A Shorter Summary Is Not One Edit — Name the One You Want

Cartographers needed 24 named operators to say what happens to a map when it gets smaller. “Make it shorter” uses none of them. Text has its own list: a 2000 study of human-written summaries found six distinct edits, from deleting clauses to rewording them more generally. Name the reduction you want, then check the output against that one.

The map list comes from Robert Roth, Cynthia Brewer and Michael Stryker, in a peer-reviewed 2011 paper in Cartographic Perspectives. They sorted the operators into “four higher-level categories of multiscale mapping operators: content, geometry, symbol, and label,” each with “a formal definition” and “a sample illustration” ¹. Not all 24 shorten anything, and only a handful translate to prose.

The text list comes from Hongyan Jing and Kathleen McKeown at Columbia University, who “manually analyzed 30 articles and their corresponding human-written summaries” ². In a study published in 2024, a Stanford and Columbia team that included McKeown used those operations to compare a language model with freelance writers on the same news task. The two made measurably different edits ³.

Our post on how an AI summary says more than your notes do looked at what goes wrong in a summary. This one asks the question before it: which edit did you want?

The study, briefly: 24 names for a smaller map

Roth, Brewer and Stryker’s 2011 typology names 24 operators for keeping a map legible as its scale shrinks. Four cover content, seven geometry, nine symbols and four labels. The total is our count of their list, since the paper never prints the number. Its premise is that reduction is a set of decisions, not one.

The introduction puts it bluntly: “As any well-trained cartographer will tell you, there is no one-click solution that automatically discriminates essential geographic information from irrelevant or excessive detail” ¹. A shrinking map has to lose something. The typology says what kind of loss.

Three definitions show why the names matter. Eliminate “removes features once a scale is reached where they become illegible or no longer fulfill their intended purpose.” Aggregate is the “replacement of many related features with a representative feature of increased dimensionality.” Reclassify “alters the way that features are organized in the representation based upon their attributes in order to improve map legibility” ¹.

Each produces a smaller map. Each produces a different one.

The most telling line concerns the word simplify. Its use, the authors write, “has evolved from a more generic descriptor of any action that reduces detail or data volume (Robinson et al., 1978) to its present-day, narrow focus on eliminating points” ¹. The field had to narrow a vague word before it could use it. “Shorter” has never been narrowed.

Not every operator is a cut. Add and add labels insert things, enhance and exaggerate make a feature more prominent, and the colour, transparency and rotation operators have no counterpart in a sentence ¹. Even reorder means the stacking order of map features. What transfers is the method: split one vague request into named operations, each with a definition you can check.

What they actually measured: six edits inside a human summary

Jing and McKeown compared 30 news, medical and legal articles with their human-written summaries. They found six operations that turn source sentences into summary sentences — sentence reduction, sentence combination, syntactic transformation, lexical paraphrasing, generalization or specification, and reordering. These “can be used alone or together,” so one summary sentence can carry several edits.

The corpus was small — 15 general news reports, 5 medical and 10 legal documents ². Two operations matter most for a note. In sentence reduction, “the deleted material can be at any granularity: a word, a phrase, or a clause.” Generalization or specification means: “Replace phrases or clauses with more general or specific descriptions” ².

Their example deserves a slow reading. The phrase “a proposed new law that would require Web publishers to obtain parental consent before collecting personal information from children” became “legislation to protect children's privacy on-line” ². The second phrase is shorter and true. It has also lost the mechanism — the consent, the publishers, the collection.

People also write from nothing. In an automatic analysis of 300 human-written abstracts, the authors “found that 19% of sentences in the abstracts were written from scratch” ². Roughly one sentence in five had no source sentence behind it.

More than two decades later, Tianyi Zhang and colleagues gave a model and freelance writers the same news-summary task and read the edits through that lens. Their verdict: “human-written summaries contain more lexical paraphrasing and sentence reduction whereas the Instruct Davinci model has more direct copying from the article” ³.

The model was a 2022-era OpenAI system. Its summaries scored 0.92 on coverage and 12.1 on density, against 0.81 and 2.07 for the writers’ — higher means closer to the article’s own wording ³.

Same task, different edits.

Neither was simply better: “every rater has their own consistent preference for either Instruct Davinci or the freelance writers” ³.

One word, three meanings: generalization

In our post on over-generalization, generalization is a defect: a summary widening a hedged claim into a sweeping one. In cartography it names the craft of reducing a map. In Jing and McKeown’s list it is one named summarizing operation. The move is similar in all three — fewer specifics, a broader description. The difference is whether it overshoots.

That matters, because the two posts could read as a contradiction. The earlier one reported a study in which asking a model for accuracy made over-generalization worse. It concluded: “The defense is not a better prompt; it is a kept original.” This post does not offer a better prompt either. It offers a name for the edit you wanted, which gives you something specific to check the output against.

Through Jing and McKeown’s list, that defect is operation five overshooting. Naming the operation does not stop the overshoot — it tells you where to look.

From here on, this post says reduction operation and leaves generalization to the sources.

Why this matters for your notes: “shorter” is a length, not an edit

Asking for a shorter summary sets a length and leaves the edit to the model. The 2024 study shows the gap: one task produced copying from the model and reduction from people. The US guideline for abstracts works differently. It names distinct kinds of summary, and it sets length by document type rather than by “shorter.”

It is the default request. A Hacker News commenter put it plainly in December 2024: with ChatGPT, “you can iteratively improve text,” and the first example given was “make it shorter” ⁴.

Length is a weak instruction even for length. The 2024 researchers found the 2022-era model “consistently produces summaries that exceed a given word limit,” so they asked for 25 words to get summaries averaging 50 ³. A word count is not an operation. It does not say what to remove.

ANSI/NISO Z39.14 is the 1997 guideline for abstracts, reaffirmed in 2015. It refuses to treat “summary” as one thing: “The term abstract should not be confused with the related but distinct terms: annotation, extract, summary, and synoptic” ⁵. Five named products — where a chatbot sees one.

It then splits the abstract itself in two. Informative abstracts “state the purpose, methodology, results, and conclusions presented in the original document.” An indicative abstract should “describe the purpose or scope of the discussion or descriptions in the document” ⁵.

For a note, that is the most useful split on the page.

“What is this note about?” and “What did this note find?” are two requests. Both come back shorter.

Even the standard’s lengths follow the document type. When no length is specified, it suggests about 250 words for papers and articles, 100 for notes and short communications, and 30 for editorials and letters to the editor ⁵.

A working vocabulary for shorter

The table below is our translation, not any source’s. Each row pairs a named operation from the map or text literature with a sentence you can type and a check you can run on the output. Two rows are marked risky, because by definition they can produce wording your note never contained.

Typify sounds like the gist operator, so it needs its full definition. Unlike eliminate, “the typify operator uses only the spatial characteristics of the features to generate the new arrangement of symbols that were not from the original set” ¹.

On a map, an invented symbol is a legible convention. In a summary of your note, it is a sentence nobody wrote.

The safest row is the first, and it is not new here. Make the AI quote your note before you believe it already asked for one named operation: to “return the extractive form — the exact words from the note — first.” The abstracting standard defines the same product: “Extract – One or more portions of a document selected to represent the whole” ⁵.

An extract can be checked with a search. That is its virtue.

How to apply this tomorrow

Replace “shorter” with the name of one operation. Decide whether you want what the note is about or what it found, state the length separately, and check the result against the operation you named. The name is not a guarantee that the model performs it. It is the standard you hold the result to.

1. Pick one verb from the table. Remove, merge, regroup, tighten or reorder. One operation per request gives you one thing to check.
2. Say “about” or “found.” An indicative summary of a meeting note lists the topics; an informative one states the decisions ⁵.
3. Give the length as its own instruction. Expect misses from models and people alike — even “instructing writers with a sentence limit often resulted in summaries that differ significantly in length” ³.
4. Check against the operation you named. After remove, every sentence should appear in the note. After regroup, the count should not change. After tighten, no claim should be missing.
5. Quote-check anything typified or generalized. The standard’s rule is the right test: “Do not include information or claims not contained in the document itself” ⁵.

A mismatch is a finding — something you can point to.

The caveats

Maps are not text, and the map half of this post is an analogy. The text half stands on its own evidence, but the model studies used 2022–23 era systems summarizing news, not anyone’s notes. Naming an operation also does not make a model perform it: a 2024 benchmark found far more factual errors when summaries had to meet stated requirements.

Yixin Liu and colleagues concluded that “instruction controllable text summarization remains a challenging task for LLMs” ⁶. With stated requirements, the error rates of text-davinci-002 and gpt-4-0314 were 36% and 14%, against 1–2% for text-davinci-002 on generic news summaries ⁶. A named operation is a check, not a safeguard.

The cartographers flagged the same risk in their own tools: “trained and untrained cartographers alike would be able to generate inappropriate multiscale map designs more quickly and more easily” ¹. They add that “no single ScaleMaster provides a ‘best’ solution” and that “the science of multiscale mapping needs to catch up to its practice” ¹. The typology is a framework, and its authors say so.

Jing and McKeown were careful too: “we make no claim that humans use the same cut and paste operations” ². Their six operations describe the summaries — how writers think is another matter. People invent as well, and the 2024 judges split on which summaries they preferred.

The operation-to-prompt table is ours. No source above recommends a prompt, and none studied personal notes.

Frequently asked questions

These are the questions people ask when a summary comes back shorter but wrong. One answer runs through all of them: “shorter” describes a length, not an edit. Name the reduction operation you want and check the output against it, knowing that the name alone will not make the model comply.

What should I ask for instead of “make this shorter”? Name one operation: remove what does not serve a purpose, merge related points, regroup by a new key, tighten without dropping claims, or reorder. Cartographers use 24 named map operators, not all of them reductions; summarization research names six text edits. Check the result against the one you asked for.

How do I tell an AI what kind of shorter I want? Give the operation, the purpose and the length as separate instructions. Say whether you want what the note is about or what it found, which the abstracting standard calls indicative and informative. Treat the length as a target the model may miss.

Why does the AI summary add things that weren't in my note? Some reduction operations create new wording by design. Broader descriptions and representative examples can produce sentences the note never contained. People do it too: 19% of sentences in 300 human-written abstracts were written from scratch. Check any sentence you cannot trace back to the note.

Why are AI summaries too short to be helpful? Often the request asked for less text, not for a kind of summary. An indicative summary says what a document is about; an informative one states its purpose, methods, results and conclusions. If you got the topic but not the findings, ask for the informative kind by name.

Why does the summary miss the interesting points and numbers? It probably removed when you wanted it to tighten. Removing drops whole points; tightening keeps every claim in fewer words. Ask for fewer words with the same claims and every number kept, then count the claims and figures against the original note.

Is asking for bullet points a way to make a summary shorter? Bullet points are a layout, not a reduction. They say how the output looks, not what is removed, merged or kept. Pair the format with an operation, such as one bullet per decision in the note’s own words, so you know what each bullet should contain.

Does naming the operation make the AI follow it? Not reliably. A 2024 benchmark found more factual errors in summaries with stated requirements than in generic ones, including 14% for a March 2023 version of GPT-4. The named operation is what you check the output against, not a guarantee. Keep the original so there is something to check.

If “shorter” can mean any of the eight rows above, then every summary you accept was an edit someone chose. The note stays the source, as let the AI summarize, but keep the original argues; the named operation is what you hold the summary to. A reduction you did not name is a reduction you cannot check.

MNMNOTE keeps each note as plain Markdown on your own device, so the original stays open beside any summary you check against it — mnmnote.com.