Reading and Researching Faster Without Losing Accuracy

The main question: does it hurt or help?

The first measurement. Seven experiments, 10,462 participants in total. The most decisive of them is the second: both groups received the same information, one as a summary, the other as a set of links.

So the explanation is not how much information there was, but the form in which it had to be processed. Whoever got it ready-made spent less time on it, learned less, and could pass on fewer facts. In a separate round, independent raters judged the advice written from the ready-made summary less informative and less trustworthy.

The second measurement. 124 participants were given the same research paper, one group with a summary and the other without, and then answered comprehension questions.

The second row is the point: readers who had a summary also answered better on the questions the summary never covered. It did not replace the reading — it prepared it. One of the authors put it this way: the summary is like a starter that leaves you hungry for the main course.

The resolution. The two experiments are not measuring the same situation:

This whole post in one sentence: an AI summary hurts when it stands in for the reading, and helps when it leads into it.

How faithful is the summary itself?

There are measurements for two tasks of very different difficulty, and the gap between them is wide.

The easy case — summarising one document you hand over. In a public benchmark of 7,700 articles, the hallucination rate at the top of the field runs between 1.8% and 4.1%. That sounds good, but two caveats come with it. First: the benchmark explicitly asks the model to rely only on the facts present in the document — the easiest task imaginable. Second: one of the well-performing models has an answer rate of only 80.7%, meaning it stays silent in one case out of five. Part of that low hallucination rate is manufactured by refusal — these two numbers always have to be read as a pair.

The hard case — news and multiple sources. A study run by public service media organisations had more than 3,000 AI answers assessed by professional journalists, across 22 organisations, 18 countries and 14 languages, on four major assistants. The picture as of April 2026:

• 45% — the answer contained at least one significant error;
• 31% — a serious sourcing problem;
• 20% — a serious accuracy problem.

For one of the tools examined, 76% of the answers contained a significant problem, overwhelmingly because of bad sourcing. The report calls the errors systemic rather than isolated — and states that this holds regardless of language, country or platform.

The weakest point: sourcing

This is where the gap is widest between what the tool shows you and what it actually knows.

A university research centre tested eight search assistants with 1,600 queries: they submitted an excerpt from an article and asked where it came from. More than 60% of the queries came back with a wrong answer. The best tool was wrong 37% of the time, the worst 94%.

The researchers' own wording is the most important part of all of it: the wrong answers were delivered with "alarming confidence", rarely using any hedging expression at all. Confidence does not correlate with correctness.

A measurement from April 2026 shows the same thing from another angle. Across more than four thousand searches it examined the generated summaries that appear at the top of the results page:

• the answer was correct in 91% of cases on the question itself;
• but only 39% were at the same time correct and supported by the sources cited;
• 67% of the claims were backed by the sources cited — meaning every third claim is unsupported;
• and the ratio is getting worse: over the same half year accuracy rose from 85% to 91%, while within the correct answers the share of unsupported ones rose from 37% to 56%.

(This measurement was carried out by a company that itself sells a hallucination-detection model — that caveat belongs next to the number. The search provider concerned disputed the methodology, arguing that the question set used does not reflect real searches.)

The common denominator across the news study, the sourcing test and the search-summary measurement is the same: the answer as a whole holds up, its parts do not. Which is exactly why "I read it, it looks fine" does not work.

How much do we actually check?

There is a number for this too, and it is less flattering than we would think.

A study surveying 319 knowledge workers collected 936 real cases of AI use at work. In 555 of them (roughly 60%) some form of critical activity took place — which means in two cases out of five there was no checking at all.

And the most telling detail: out of the 319 respondents, 114 said they also compare the output against an external, trustworthy source. Barely more than a third.

The study's main correlation is this: the greater the trust in AI, the less critical thinking — the relationship is strong and significant. The greater someone's confidence in their own expertise, on the other hand, the more of it there is.

On top of that comes the finding that casts doubt on every self-estimate in this area: in a controlled experiment the participants felt 20% faster while the measurement showed them 19% slower. Thirty-nine percentage points. There is more on that in the real time savings post, broken down by type of task.

The order the research supports

Five steps. Behind each one stands one of the measurements above.

1. YOUR OWN QUESTION FIRST, IN WRITING, NO AI
   What you want to find out, and what you know already.
   Two sentences is enough. This step gives you depth.

2. THE SUMMARY GOES BESIDE THE ORIGINAL, NOT INSTEAD
   That is exactly the difference between the two studies.

3. CHECK IT CLAIM BY CLAIM, NOT AS A WHOLE
   "Which source does this sentence come from, and where?"
   The answer is 91% right, the claims 67%.

4. NUMBER, DATE, NAME, QUOTE: A SEPARATE PASS
   These four go wrong most often, and these four
   are the ones you pass on.

5. CONFIDENCE IS NOT A SIGNAL
   The wrong answers that were measured sounded sure.
   If there is no "probably" in it, that means nothing.

A copy-paste instruction for the third step:

Summarise this material, but after every claim put in
square brackets which part of the source it comes from
(page, section heading, or the first few words).

Anything you cannot place exactly, do not write down.
At the end, list separately what is NOT in the source
in answer to my question.

The last sentence is the most important one. In an office setting the commonest fault of summaries is not the lie but the omission — and none of them flags what is missing on its own.

Common misconceptions

• "I read the summary, I know what is in it." The large experiment series refutes precisely that: even with identical information content, the people who worked from a summary could pass on fewer facts.
• "If it gives a source, it has checked it." No. The measurement found that a third of the claims are not supported by the sources named, and that sourcing is getting worse while accuracy improves.
• "Low hallucination rate, good model." Look at the answer rate next to it. A model that declines one question in five is buying its low error rate cheaply.
• "I take a closer look whenever something seems off." According to the measurement, the wrong answers sound confident. Nothing sets the suspicion off.
• "I do check, though." Barely more than a third of the knowledge workers surveyed compared the output against an external source. Here too, our judgement of our own habits errs upwards.

What to take away

The AI summary is not a question of speed or of accuracy — it is a question of order. If it arrives before your own question, it takes away part of the thinking, and by the measurement you will not even notice. If it arrives alongside the original, it helps you read more deeply.

And on the checking, one thing is worth remembering: do not look at the whole, look at the parts. The answer is very probably correct. Its individual claims are considerably less probably supported — and a citation being there is not the same as a citation being right.

What that difference means in a contract or an offer, where a wrong citation costs money, is set out with numbers in the document drafting post; the note-taking side of it is where the note-taking post gives you a recipe.