Why your AI does that
Pick the thing that annoyed you. Every page shows the mechanism on your own text — no screenshots of someone else's chat, no hand-waving.
Memory and context
Why a chat drops what you said, and what "context" actually means.
Why AI misses things in the middle of a long document
Where a fact sits inside a long input changes how likely the model is to use it. In the published measurements below, accuracy is highest when the answer is at the very start or the very end and lowest in the middle — the same question, the same amount of text, only the position moved. The middle is not skipped on purpose; it is simply where retrieval is weakest.
Why AI stops following my instructions in a long chat
Your instruction is one small block inside a much larger request. With a vendor system prompt in front of it and a thousand tokens of conversation behind it, a one-line rule is about 0.2% of what the model reads. Measured performance on long inputs also falls as that request grows, so the same rule holds worse at turn forty than at turn two.
Why ChatGPT forgets what I said earlier in the chat
A chat has a fixed size, counted in tokens, and that size has to hold the whole conversation plus the reply. When the conversation outgrows it, the oldest messages simply stop being sent — the model is not forgetting them, it is not being shown them. Nothing you say inside the chat makes the box bigger.
Why my document is too long for the chat
Length limits are counted in tokens, not pages or words, and each model has its own window. The same text is 66 tokens for GPT-4o and about 122 for Claude Opus 5, because the two split it differently. If the input on its own is bigger than the window, the request is refused outright rather than trimmed — which is why the refusal arrives before anything is read.
Accuracy and made-up answers
Why confident answers can be wrong, and where the confidence comes from.
Why AI can't count the letters in a word
Your text is cut into tokens before the model reads it, and a token is usually a chunk of several letters — inside a normal sentence, strawberry is one single token. Counting letters means counting inside something the model was never shown, so it answers from the shape of similar text instead of by looking. Separating the letters is what actually fixes it.
Why AI does not know about recent events
Training ends on a date, and after that the model is a fixed thing that learns nothing from use. Anything later reaches it only if you or the app puts it into the request. The published dates also mean less than they look: Anthropic gives two different cutoffs per model, and Google describes one model's knowledge as reliable to March 2026 in some areas and only to January 2025 in others.
Why AI does not know what today's date is
There is no clock inside a model. A date reaches it only if the surrounding application writes one into the request, and Anthropic documents doing exactly that in its chat apps — while stating that the same system prompt does not apply to its API. So the same model can know today's date in one place and be a year adrift in another.
Why AI invents sources and citations
A citation is written the same way as everything else: one word at a time, each chosen because it fits what came before. A plausible author, a plausible year and a plausible journal assemble into a reference that has never existed, and nothing in that process ever checks. Research from OpenAI attributes the guessing to how models are trained and graded, not to a defect in one model.
Unpredictable answers
Why the same question does not give the same answer twice.
Quality and degradation
Why answers get cut off or get worse the longer you go.
Why AI answers get worse the longer the chat gets
This is not the chat running out of room. Measured on inputs that fit comfortably, accuracy falls as the input grows — one study found drops of 13.9% to 85% even when the model could retrieve every relevant fact perfectly. Length alone costs quality, which is why a fresh chat with a short written brief often outperforms the long one you have been building all day.
Why the AI answer stops in the middle
A reply has its own budget, separate from how much the model can read. GPT-4 Turbo produces at most 4,096 tokens per reply while reading 128,000; Claude Haiku 4.5 produces 64,000 and reads 200,000. Running out of reply budget stops the text mid-sentence, and that smaller limit is usually the one you hit. Saying continue picks it up from where it stopped.
The recurring ideas behind these answers live in Learn.