AI News Summaries: What They Get Right and Where They Mislead

A summary tells you an exchange is listing a token. The article it was built from says the exchange is reportedly in early discussions about a potential listing, according to two people who were not authorised to speak publicly.

Both statements are about the same event. Only one of them is worth acting on, and it is not the short one.

The Crypto App ships AI summaries in its news feed. This page is not going to pretend they are a solved problem, because the failure modes below are properties of summarisation itself rather than something any particular vendor has engineered away. Knowing where the compression loses information is what makes the feature useful instead of dangerous.

What are AI summaries genuinely good at?

Four things, and they are not trivial.

Deduplication. The single largest problem with a crypto feed is that one event produces forty items. Collapsing those into one entry that says "these forty headlines are about the same funding round" is real work that saves real time, and it is the job summarisation does best because it requires only recognising similarity.

Answering what happened while you were asleep. Crypto markets do not close, and the overnight catch-up is genuinely well served by compression. You want to know whether anything happened, not to read eleven articles about whether it did.

Stripping boilerplate. Crypto articles carry an unusually high ratio of restated background to new fact. Two paragraphs explaining what a Layer 2 is, one paragraph of news, a closing paragraph of context. A summary that returns the middle paragraph has removed padding rather than information.

Translation. If English is not your first language, reading a governance proposal or a regulatory notice in your own is not a convenience, it is the difference between reading it and not. The app supports 12 languages, and this is the use case where compression costs least, because the alternative was not reading the original either.

None of that is marketing. Triage is a genuinely hard problem and summarisation is genuinely good at it.

How does summarising change what a claim is saying?

By removing the words that carry the uncertainty, because those are the cheapest words to remove.

Look again at the example at the top. The hedges in that sentence are reportedly, early, potential, unnamed sources, and not authorised to speak. Every one of them is grammatically optional. You can delete all five and the sentence still parses, still describes the same event, and reads considerably better. A system optimising for brevity will delete them almost every time, and the result is a claim that has been silently promoted from rumour to fact.

This is the single most consequential failure, because in crypto reporting the hedges frequently are the information. The difference between "has partnered with" and "has signed a non-binding memorandum of understanding regarding a potential partnership" is the entire content of the story.

Attribution goes the same way. "According to a research note from a firm that also holds a position in the token" compresses to "analysts say." The conflict of interest was the most useful thing in the sentence, and it was in a subordinate clause, which is exactly where compression cuts. In a sector where the person making a claim very often owns the thing they are claiming about, losing the claimant loses most of the signal. This is the same structural problem as paid placements: the disclosure lives in a part of the document that does not survive being passed along.

Two more losses are worth naming. Negations and conditionals are fragile: "the regulator said it would not object provided three conditions were met" compresses very naturally into "regulator approves," which is a different claim about a different world. And numbers are where summarisers err most often while being the facts most likely to move your decision, so a percentage, a token amount, a deadline or a jurisdiction read only in a summary should be treated as unverified.

Then there is tone. Generated summaries have one voice, flat and declarative, and it does not vary with the reliability of the input. A human newsroom signals doubt structurally, through where a story is placed, how long it runs, and whether the headline is a question. A summary of a court filing and a summary of a Telegram screenshot read identically. You lose an entire channel of information you were not aware you were using.

What happens when forty outlets rewrite one source?

The failure that a wider feed makes worse rather than better.

One outlet publishes a claim. Within four hours, thirty-nine others have rewritten it, each citing the first, some citing each other. A summariser now sees forty documents asserting the same thing. Corroboration across many sources is a reasonable heuristic in most domains, and here it produces exactly the wrong answer: there is one source and thirty-nine copies, and the copies contain no independent verification whatsoever.

Worse, the copies tend to be more confident than the original, because each rewrite drops another hedge. The chain degrades in one direction. By the tenth rewrite, "sources suggest" has become "it has been confirmed," and no one lied at any step.

This deserves saying plainly about our own product. The Crypto App's feed draws on around 68 publishers. A curated set is better than a scrape, and naming the publisher on every item is the specific thing that lets you trace a claim back. But breadth of publisher coverage is not breadth of reporting, and a summary that aggregates across a feed will reflect how many outlets copied a story rather than how well founded it was. That is a limitation of the format, and no number of publishers fixes it.

When should I go read the original?

One rule covers most of it: if the summary would change what you hold, open the source. If it would not, the summary was sufficient and you have lost nothing.

Beyond that, four triggers are worth memorising. Open the original when the summary names no one, because unattributed claims are where the sourcing was dropped. Open it when the summary is about something unresolved yet contains no hedging at all, because unresolved situations have hedges and their absence means removal rather than certainty. Open it when a number is doing the work. And open it when the summary describes an intention rather than an event, since the gap between announcing and doing is where most disappointment lives, as when news actually moves a coin covers in more detail.

This is also the argument against the weekly recap format that summarisation makes so easy to produce. A digest answers "what happened" well and "does this change anything I hold" not at all, and by the time you read it the price has already responded to whatever mattered. If the useful trigger is a change in state rather than the passage of seven days, an event-based alert or a News widget does more for you than any recap, because both are tied to something occurring.

Use summaries the way you would use a table of contents. They are excellent at telling you which of the day's forty items are actually six, and which of those six you should read properly. They are poor at telling you what those six mean, and they are worst precisely where the stakes are highest, which is on claims that are contested, conditional or quantified. Read the summary. Then, when it matters, read the thing.