In 2023, a group of researchers at Stanford, Berkeley, and Samaya AI published a paper with a title that has since become a meme in machine learning circles. It was called Lost in the Middle: How Language Models Use Long Contexts.
The finding was simple and uncomfortable. When large language models are given long input documents to work with, they do not read them evenly. They pay attention to the beginning. They pay attention to the end. And they quietly underperform on everything in the middle.
The researchers tested this across multiple tasks and multiple models. They watched performance fall off a cliff when the answer was buried halfway through the document, even when the document was well within the model’s stated context window. The pattern produced what they called a U-shaped performance curve. Accuracy was high at the start, high at the end, and sagging in the middle. The middle was being lost.
This was not a small effect. For some tasks, models were 20 to 30 percentage points less accurate when relevant information sat in the middle of a long document than when the exact same information sat at the boundaries. The phrase Lost in the Middle stuck because it named, with painful clarity, a limitation almost no one had noticed and almost everyone had been quietly paying for.
Here is the part that makes the finding interesting. Lost in the Middle is not really an AI problem.
It is a problem of any system, biological or digital, that has to read something complicated end to end and still understand what was in the middle.
How AI Loses the Middle
When you give a large language model a very long input (a report, a policy document, a multi-chapter brief, a conversation history of thousands of words) the model does not actually treat all of it with equal attention. Its internal attention mechanism puts more weight on what comes first and what comes last.
Researchers have since traced this back to an intrinsic property of how these models were trained. They develop what is called a U-shaped attention bias. The tokens at the start of the input get attention. The tokens at the end get attention. Everything in between gets quietly under-weighted, regardless of how important it actually is.
This matters because the longer the input, the more middle there is. A short prompt has almost no middle. A 75,000-word document has an enormous middle. And it is exactly the long documents (the policy briefs, the country assessments, the consolidated reports, the multi-stakeholder analyses) where the most consequential information often sits buried halfway down, in the section that nobody felt was important enough to put on page one.
AI engineers have responded with new techniques. Position-aware training. Calibration mechanisms. Better retrieval pipelines that pull the relevant passage out of the middle and re-present it at a position the model will actually read. An entire subfield has grown up around making sure the middle does not get lost.
But here is what struck me when I read the research. The same U-shaped curve has been documented in human cognition for more than sixty years. Psychologists call it the serial position effect. When people are asked to remember a list of items, they remember the first few (primacy) and the last few (recency) with high accuracy, and they forget the middle. The shape is the same. The cause is different. The implication is identical.
If a machine reading a long document is biased toward the beginning and the end, and a human reading a long document is biased toward the beginning and the end, then we have a problem larger than either.
How Humans Lose the Middle
Humans get lost in the middle of complicated things all the time. We just rarely name it.
A senior leader reads the executive summary of a long report carefully. She reads the recommendations at the end carefully. The forty pages in the middle that contain the actual evidence, the methodological caveats, the nuanced findings that complicate the headline conclusions, those get skimmed. By the time she sits in the decision meeting two weeks later, what she remembers is the opening framing and the closing recommendation. The middle is gone.
A board member reviews a hundred-page strategy document. She remembers the first chapter, which set up the problem, and the final chapter, which proposed the response. The four chapters in between, which contained the actual analysis that connected the problem to the response, have faded. When she votes on the strategy, she is voting on the bookends. The middle, which is where the reasoning lived, has not made it into the decision.
A team listens to a long presentation. They remember the opening hook and the closing call to action. The substantive section in the middle, where the speaker walked through the evidence, has compressed in memory into a vague impression of “they talked about data for a while.”
These are not failures of intelligence or commitment. These are predictable cognitive patterns. Human attention, like model attention, is U-shaped. We over-weight what we encountered first and last. We under-weight what sat in the middle.
And we make consequential decisions on the basis of that under-weighting all the time.
How Organisations Lose the Middle by Default
Once you understand that individual humans get lost in the middle, the next step is uncomfortable. Organisations, which are made of humans reading documents and making decisions, get lost in the middle by default. The effect compounds at scale.
Consider a development finance institution evaluating a portfolio of proposed investments. Each proposal is hundreds of pages. The investment committee reads the front (the deal summary, the financial structure) and the back (the recommendations, the conditions). The middle, where the country risk assessment lives, where the partner capacity diagnostic sits, where the contextual analysis explains why this design might or might not actually work in this environment, is read in pieces and remembered in fragments.
The investment goes ahead. Eighteen months later, the programme is underperforming. The post-mortem identifies issues that were, in fact, raised in the middle of the original proposal. They were there. Nobody quite remembered them at the moment the decision was made.
Consider an international NGO designing a new five-country programme. The country analysis for each context runs to dozens of pages per country. The design team reads the opening framings and the closing recommendations carefully. The middle sections (the local power dynamics, the institutional history, the trust patterns, the unwritten rules of how things actually move) get absorbed in fragments. The eventual design treats the five countries as more similar than they are, because the differences that mattered were in the middle of each country analysis.
Consider a corporate acquisition. The due diligence document runs to thousands of pages. The board reads the executive summary and the deal recommendation. The cultural assessment, the integration risk analysis, the chapter on the acquired company’s informal power structure, all sit in the middle. Six months into the integration, the issues that surface are the ones the middle had quietly warned about.
In each case, the organisation did not lack information. The information was there, in the middle of the document. The organisation lacked the discipline to read the middle as carefully as it read the beginning and the end. That gap is institutional Lost in the Middle. It is one of the most under-named forms of organisational decision failure.
Reading the Middle Is a Core Capability
This is why I have come to believe that reading the middle is not a nice-to-have skill. It is one of the core capabilities of context intelligence in the era we are entering.
The documents we are being asked to interpret keep getting longer. The contexts we are being asked to operate in keep getting more layered. The decisions we are being asked to make are increasingly based on synthesis across long inputs, multiple sources, and stretched timelines. In that environment, the system (machine, human, or organisational) that can read the middle well has a decisive advantage over the system that cannot.
AI engineers know this. They have built an entire discipline around making sure the middle does not get lost. They retrain models. They calibrate attention. They build retrieval systems that surface the buried passage. They treat the middle as engineering work.
Most organisations have not made the equivalent move. There is no organisational role for middle-reading. No diagnostic that surfaces what the middle of the document said that nobody quite remembered. No discipline that systematically counteracts the cognitive bias that puts the first page and the last page in charge of the decision.
This is exactly what context intelligence is for. And it is exactly the gap the MAPS framework is designed to close.
When you read an operating environment through the four dimensions of MAPS (Mind, Authority, Power, Systems and Social Trust) you are deliberately reading the middle. You are reading the layer of how decisions actually get made, where trust lives, where power is exercised, what cognitive frames people are working from. None of that is in the executive summary of any document. None of it is in the recommendations. All of it is buried in the middle of the operating reality, in the layer between the headline narrative and the formal conclusion.
MAPS is, in a sense, a method for refusing to let the middle be lost. It pulls the contextual layer up to the surface where it can actually shape the decision. It does for organisations and individuals what attention calibration and retrieval pipelines do for AI systems.
The Pattern, One More Time
AI systems get lost in the middle of long documents because of how their attention is structured.
Humans get lost in the middle of long documents because of how memory is structured.
Organisations get lost in the middle of long decisions because they are made of humans reading long documents.
Same U-shaped curve, three scales. AI engineers have built a discipline to compensate for the curve on their side. Context intelligence, with MAPS as its structured form, is the equivalent discipline on the human and organisational side.
If you sit in a role where consequential decisions rest on long, complex inputs (investment committees, programme design boards, executive teams reviewing strategy documents, senior leaders interpreting country analyses, anyone managing a portfolio of complicated commitments) the question I would ask is this. What is in the middle of your most important documents that you are not currently reading? And what is the cost of acting on the bookends alone?
The middle is where the context lives. Reading it is no longer optional.
Catherine Jura Sentamu is the Founder of Edify Learning Forum and the author of The Context Tax: Turning Context into a Measurable and Actionable Intelligence Layer. She is the creator of the MAPS Framework: Mind, Authority, Power, and Systems and Social Trust.
Note on sources: The Lost in the Middle phenomenon was first documented in Liu et al., “Lost in the Middle: How Language Models Use Long Contexts” (arXiv, 2023). Subsequent research has linked it to an intrinsic U-shaped attention bias in transformer models, and to the long-established serial position effect in human memory first described by Murdock (1962).
