she'd been given was fine. It was advice for a puppy. Hers is a specific puppy, and that turned out to be the whole problem.
I'm Luke Buchanan, Owner of The Toe Beans Co and Sydney's Puppy Trainer, and I spend most of my working week in people's homes watching what actually happens. The thing that fails almost never fails because the technique was wrong. It fails because the technique was right for a different dog.
Where the phrase came from
"ChatGPT dog" isn't something trainers say. It came from an owner, which is exactly why it's a good phrase. What she'd worked out was that her puppy was being trained by a source that had never seen him.
AI is genuinely useful for this. So are forums, books, and the better end of YouTube. If you want to know what a crate is for, or roughly when a second vaccination lands, you'll get a clean answer in ten seconds and it'll be correct. I use these tools myself. We built one, which I'll come back to.
What none of them can do is watch your dog. And most puppy problems aren't knowledge problems. They're observation problems.
Three ways generic advice goes wrong
These are patterns I see over and over, with details changed so nobody's identifiable.
The Groodle genetic lottery, same breed, opposite behaviour
Crossbreeds are where general advice comes apart fastest, and Groodles are the clearest example. Our own breed notes say it plainly: size and traits are unpredictable, especially in first generation crosses.
I've seen littermates go to two homes a few streets apart, both owners doing the same sensible things, and one puppy settles into a crate inside a fortnight while the other is still waking the neighbours at 4am six weeks later. Same breeder, same crate, same advice. The difference is arousal baseline, and there's no breed page in the world that can tell you where your dog sits on it.
A breed guide describes tendencies across a population. It was never written about your dog, and with a crossbreed the gap between those two things is at its widest.
The food-motivated technique that backfired on a smart dog
A very food driven dog looks like the easy one. It usually isn't.
The standard advice is simple enough. Reward what you want with food. Correct, as far as it goes. But hand that to a clever dog with a big appetite and something else happens. The dog works out that food arrives when it does something, so it starts offering behaviour at you constantly to make food appear. Nudging. A sit you never asked for, then another one, faster and closer.
What's happened there is a reversal. The dog is now conditioning the handler. I've watched an owner get up off the sofa to fetch a treat because their dog sat down and stared at them, and both of them thought this was training.
The other half of it is arousal. Visible food pushes a food driven dog over threshold quickly, and above threshold very little goes in. Owners read that as stubbornness and raise the stakes with better food, which raises arousal again. In some dogs that loop ends somewhere worse, with a dog that starts treating food as something to compete for and begins guarding the bowl.
What we do instead is put food back where it belongs. Eat something yourself before you feed the dog. Ask for a sit, a wait, then release. The bowl goes down when the dog is genuinely calm, never as a bribe to get it there. Pick the bowl up the second the dog walks away from it. And don't chase your dog's mood by swapping foods, because that teaches the dog it controls what goes in the bowl.
Food is one tool inside a leadership structure. It was never meant to be the structure.
The high-pitched recall cue a herding puppy read as a chase invitation
Nearly every recall guide tells you to sound exciting. Happy voice, high pitch, big energy, be the best thing in the park.
Try that with a young kelpie or a border collie and watch what happens. A lot of them don't come in. They arc. They circle wide, drop into a crouch, get faster, and the owner sounds more excited, which makes it worse.
The dog is answering a different cue to the one you think you gave. High pitch and fast movement is chase language, and to a dog built to move things, that reads as an invitation. It's the same mechanism behind why chasing a dog that's stolen a sock never gets the sock back. Your excitement is the reward.
The version that works is duller than people expect, and you can read the full method in our guide to recall training that actually works. Call once, level and flat, then stop talking. Repeating the cue teaches your dog that answering is optional. A whistle beats a voice for most dogs because the tone never leaks how you're feeling. Long line on a harness, never a collar. Step on the line so the distance can't grow, take the collar gently when the dog arrives, reward, then release the dog straight back to whatever it was doing. That release is the part everyone skips, and it's the part that stops recall meaning the fun is over.
One more, and this is the one that breaks recall permanently. Never call your dog to do something it dislikes. Not for a bath, not for nail clipping, not for the end of the park.
Why AI, forums and generic guides can't diagnose
A general source can only answer the question you asked it. That sounds obvious. It's the entire thing.
When I walk into a house for the first time, I'm not really listening to what the owner says the problem is. I'm reading what sits around it.
- Where the dog actually is on the arousal scale right now. Most owners think their dog is at a 3. It's usually a 6, and above a 5 nothing meaningful gets learned.
- What the dog believes its job is. Two dogs barking at the same door can be two completely different problems with two different fixes.
- What's already been reinforced by accident. This is the big one, and it's invisible from the inside, because it's built out of a hundred small moments nobody thought were training.
- Who's home, who feeds the dog, and who caves at 6am.
- Whether there's a second dog, which changes the feeding plan completely.
- Age and stage. A twelve week old puppy and a nine month old adolescent get different answers to the same question.
None of that is knowledge a chatbot lacks. It's information nobody has given it, because nobody in the house can see it either.
There's a second problem underneath that one, and it's the one nobody talks about. A general model has read everything ever written about dog training. Roughly half of that is correction based. A decent chunk of it is twenty years out of date, because the dog training industry runs about twenty years behind human behavioural science and the internet has kept every version of it. When a model averages all of that, you get the middle of the pile. Sometimes the middle of the pile is a prong collar.
When generic advice is fine, and when it isn't
Generic advice works for facts and fails for anything that depends on your dog's current state. I'd rather be honest about that than pretend every question needs a trainer.
Generic is fine for facts. What a long line attaches to. When vaccinations are due. What the socialisation window is and when it shuts, which is 6 months, not the 16 weeks you'll still see repeated across half the guides online. Ask a chatbot what to expose a puppy to before 6 months and you'll get a decent list, because that list barely changes between dogs.
It falls over the moment the answer depends on your dog. Anything to do with arousal, fear, guarding, reactivity, or a pattern that's already been running for weeks. A general source can tell you to reward calm. It can't tell you your dog hasn't been calm once in the last fortnight and you've been rewarding something else entirely.
So use the tools. They're good at facts. But the second the answer has to fit one particular dog, a source that has read everything is the wrong source, and that part is actually fixable. You narrow what it reads.
So we built one that only reads our method
Delphi is a chatbot inside our community that answers using The Toe Beans Co's own training material and nothing else, so what it tells you matches the course you're already following.
That's the whole idea, and it's a smaller idea than it sounds. The open internet isn't in there. It reads one library: The Big 8, the Weekly Journey, the Problem Library, written in the same words your trainer uses in your kitchen.
Here's what that changes when you're actually sitting there at 11pm with a puppy screaming in a crate.
- Ask it about crate crying and you get the sequence from The Big 8, the same one you were shown in week one.
- No prong collars. No e-collars. No correction-based fix at all, because none of that exists in what it reads. A general model will hand you all three without blinking, and it isn't doing anything wrong, it's just reporting what most of the internet says.
- The material it answers from is the material I work from. So when you message us the next morning, nobody has to undo anything first.
- And it's awake at odd hours, which matters more than it sounds, because the questions that genuinely derail people arrive at 11pm on a Tuesday rather than at 2pm on a working day.
That last one is the reason it exists at all. The questions that break people aren't the big ones. They're small and urgent and they happen when nobody's available, and by the time you can ask a person on Thursday you've already spent three nights doing the wrong thing consistently enough for the dog to have learned it.
It's new, and we're still feeding it. If it doesn't know something yet, it says so rather than filling the gap.
What the bot is good for, and what still needs a person
Delphi is a support tool, not a replacement for a trainer who can see your dog. That distinction matters, so I'd rather draw it now than let you find it out at the wrong moment.
Where it earns its place: a fast answer at an odd hour, checking you've remembered a sequence correctly, and the small question that isn't worth a phone call but is stopping you sleeping. Consistency, mostly. It'll tell you the same thing the course told you, which sounds unremarkable until you've spent a fortnight collecting four different answers from four different websites and lost track of which one you're currently following.
Where it stops: anything where somebody needs to look at the dog. Delphi can't see your kitchen. It can't watch your puppy's body language shift in the three seconds before the bite. And it can't tell you that your dog is sitting at a 6 when you're certain it's at a 3, because you're the one typing, and you're the one who thinks it's a 3.
Aggression, real fear, guarding, or a pattern that's been running for months all go to a person. It won't diagnose anything and we've built it not to try.
That's the same reason every Complete Puppy Program starts with a session in your home rather than a video call or a questionnaire. Your kitchen, your hallway, your dog. I watch a normal hour and we build the plan off what's in front of us, which is usually not the thing you emailed about. Then comes group puppy school, which does the thing a home session can't, because a puppy that sits beautifully in your kitchen and forgets its own name the moment another puppy walks past hasn't learned much yet. The bot sits alongside all of that, filling the gaps between sessions. It was never meant to replace the room.
Want help putting this into practice?
The Toe Beans Co runs a free SKOOL community where Sydney dog owners get training guides, Q&As, and direct support from Luke. It's free to join.
If you'd rather someone actually looked at your dog, the Complete Puppy Program starts in your home and runs puppy school across Paddington, Zetland, Marrickville, Alexandria and Neutral Bay. Start with a free meet and greet and we'll tell you honestly whether you need us.
Frequently asked questions
Q: Can I use ChatGPT for puppy training advice at all?
A: Yes, and it's good at facts. Vaccination timings, what equipment does what, what to expose a puppy to before 6 months. Where it struggles is anything that depends on your dog's current state, because it can't see your dog and neither can the person typing the question.
Q: Isn't dog training basically the same for every dog?
A: The principles are. The plan isn't. Two puppies from the same litter can need genuinely different starting points depending on arousal baseline, what's already been reinforced at home, and how many people live in the house.
Q: Is your training bot just ChatGPT with a different name on it?
A: No. Delphi only reads The Toe Beans Co's own training material, so it answers from The Big 8, the Weekly Journey and the Problem Library rather than from the open internet. That's why it won't suggest a correction-based fix, and why it won't contradict what you were taught in the course.
Q: Can a chatbot replace a dog trainer?
A: No, and ours isn't built to. A bot is good for fast answers at odd hours and for keeping you consistent with what you've already been taught. Anything that needs eyes on the dog, including aggression, fear and guarding, still needs a person in the room.
