You upload a 40-page lease, a stack of medical records, or a loan agreement, and ask an AI chatbot to summarize it, flag the risky clauses, or answer a question buried somewhere inside. It replies instantly, in clean, confident prose. But can AI read long documents as thoroughly as it sounds like it can? Research into how these models actually behave over long text says: not evenly. There’s a well-documented blind spot right in the middle of whatever you feed it — exactly where the clause you most need to catch tends to live.
This isn’t a reason to stop using AI on long documents; it’s genuinely useful for a first pass. It’s a reason to know where it’s weakest, so you know when to stop trusting the summary and go read the page yourself. No fear-mongering here — just the finding, why it matters most for contracts and medical and financial paperwork, and the concrete habits that keep a human in charge of anything consequential.
What “Reading” a Long Document Actually Means for AI
When you paste a long document into a chatbot, it gets broken into tokens — small chunks of text — and loaded into the model’s context window, the amount of text it can hold “in mind” at once. Context windows have grown enormously: what used to be a few thousand tokens is now routinely hundreds of thousands, with some vendors advertising windows in the millions. That growth is a genuine engineering achievement, because attention gets more expensive the longer the input gets — doubling the text roughly quadruples the compute cost, so fitting a whole novel or case file into one prompt at all is a hard-won result, not a given.
But here’s the part that doesn’t make it into the marketing: being inside the context window is not the same as being read with equal care. The model technically has access to every token you gave it. Whether it uses every part of that text with equal reliability is a separate question — and the honest answer, backed by research, is no. (For the “how big can the window get” side of this story, see our look at what a million-token context window actually gets you.)
The “Lost in the Middle” Effect
The clearest evidence for this comes from a widely cited 2023 study, “Lost in the Middle” (Liu et al., published in the journal Transactions of the Association for Computational Linguistics, arXiv 2307.03172). The researchers ran a simple, clever test: take a long stretch of text, hide the answer to a question somewhere inside it, and move that answer’s position around — sometimes near the start, sometimes dead center, sometimes near the end — then measure how often the model finds it.
The result was a U-shaped curve. Accuracy was highest when the answer sat near the beginning or the end of the context, and it dropped — sometimes substantially — when the answer was buried in the middle, even though nothing about the difficulty of the text itself had changed. The model wasn’t running out of room or truncating anything; it had the whole passage. It simply used the start and the end far more reliably than everything in between. And this pattern showed up even in models specifically built and marketed for long-context use — a bigger window didn’t make the effect go away.
Think of a student cramming a chapter the night before an exam: the opening hook and closing summary stay sharp, and everything sandwiched in between blurs. AI models seem to develop a version of the same habit — except unlike the student, they don’t announce when they’re guessing.
Why the Middle Matters Most in a Contract, Medical File, or Loan Agreement
This would be a minor curiosity if the important part of a document were reliably at the start or the end. It usually isn’t. The first page of a contract is typically boilerplate — parties, definitions, recitals. The last page is signatures. The clause that actually changes your obligations — an auto-renewal term, an arbitration requirement, a liability waiver, a rate-adjustment trigger — is routinely somewhere in the middle third of the document, exactly the zone the research says gets the least reliable treatment.
The same shape shows up outside contracts. A drug-interaction warning can sit mid-page in a multi-page medical report. A coverage exclusion can sit mid-section in an insurance policy. A hidden fee schedule can sit mid-document in a loan agreement. None of these are edge cases you have to go looking for — they’re the ordinary structure of long-form paperwork, which tends to open and close with the routine parts and bury the consequential parts in between. That’s precisely the layout where “the AI summarized it for me” is least trustworthy and the stakes of missing something are highest.
Try It Yourself: A Two-Minute Trust Test
You don’t have to take this on faith. Take a document you already know well — an old lease, a syllabus, a long email thread — and ask the AI two questions: one whose answer sits near the beginning or end, and one whose answer sits in the middle third. Many people who try this see exactly the pattern the research describes: the edge question gets answered crisply, and the middle one comes back vague, generic, or just wrong — often in the same confident tone as the correct answer. That tone is the trap; nothing about how an AI phrases a wrong answer reliably signals that it’s wrong, which is exactly why a spot-check habit beats a gut feeling about how sure it sounds.
Four Habits That Keep a Human in the Loop
None of this means you should avoid AI for long documents — it means you should use it the way you’d use a very fast, occasionally unreliable research assistant, not a final reviewer. A few concrete habits do most of the work.
- Spot-check the middle, not just the summary. Don’t just skim the AI’s output and move on — open the document to a page or section in the middle third and confirm the summary actually reflects what’s there. That’s the zone most likely to be glossed over.
- Ask for a direct quote, not just a paraphrase. Instead of “summarize this contract,” ask it to “quote the exact sentence about early termination, and say what page or section it’s on.” A quote is checkable against the source; if the AI can’t produce an exact one, or invents one that isn’t actually there, that’s a hallucination warning sign.
- Break long documents into chunks yourself. For anything genuinely long and important, split it into a few sections and ask about each one separately, instead of one pass over the whole thing. Every chunk you ask about is now near the start or end of its own smaller window.
- Keep a human on anything consequential. AI-only review is fine for a quick summary. For a document you’re about to sign, a medical decision, or a financial commitment, treat the AI’s read as a first pass a person — you, a lawyer, a doctor, an advisor — still reviews before anything is finalized.
A Bigger Context Window Doesn’t Fix This
It’s tempting to assume the next model with an even larger advertised context window will simply solve the problem. That’s not really how it works. Much of the engineering effort behind giant context windows — techniques like Ring Attention, which shards a long sequence’s attention computation across many chips, or serving-side systems that manage how a model’s memory of past text is stored — is about making it possible and affordable to hold that much text in front of the model at all. That’s the story covered in how AI makes attention efficient: a real, hard problem, and solving it is a genuine achievement.
But holding more text and reasoning evenly across all of it are two different problems, and progress on the first doesn’t automatically deliver the second. “Lost in the middle” showed up in models already built for long context — so a bigger number on a spec sheet isn’t, by itself, a promise that nothing important gets glossed over. Treat a large context window like a large library: a book being on the shelf doesn’t mean it’s been read carefully.
The Bottom Line
AI is genuinely good at digesting long documents faster than any human could — and genuinely uneven about which parts of that document it treats with real care. The “lost in the middle” finding gives you a specific, testable reason for that unevenness: models reliably favor the start and end of long text and lose reliability toward the middle, which is usually exactly where the clause that matters is hiding. The fix isn’t to distrust AI wholesale; it’s to spot-check the middle, demand quotes you can verify instead of paraphrases you have to take on faith, chunk genuinely important documents yourself, and keep a human making the final call on anything consequential. Do that, and you get the speed of AI-assisted document review without inheriting its blind spot.
❓ Frequently Asked Questions
Can AI read long documents accurately? It can process the whole thing — every part of a document you paste in is technically inside the model’s context window — but research shows accuracy isn’t even across that window. Models are demonstrably more reliable at using information from the start and end of a long document than information buried in the middle, a pattern known as “lost in the middle.”
What is the “lost in the middle” problem? A finding from a 2023 study (Liu et al., arXiv 2307.03172) showing a U-shaped accuracy curve: AI models retrieve information best when it’s near the beginning or end of a long context, and noticeably worse when it’s buried in the middle — even though the material’s difficulty hasn’t changed.
Does a bigger context window fix this problem? Not by itself. A larger window mostly solves a different problem — making it technically possible and affordable to hold more text in front of the model. “Lost in the middle” has been observed even in models built for long-context use, so a bigger advertised window isn’t the same as even, reliable reasoning across all of it.
How can I tell if an AI’s summary of a long document is missing something? Spot-check a section from the middle third against the summary rather than trusting the overall gist, and ask the AI to quote the exact sentence a claim is based on — if it can’t produce a verifiable quote, or the quote doesn’t match the source, treat that part with suspicion.
Should I use AI to review a contract or medical document at all? Yes, as a fast first pass — it’s genuinely useful for spotting issues and organizing a long document quickly. But for anything consequential (something you’re about to sign, a medical decision, a financial commitment), keep a qualified human reviewing the parts that matter before you act on it.
📚 Further Reading
- How AI Makes Attention Efficient — the engineering playbook behind big context windows, and why “bigger” and “reliable” are different problems.
- 1M-Token Context Windows: What the Number Doesn’t Tell You — a closer look at what a headline context-window number does and doesn’t guarantee.
- Why Long Context Is Expensive — the O(n²) math that makes large context windows a genuine engineering achievement.
- What Not to Share With AI: A Chatbot Safety Checklist — the broader habits for using AI tools safely with sensitive documents.
📚 Recommended Resources
Disclosure: some links below are Amazon affiliate links. If you buy through them, I earn a small commission — at no extra cost to you — which helps keep this site running.
Artificial Intelligence: A Guide for Thinking Humans
Melanie Mitchell · Book · ~$28 on Amazon
The right companion for this article’s core lesson — how to separate what AI genuinely does well from what it only sounds confident about. Mitchell is superb on why fluent, plausible output earns more trust than it’s actually owed, which is exactly the trap “lost in the middle” sets.
Build a Large Language Model (From Scratch)
Sebastian Raschka · Book · ~$50 on Amazon
For readers who want to understand why a model’s attention behaves this way — built from the ground up, hands-on, covering the attention mechanism and context handling that sit underneath every long-document summary an AI gives you.
Natural Language Processing with Transformers
Lewis Tunstall, Leandro von Werra, Thomas Wolf · Book · ~$60 on Amazon
A practitioner-level guide to how Transformer models process real text end to end — useful if you work with AI document tools regularly and want a deeper grip on their capabilities and limits, not just the marketing.