Google released Deep Review, which analyses technical documents. It is built to read the way a reviewer reads, not to summarise.
Google Research published Deep Review for analysing technical documents. Rather than shortening a document, it is meant to surface whether claims line up with their evidence and whether verification is missing — moving part of the slow work of peer review onto the machine. PAINT, a framework that structures multilingual expressions of pain from medical licensing exams, appeared on arXiv the same week, but its subject is representation learning for medical language and it runs on a separate track.
Summarisation reduces the effort of reading; it does not reduce the judgement of whether something is worth reading. Deep Review aims past that, having the machine point at whether a document actually carries its claims. The swing is largest in domains where checking correctness costs expertise and time, such as technical documents and papers. What is unsettled is who owns the call when the machine says the evidence is thin. Note: no third-party assessment of Deep Review's accuracy has appeared.
Whether Deep Review's flags get folded into a real review process. Whether anyone audits its misjudgements, or argues about where responsibility for a flag should sit.