Three months of speaking practice with a chatbot instead of a tutor - did anyone's real conversations get better?

I use both and the difference is exactly the thing you have noticed. A model will understand you through a terrible accent, wrong stress and a missing preposition, because it is very good at inference - so you never get the feedback signal that you were unintelligible, which is the signal that actually forces your pronunciation to move. A tutor's face does that in half a second and for free. Keep the nightly practice for fluency and drilling, and buy one human hour a week purely to be misunderstood in.

I did three months of app speaking, felt genuinely confident, then took a placement test at a school and got put a level lower than I expected. The gap was listening under pressure and dealing with people who talk over each other, neither of which I had practised once. Not wasted time, just not the skill I thought I was building.

I counted mine for six months: roughly ninety hours with a model and about twenty two with a tutor, and the tutor hours moved my speaking more per hour by a wide margin. What the model hours were good for was volume and having no fear - I turned up to lessons already used to opening my mouth, which meant we never spent time on nerves. The cost ratio makes the combination sensible rather than either alone.

Partly disagreeing. Most people use these things on the default setting, which is a relentlessly encouraging conversation partner, and then conclude the technology does not work. Tell it explicitly to interrupt you when you make an error, to refuse to guess your meaning, to answer at normal native speed with no simplification, and to end every session with a list of your five most repeated mistakes. It is a different tool with those instructions.

Record five minutes of yourself talking to it, then five minutes of yourself on a real call, and listen to both. That comparison will tell you more than the thread will.