Quick answer: when you go over an AI tool's context window, one of three things happens, depending on the tool: it refuses the input with an error, it cuts off part of what you pasted, or it drops or compresses earlier parts of the conversation. The last two are the risky ones, because the AI may answer confidently without the information it lost.
What the context window covers
Everything the model can consider at once: your current message, earlier messages, any pasted documents or uploaded files, system instructions, and the reply it's writing. All of it is measured in tokens, roughly 1.3 per English word.
1. A clear error. Developer APIs and some apps reject input that's too long with a message saying so. Annoying, but at least you know.
2. Truncation. Some tools quietly read only the first part of a long paste or file. You ask about chapter 9 and the AI only saw chapters 1 to 6.
3. Forgetting the start of a conversation. In long chats, some tools drop or summarize older messages to make room. Instructions you gave at the beginning stop being followed.
Signs you've hit the limit
- The AI says something isn't in the document when you know it is
- It stops following instructions or a format you set earlier
- Answers about the end of a document are vague or wrong
- It contradicts details from early in the conversation
- The app shows a message about length, or offers to start a new chat
How to fix it
Check the size first. Paste the text into the context window calculator to see its estimated tokens and what share of a given window it uses.
Send less. Paste only the relevant sections. Remove boilerplate like headers, footers, tables of contents and repeated legal text.
Clean the text. Transcripts with timestamps, or text full of formatting, waste tokens. The transcript cleaner and invisible character remover help.
Split and summarize. Process a long document in sections, ask for a summary of each, then work from the summaries.
Start a fresh conversation. Paste a short summary of what matters plus your instructions, rather than continuing a very long thread.
Restate key instructions. In long chats, repeat important rules near your latest message.
Bigger windows help, but don't solve everything
Larger context windows make room for whole books and codebases, but models can still pay less attention to details buried in very long inputs. Even when everything fits, focused input usually gets better answers.
For how many pages fit in common window sizes, see how many pages can ChatGPT, Claude and Gemini read?
A worked example
You paste a 70,000-word contract bundle into a chat with a 128,000-token window and ask about the termination clause near the end.
- The bundle is about 93,000 tokens.
- Your earlier conversation and instructions add about 10,000.
- The reply needs room too, say 4,000.
That's around 107,000 of 128,000 tokens, which fits, but leaves little room for follow-up questions. Three more long exchanges could push the chat over the limit, and the tool may begin dropping the start of the conversation, which here is the contract itself.
Better approach: paste only the termination and related definitions sections, about 6,000 words, in a fresh chat. The answer is faster and more focused.
Estimating before you paste
Use the context window calculator with:
- Your text or its word count
- The window size of your tool
- A reserve for the reply and follow-ups, often 10% to 25% for long conversations
If the result shows more than about 70% used before you've asked anything, plan to split the work.
Some apps keep long chats going by compressing older messages into a summary. You might notice:
- The AI remembers the gist of early messages but not exact details or numbers
- Quoted text from earlier in the chat comes back slightly different
- It asks you to repeat information you already gave
When accuracy matters, paste the important material again or start a fresh chat with a short brief.
Writing a good fresh-start brief
When you move to a new chat, include:
- The goal: what you're trying to produce
- Key facts and decisions so far, as a short list
- Rules and format you want followed
- The material for the next step only
A 300-word brief often works better than 50,000 words of old conversation.