Why I Still Read the Sources Myself (Even When the AI Summarized Them Well)
The summary was accurate. I read the paper anyway, and the gap between the two is the whole argument of this post.
Last month a summary of a research paper, produced by a tool I trust, was completely accurate. Every claim it made was in the paper. I still spent forty minutes reading the actual study, and I want to defend that apparently irrational use of time.
Accuracy is not the same as emphasis. The summary correctly reported that the intervention improved outcomes by 12 percent. What it compressed away was that the effect appeared in one subgroup and vanished in two others, and that the follow-up period was shorter than the disease's actual timeline. Nothing in the summary was false; the shape of the truth was different from its summary of it.
Summaries optimize for compression, and compression has a fixed budget. Someone, some model, some weighting decided which details were the 5 percent worth keeping. For decisions I do not care about, I am happy to inherit those choices. For anything I will act on, repeat, or be asked about, the emphasis layer is where the value is, and emphasis cannot be outsourced without inheriting someone else's judgment about what matters.
There is also the memory effect. I remember the papers I read; I do not remember their summaries. Reading slowly is how the material becomes mine rather than something I am quoting. The summary tells me whether to invest that reading time, which is genuinely valuable, but it is a film trailer, not the film.
So my workflow stays two-step and I have stopped feeling inefficient about it. Summaries decide what deserves the forty minutes. Reading earns the right to have an opinion. Tools that promise to skip the second step are not saving time; they are renting me a confidence I have not earned.