Your AI just got dumber — and you fed it too much

Why the AI copilots enterprises are betting billions on get quietly worse the longer, harder, and more usefully you use them — and why almost no one is measuring it.

Context rot: the silent bug in every AI rollout

Picture this: your AI copilot nails a simple question in seconds. Two hours into a complex compliance review — the exact moment you need it most — it starts missing things a first-year analyst wouldn't. Nothing crashed. No error message appeared. The model simply got worse at its job, in real time, without telling anyone.

That is context rot, and if your organisation is scaling AI copilots, agents, or assistants, it is almost certainly happening inside your workflows right now.

The problem: bigger context windows were supposed to fix everything

For two years, the AI industry sold context window size as the headline metric — 32K tokens, then 200K, then a million. The pitch was simple: paste in the whole contract, the whole codebase, the whole case file, and the model will reason over all of it. Enterprises believed it and built accordingly, dumping entire document sets into single prompts.

The evidence says otherwise. In a widely cited 2023 study, Liu et al. (Stanford, UC Berkeley, and Meta AI) tested open and closed models on multi-document question answering and key-value retrieval tasks. They found that accuracy on retrieval tasks fell by more than 30 percentage points when relevant information moved from the start or end of the context to the middle — a U-shaped degradation curve that held across every model they evaluated, including those explicitly fine-tuned for long-context use.

More recently, Chroma Research's 2025 benchmark of 18 frontier models — including GPT-4.1, Claude 4, and Gemini 2.5 — found that all 18 showed measurable degradation as input length grew, even on relatively simple retrieval tasks and even when the context window was nowhere near full. Adobe Research's NoLiMa benchmark (Modarressi et al., ICML 2025) pushed further: when questions and target information shared minimal literal overlap — forcing the model to infer rather than pattern-match — the majority of 13 models advertising 128K-token windows fell below half their short-context accuracy by just 32,000 tokens. Even GPT-4o, one of the stronger performers, dropped from 99.3% to 69.7%.

The information wasn't missing. The model simply stopped paying reliable attention to it.

Context rot: accuracy falls as input grows

Even when the answer is present and easy to find

Context rot: accuracy falls as input growsIllustrative line chart. Task accuracy falls from about 96% at a few hundred tokens to about 41% at 128,000 tokens of context.304050607080901001001K10K100KSame task.More tokens.Worse answer.200 tokens: 96% accuracy1K tokens: 91% accuracy4K tokens: 84% accuracy8K tokens: 76% accuracy16K tokens: 68% accuracy32K tokens: 58% accuracy64K tokens: 49% accuracy128K tokens: 41% accuracyContext length (tokens, log scale)Task accuracy (%)

Illustrative curve based on published long-context benchmarks (Chroma Research 2025; Liu et al. 2023; Modarressi et al., ICML 2025).

View the data
Context length (tokens)Task accuracy
20096%
1K91%
4K84%
8K76%
16K68%
32K58%
64K49%
128K41%

Why there is no simple fix

Context rot is not a conventional software bug you can patch overnight. It emerges from well-documented architectural tendencies in how transformer models allocate attention. Three documented failure modes compound each other:

  • Attention dilution. Every additional token competes for a fixed, softmax-normalised pool of attention. Bansal et al. (Meta, Harvard, and OpenAI researchers, 2025) formalise this as “score dilution” and prove that, as context length grows, the attention mass assigned to a relevant token vanishes unless its distinguishing signal grows at least logarithmically with the number of competing tokens. In plain terms: the more you add, the harder the signal has to fight to be heard — and past a point, it can't.
  • Positional bias. Models attend far more reliably to the start and end of a context than the middle. Liu et al.'s “lost-in-the-middle” U-curve has been replicated across independent studies, including follow-up work by Xiao et al. (ICLR 2024) on attention sinks, and calibration research by Hsieh et al. (“Found in the Middle,” ACL Findings 2024) that recovered up to 15 accuracy points by explicitly correcting for positional attention bias. Even models built for long context retain this bias.
  • Noise accumulation. In multi-step agent workflows, every file read, search result, and exploratory dead end stays in the window, competing with the answer that actually matters. Chroma's 2025 study found that semantically similar distractors caused degradation beyond what length alone predicted — and that models sometimes performed worse on logically coherent documents than on shuffled text, because coherent structure creates more plausible-seeming false leads.

The ‘lost in the middle’ effect

Where a fact sits matters as much as whether it’s there

The lost-in-the-middle effectIllustrative U-shaped curve. Retrieval accuracy is about 90% for facts at the start or end of a document and about 60% for facts in the middle.556065707580859095100Startof documentMiddleof documentEndof documentHigh recallHigh recallBlind spotStart of document: about 90% retrieval accuracyMiddle of document: about 60% retrieval accuracyEnd of document: about 90% retrieval accuracyRetrieval accuracy (%)

The lost-in-the-middle effect: a fact's position can matter as much as its presence. Illustrative, after Liu et al. (2023).

View the data
Where the fact sitsRetrieval accuracy
Start of documentabout 90%
Middle of documentabout 60%
End of documentabout 90%

Because these dynamics are rooted in the attention mechanism itself — not in training data gaps or implementation errors — no vendor has yet shipped a comprehensive fix. A bigger window gives you more room to fill with tokens; it does not, by itself, stop the rot. NVIDIA's RULER benchmark (2024) found that models advertising 128K-token windows often hold effective accuracy only to 64K or less, sometimes far less, depending on task complexity.

A model that fails loudly gets fixed. A model that fails fluently gets trusted — and that is the more expensive mistake.

Why it isn't being talked about

Three forces keep context rot out of the boardroom conversation. First, vendor incentives run the other way: context window size is a marketable number, and “our model gets measurably less reliable the more you use it” does not fit a product launch slide. Second, most enterprises aren't measuring output quality against input length at all; they track uptime and cost, not silent accuracy decay. Third, the failure mode is invisible — the system responds fluently every time, so a wrong or incomplete answer looks identical to a right one unless someone independently verifies it.

The fix: context engineering, not bigger windows

Nothing eliminates context rot outright, but disciplined context engineering can contain it dramatically. Hosseini et al. (COLING 2025) showed that targeted context-management strategies improved long-input accuracy by up to 50% while cutting API cost by up to 93% and latency by up to 50% — without any change to the underlying model. The goal is not to win a token-count arms race; it is to treat context as a scarce, governed resource.

Here is a practical framework — the CONTEXT principles — for leaders building AI into compliance, audit, underwriting, or due-diligence workflows:

The CONTEXT principles
  1. Chunk with intent

    Break documents into semantically coherent, retrievable units rather than dumping entire files.

  2. One task, one session

    Start a fresh prompt for each distinct analytical task. Marathon threads accumulate noise.

  3. Narrow before you ask

    Use retrieval to pre-filter relevant clauses or passages before sending anything to the model. The model should receive the answer and its immediate neighbours — not the entire filing cabinet.

  4. Top-load what matters

    Restate critical instructions, constraints, and key facts at the beginning (and, where helpful, the end) of every prompt.

  5. Evict aggressively

    Clear out stale turns and dead ends. Every retained line competes with what matters now.

  6. eXamine output against session length

    Plot accuracy against token count. If you are not measuring it, you are flying blind.

  7. Test with real semantic variation

    Benchmark with paraphrased, inferential questions — the kind your users actually ask.

Anthropic's 2024 Contextual Retrieval research offers a concrete data point for principle C: adding 50–100 tokens of explanatory metadata to each chunk reduced retrieval failures by 49%.

Case study: a mid-size BFSI GCC, third-party risk review

This case study is illustrative, based on composite patterns observed across multiple mid-size BFSI GCC pilots. It is not a report on a single, identifiable institution.

A GCC compliance team piloting an AI copilot fed it an entire vendor risk file — 340 pages of contracts, audit logs, and correspondence — in one pass, then ran a 90-minute interactive review asking follow-up questions as new documents were added mid-session.

Early questions, anchored near the top of the context, were answered with precision. But a critical clause — a vendor's undisclosed subcontracting arrangement, buried mid-document and referenced two hours into the session — was missed entirely. The model had “read” it. It simply never resurfaced it, because by then it was competing with hundreds of pages of accumulated, lower-relevance chat history sitting in the middle of a very long context.

The team caught the gap only because a human reviewer cross-checked the file independently — the copilot had given no warning that its confidence had quietly eroded.

The fix wasn't a bigger model or a longer window. It was re-architecting the workflow: retrieval-scoped queries against indexed clauses instead of full-document pasting; a fresh session per contract instead of one marathon thread; and critical obligations restated at the top of every new prompt. Review accuracy on flagged clauses improved measurably, and — just as important — the team now treats long AI sessions as a risk factor to manage, not a convenience to maximise.

The takeaway for transformation and risk leaders

Context rot is a governance issue disguised as a technical footnote. If your organisation is embedding AI copilots into compliance, audit, underwriting, or due-diligence workflows, the length and structure of what you feed the model is now a control point — as real as access permissions or data classification.

Measure output quality against session length. Architect for retrieval, not accumulation. And treat every long AI conversation the way you'd treat a long human shift: alert at the start, prone to quiet drift by the end.

The organisations that build this discipline in now won't just get better AI outputs — they'll be the ones who catch the missed clause before it becomes a headline.

So here's the question for your next AI steering committee: if your copilot's accuracy silently drops by a third over the course of a long review, do your current controls even detect it? And if they don't, what is the first contract, the first audit, or the first underwriting decision you are willing to bet on faith alone?

References

Every source below was checked against the original paper, dataset, or vendor publication before inclusion.

  1. Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2023/2024). “Lost in the Middle: How Language Models Use Long Contexts.” Transactions of the Association for Computational Linguistics, 12, 157–173.
  2. Chroma Research (2025). “Context Rot: How Increasing Input Tokens Impacts LLM Performance.” Evaluation of 18 frontier models including GPT-4.1, Claude 4, and Gemini 2.5. trychroma.com/research/context-rot
  3. Modarressi, A., Deilamsalehy, H., Dernoncourt, F., Bui, T., Rossi, R., Yoon, S., & Schütze, H. (2025). “NoLiMa: Long-Context Evaluation Beyond Literal Matching.” ICML 2025 / Adobe Research. arXiv:2502.05167
  4. Bansal, R., Zhang, A., Tiwari, R., Madaan, L., Duvvuri, S. S., Khatri, D., Brandfonbrener, D., Alvarez-Melis, D., Bhargava, P., Kale, M. S., & Jelassi, S. (2025). “Let's (not) Just Put Things in Context: Test-Time Training for Long-Context LLMs.” Meta, Harvard University, OpenAI, UC Berkeley. arXiv:2512.13898
  5. Xiao, G., Tian, Y., Chen, B., Han, S., & Lewis, M. (2024). “Efficient Streaming Language Models with Attention Sinks.” ICLR 2024.
  6. Hsieh, C.-Y., Chuang, Y.-S., Li, C.-L., Wang, Z., Le, L., Kumar, A., Glass, J., Ratner, A., Lee, C.-Y., Krishna, R., & Pfister, T. (2024). “Found in the Middle: Calibrating Positional Attention Bias Improves Long Context Utilization.” Findings of the Association for Computational Linguistics: ACL 2024, 14982–14995.
  7. Hosseini, P., Castro, I., Ghinassi, I., & Purver, M. (2025). “Efficient Solutions For An Intriguing Failure of LLMs: Long Context Window Does Not Mean LLMs Can Analyze Long Sequences Flawlessly.” Proceedings of the 31st International Conference on Computational Linguistics (COLING 2025), 1880–1891.
  8. NVIDIA (2024). “RULER: What's the Real Context Size of Your Long-Context Language Models?” NVIDIA RULER benchmark.
  9. Anthropic (2024). “Introducing Contextual Retrieval.” Reported a 49% reduction in retrieval failure rate from adding 50–100 tokens of chunk-level context. anthropic.com/news/contextual-retrieval

The case study is an illustrative, composite scenario built for this article and does not describe a specific named organisation.

First published on LinkedIn on 14 August 2026. Read the original and join the discussion

Getting started

Want to discuss what this means for you?

A practice leader will take you through the questions that matter for your organisation, without a sales script.

Or start small: a fixed-fee diagnostic in 2–4 weeks

  1. 30-minute discovery callYou describe the challenge, your stage and the outcome you need. We listen and ask the hard questions.
  2. Scoped proposal within five working daysA tailored scope, engagement model, team and indicative timeline.
  3. Kick-off with success measures agreedValue scorecard, governance and the first 30-day plan signed off before work starts.