AI summaries: the question is not how much they cut, but whether you can get it back
In a Feature Article published in eLife on 14 October 2026, Giovanni Spitale and Federico Germani, co-directors of the ITE Lab at the Institute of Biomedical Ethics and History of Medicine (University of Zurich), set out a criterion for evaluating AI-generated syntheses of the scientific literature — and argue that the criteria currently in use cannot see the problem that matters most.
The article takes its title from Douglas Adams: a supercomputer returns “42” as the Answer to the Ultimate Question, but the question itself has been lost. An answer severed from its question and its context is inert — it cannot be used, and it cannot be checked. The authors argue that machine summarisation reproduces this structure. Compression works by leaving things out, which is pure efficiency up to a threshold; beyond it, the summary loses the conditions under which the original finding holds. For scientific findings, the authors note, those scope conditions are not background a reader can supply: they are part of the finding.
Two consequences follow. First, since the threshold depends on the reader and on the question being asked, no single level of compression can be correct for everyone — which is why “superhuman” synthesis is the wrong benchmark, and why existing evaluations miss the issue. Second, and more consequentially, the authors predict a form of drift: a summary compressed for one purpose carries no record of that purpose, and may later be reused where the omitted information was decisive. Each such failure begins as a correct use of the tool, so better models will not prevent it.
Their proposal is to make recoverability a reported property of machine-generated synthesis, through three measures that do not require knowing the reader’s question: coverage, the proportion of claims linked to a source; depth, the effort needed to reach the passage that warrants a claim; and warrant, whether the cited source in fact supports the citing sentence. Recent evaluations of deep-research agents show why the third matters: link validity above 94% alongside factual support as low as 39%.
The authors note that journals already require the human equivalent of this standard, in the ICMJE recommendation that quoted material carry full attribution. Reporting recoverability applies the same principle to machine output, making the traceability of claims measurable rather than a matter of citation practice.
Spitale G, Germani F (2026). AI in Science: The 42 problem in AI-assisted research. eLife — open access, CC-BY.