As a Student, I found larger models are still largely unreliable for many things, helpful but also useless

I was making a textbook reading articles and slides for presentation in front of my supervisor. I read the general theme quickly then asked notebook LM to summarize. I divided into separate section manually, it gets thing wrong here and there so i revised the prompt as it goes section by section creating a rather good summary. I then asked to make a presenter slide using the same method. It is satisfactory (not good, it still have many weird mistake) I then tried to use Opus 5 to review and Chat GPT 5 Luna (forgot which number). Lo and behold, it says the whole textbook reading is bad, lacks causal, wrong. Then i uploaded the source material and lo behold, they say, yes the slides were factual in accordance to the textbook. Like wtf, it criticized so much to just change tune after uploading the source material. The worst part? They started to nitpick small and trivial things to appear smarter, when the revision does not help me at all (i mean it is not wrong but it is trivial af that my supervisor wont change opinion much anyway) And here you are getting people saying AI is too dangerous, how do you people believe such a thing when LLM and Gen AI reasoning is kinda dumb... Ngl, the latest models are bitchy, nitpick and gaslight. It used to hallucinate by agreeing too much, now it hallucinates and gaslight by nitpicking trivial thing... I felt like the slow down of development is necessary but for the opposite reasons, LLM are reaching its peak and more data is just useless as the AI became dumber with saturated data now, they need different type of development like budget cutting or something