The Summary Is a View, the Note Is the Source

An AI summary is a lossy view of your note. The note is the source. Treat the summary as derived: regeneratable, defeasible, never the record. Keep the original bytes in a sibling file. The discipline is small: one note, one summary, the same folder. The payoff is that a wrong summary stays reversible; an irreversible one is the failure mode.

The shape is older than the language we use for it. PostgreSQL's documentation puts it in two sentences: "Making liberal use of views is a key aspect of good SQL database design. Views allow you to encapsulate the details of the structure of your tables, which might change as your application evolves, behind consistent interfaces"¹. The summary is the view; the table is what you keep. The view exists to be regenerated against the table — never to replace it.

What an AI summary actually is

An AI summary is a probabilistic compression of a text into a shorter text. It is not a digest, not an index, and not a record. It is a model's guess at what mattered, written in the model's voice, sometimes accurate, sometimes wrong, never authoritative.

The original document is the only place from which the summary can be re-derived. Without the original, you cannot tell which summary lied. Ted Chiang put the analogy in mainstream prose in The New Yorker in 2023: "Think of ChatGPT as a blurry JPEG of all the text on the Web. It retains much of the information on the Web, in the same way that a JPEG retains much of [a higher-resolution image]"². The same shape applies one level down. A summary of your note is a blurry JPEG of your note — useful while you still have the original, dangerous the moment you do not.

The architectural analogy is cleaner. In a database, you do not delete the table after you create a view; the view exists because the table exists. PostgreSQL's docs are explicit that views are an interface over a table whose structure may change¹. The pair-file convention is the same instinct applied to plain text: write meeting-2026-06-27.md, generate meeting-2026-06-27.summary.md, keep both. The summary is consistent until the table changes; the table is the thing you actually keep.

Hallucination is the floor, not a bug to outgrow

Three Singapore researchers, Xu, Jain, and Kankanhalli, formalised the limit in 2024. They write that "Hallucination has been widely recognized to be a significant drawback for large language models (LLMs)"³ and prove that "it is impossible to eliminate hallucination in LLMs"⁴. The result is about what an LLM is, not a complaint about this generation of them.

The full theorem is sharper still: "LLMs cannot learn all the computable functions and will therefore inevitably hallucinate"⁴ when used as general problem solvers. The wrong-summary rate is not a bug a future version closes — it is a property of the architecture.

Empirical numbers anchor the floor. A 2025 study by Abeer Alessa and colleagues at UC San Diego, Quantifying Cognitive Bias Induction in LLM-Generated Content, measured that the tested LLMs "hallucinate on 60.33% of post-knowledge-cutoff questions" and "expose users to content that changes the context's sentiment in 26.42% of cases (framing bias)"⁵. Sixty point three three percent. On exactly the kind of question a research assistant is most useful for.

The same paper measured a behaviour change in the reader: "humans are 32% more likely to purchase the same product after reading a summary of the review generated by an LLM rather than the original review"⁵. The summary does not just lose information. It changes the decision the reader makes from it. The reversibility of that decision depends on the original review still existing somewhere the reader can read.

The BBC's research team measured the same failure on a real-world task in 2025. Four hundred BBC articles were summarised by ChatGPT, Copilot, Gemini, and Perplexity. The published finding: "51% of all AI answers to questions about the news were judged to have significant issues of some form. 19% of AI answers which cited BBC content introduced factual errors — incorrect factual statements, numbers and dates. 13% of the quotes sourced from BBC articles were either altered or didn't actually exist in that article"⁶. The four largest consumer assistants, on a news task, on a corpus with a known ground truth.

Even the best grounded summariser is wrong about 1.8% of the time

Vectara's public Hallucination Leaderboard tests grounded summarisation, the easiest case for the model, because the source is provided. The lowest hallucination rate as of May 2026 is 1.8% (antgroup/finix_s1_32b); OpenAI's gpt-5.4-nano posts 3.1%; gpt-5.4-2026-03-05 posts 7.0%⁷. Even on the easiest case, the floor is non-zero.

The leaderboard also surfaces a counter-intuitive operational fact: reasoning variants of the same family are strictly worse than their non-reasoning siblings on this benchmark. grok-4-1-fast-reasoning is 19.2% versus grok-4-1-fast-non-reasoning at 17.8%; grok-4-fast-reasoning 20.2% versus grok-4-fast-non-reasoning 19.7%⁷. The practical implication is plain: for summarisation tasks, turn reasoning off if the toggle exists.

Multi-document is worse. Catarina Belem and colleagues at Megagon Labs measured in 2024 that "on average, up to 75% of the content in LLM-generated summary is hallucinated" on multi-document tasks, and that GPT-3.5-turbo and GPT-4o "still generate summaries about 79.35% and 44% of the time" when asked about a topic that does not exist in the source corpus⁸. The failure scales with how much you ask the model to compress.

(Vectara publishes its own benchmark; treat it as a vendor-grounded floor rather than the industry consensus, and note that all of these numbers measure the grounded case — the model is being shown the source. When the source is not shown, the rates from Alessa⁵ and the BBC⁶ are the relevant ones.)

The lost piece is often the piece that mattered

A "low" rate hides the size of the loss it conceals. A 2025 npj Digital Medicine study by Asgari and colleagues evaluated LLM summaries of clinical notes: "Of 12,999 sentences in 450 clinical notes, 191 sentences had hallucinations (1.47%). Of which 84 sentences (44%) were major (could impact patient diagnosis and management if left uncorrected)"⁹.

One and a half percent of sentences. Forty-four percent of those errors major. The headline number does not describe the harm; the conditional does.

The same shape recurs in the general literature. A 2025 Scientific Reports paper by Liu and colleagues notes that "25% of summaries from the CNN/Daily Mail dataset, generated by traditional end-to-end pretrained language models, exhibit hallucination problems," and defines the term plainly: "hallucination is defined as the inclusion of any span in the generated summary that is not supported by the input document"¹⁰. That is the working definition this whole post sits on. A faithful summary is one whose every span is reachable from the source; an unfaithful summary is one whose spans are not. Keeping the source is what makes "reachable" checkable.

Apple Intelligence and the BBC: the news-summary case in the public record

In late 2024 and early 2025, AI-generated notification summaries from Apple Intelligence misrepresented BBC headlines: a summary stated that Luigi Mangione had shot himself; another that Luke Littler had won the PDC World Darts Championship before the final. The BBC complained publicly. Apple paused the news-notification summary feature in January 2025¹¹.

The detail is not that this happened to Apple — any vendor running uncalibrated summarisation at scale would have produced the same class of error. The detail is that the failure became public, dated, named, and conceded by the company that shipped it.

The corrective discipline is upstream of any vendor. If the summary is wrong, the question that follows is can you reach the source it summarised? Apple's pause is honest evidence that an irreversible AI summary is the failure mode. A reversible one — paired with its source, in a folder you control — is the version that survives the mistake.

The pair-file convention: small, ugly, durable

The discipline is two plain-text files in the same folder, named to the same stem: a source file ending in .md, and an AI-generated companion ending in .summary.md. The source is what you keep; the summary is what the model produced from it. Both are yours. Neither depends on an app to remain readable.

The convention in one block:

~/notes/meetings/
  meeting-2026-06-27.md           # the source you wrote or transcribed
  meeting-2026-06-27.summary.md   # the AI summary, generated against the source

That is the whole convention. The source ends in .md. The summary ends in .summary.md. They sit beside each other. They sort together. They migrate together. They are both plain text, readable by any editor in twenty years. There is no schema, no app, no toggle — only a naming rule consistent enough that you can find the pair without thinking.

What the convention buys is four things. You can re-read the source when the summary surprises you. You can regenerate the summary against a new model without losing the source. You can diff two summaries of the same source to see which model lied¹². You can grep the source for the span the summary should have been reachable from. The pair file is the artifact; the discipline is to look at the source when the view goes wrong.

The naming rule generalises¹³. interview-with-Y.md ↔ interview-with-Y.summary.md. paper-arxiv-2401.11817.md ↔ paper-arxiv-2401.11817.summary.md. book-chapter-3.md ↔ book-chapter-3.summary.md. The pair is recoverable from the filename alone, with no metadata, no database, no app-specific format. The convention is the schema.

What the pair file is not

It is not a security boundary. It does not prevent hallucination — it makes the hallucination reversible. It does not pay off unless you actually re-read the source when the summary surprises you; the habit is the artifact, not the file.

And it does not protect you from the most common failure mode of all: the summary is so well written that you do not feel surprised. Confidence in a well-written summary is a UX hazard, not a verification.

The discipline fails or shifts shape in three places, named here so they do not surprise you later. Audio and video are too heavy to keep as the "original" — pair the summary with the transcript file instead. Drafts you wrote from an AI prompt have no upstream original to keep; the prompt and the output are the only artifacts. And when the "source" you pair against was itself an earlier AI summary you trusted, you are chaining hallucinations rather than reversing them. The pair-file rule survives all three only if you read the source as primary and treat it as such.

A note on what hosting AI summarisation means for the local-first claim: if you send the source to a third-party model endpoint to generate the summary, the source text leaves your device on that path. The note you keep is yours; the round-trip you opt into is not. A meeting-2026-06-27.md paired with a meeting-2026-06-27.summary.md is local; the summary's generation is whatever your chosen provider's terms say. The discipline is about what you keep after the round-trip, not a claim that the round-trip itself was private.

How this lives alongside MNMNOTE

A pair file is just two Markdown files in a folder. In MNMNOTE, your notes are plain Markdown files on your own device, in folders you name. The pair convention is portable, version-controllable, and re-readable in any editor in 2046 — source and derived summary travel together because they are two of your notes.

MNMNOTE's bring-your-own-key approach means you pick the model that generates the summary against the source. The source stays in the folder you put it in. The summary file the model produces is a sibling, named by the same rule. Neither file is generated by the app; the convention works whether you used a hosted assistant, a notebook tool, or a local model. The point is that both files exist as plain text, side by side, owned by you, with the same lifespan as the rest of your notes.

Frequently asked questions

These six questions are the verbatim shapes of what people search before they reach for an AI summary of their own notes. The answers below are the version the pair-file convention answers in writing — once, on your terms — so the rule survives the moment of trusting a clean-looking summary you cannot audit.

Can I trust an AI summary of my notes? Not as the single record. The UC San Diego team measured 60.33% hallucination on post-knowledge-cutoff questions⁵, the BBC's 2025 study found that 51% of AI answers about news had significant issues⁶, and Vectara's 2026 leaderboard shows that even the best grounded summariser still hallucinates at 1.8% and most frontier models at 3–8%⁷. The honest framing is: a summary is a view, not a record. Keep the source so a wrong view is reversible.

Should I delete the original after AI summarises it? No. Deletion is what makes the hallucination irreversible. The original is what lets you check the summary later, regenerate it against a new model, or diff two summaries to see which one lied¹². A 1.47% sentence-level hallucination rate in clinical summaries sounds low until you read that 44% of those errors were major⁹ — the lost piece is often the piece you needed. Keep the source.

How do I summarise notes with AI safely? Use a pair-file convention: <note>.md ↔ <note>.summary.md, both plain text in the same folder. Pin the model and its settings — Vectara's leaderboard data shows that reasoning variants are strictly worse than non-reasoning siblings on summarisation, so turn reasoning off if the toggle exists⁷. Re-read the source whenever the summary surprises you, and treat the act of re-reading as the actual workflow.

Why are AI summaries wrong? Three reasons backed by primary sources: compression is lossy by design — Ted Chiang's blurry-JPEG analogy makes this concrete²; hallucination is formally unavoidable for any model used as a general problem solver — Xu, Jain, and Kankanhalli proved this in 2024⁴; and the failure scales with the amount of source material — Belem and colleagues measured up to 75% hallucinated content in multi-document summarisation⁸. None of these will be solved by a better prompt.

Can AI summaries change my decisions? Yes, measurably. Alessa and colleagues at UC San Diego found that "humans are 32% more likely to purchase the same product after reading a summary of the review generated by an LLM rather than the original review"⁵, and that summaries shifted the context's sentiment in 26.42% of cases. The summary is not a neutral compression of the source — it is a re-presentation of it, and the re-presentation moves the reader. The reversibility of that movement is the case for keeping the original.

What is the best AI for summarising without hallucination? There is no zero. The lowest published rate on Vectara's grounded-summarisation benchmark, as of May 2026, is 1.8% (antgroup/finix_s1_32b); gpt-5.4-nano is 3.1%; gpt-5.4-2026-03-05 is 7.0%⁷. The honest answer is not to find the model that will not lie — it is to pair every summary with the source it summarised. Note that the leaderboard measures the grounded case (the model is shown the source). Ungrounded recall ("summarise from memory") fails far more often⁵⁶.

This essay sits next to Steph Ango's line on files: "Apps are ephemeral, but your files have a chance to last"¹⁴. The summary is the app's artifact; the source is the file you keep. The same instinct that says grep your own notes underwrites audit your own summaries — the deeper version of this argument lives in the personal-RAG piece next door¹⁵, and the input-side companion to this one decides what you should never send in the first place¹⁶.

Chiang's punchline is the post in one line. Hold the source; ignore the rumour that the compression is the original; remember that "we aren't losing our access to the Internet. So just how much use is a blurry JPEG, when you still have the original?"¹⁷ The same question, asked one level down, of your notes, is what the pair file answers in writing. The other side of the same coin lives at mnmnote.com.