Google has launched Deep Review, its AI system for reviewing and checking papers. It’s the public version of the Paper Assistant Tool described in its June preprint: Towards Automating Scientific Review with Google’s Paper Assistant Tool. The service gives you 2 free reviews per week with 4 Argon on Medium. If you are using the API, “expected costs for a 40 page document range from ~$3.00 for Low with 3.8 Flash to $20 with 4 Argon on Extra High.” Next step “AIrXiv”, the fully automated reviewer and preprint publishing pipeline described in the preprint?
The results from a couple of quick and crude tests (comparisons written by ChatGPT):
Test 1: Hyperglycosylation is a metabolic driver of Alzheimer’s disease, Nature Metabolism
Paper: https://www.nature.com/articles/s42255-026-01538-4
PubPeer comments: https://pubpeer.com/publications/A6FC89F1344C7ECE582D68E357D6D4
Deep Review output: https://innovationideas.co.uk/wp-content/uploads/2026/10/Deep_Review_Alzheimers_Nature_Metabolism.txt
The main PubPeer criticism is pseudoreplication. Thousands of image pixels were treated as independent observations although they came from only a small number of human or mouse brains. Deep Review caught this clearly and suggested analysing at biological-specimen level or using a hierarchical model.
It also identified a much longer list of other potential issues. A couple I [ChatGPT] checked were real: the paper describes shPGM3 injections into the lateral ventricles in the text and Methods, but Extended Data figures describe hippocampal injections; and the Results give the NGI-1 dose as 2 nmol while the Methods give 5 nmol. It missed some specific points raised on PubPeer.
Test 2: Dupilumab enables glucocorticoid withdrawal in refractory multisite mucous membrane pemphigoid: a case report, Frontiers in Medicine
Paper: https://www.frontiersin.org/journals/medicine/articles/10.3389/fmed.2026.1915046/full
PubPeer comments: https://pubpeer.com/publications/4E6FA79CBD5A3D3E020FBB1A672F05
Deep Review output: https://innovationideas.co.uk/wp-content/uploads/2026/10/Deep_Review_Dupilumab_Frontiers_Medicine.txt
The PubPeer comments focus on the images, including questionable before/after photographs and OpenAI provenance signals. Deep Review missed all of this. It actually said the clinical and histopathology images aligned with their textual descriptions, and found no significant issue with the generative AI disclosure.
It did identify a number of other clinical, methodological and reporting weaknesses. For example, it pointed out that the claim of remission on dupilumab monotherapy was based on only 1 month after prednisone withdrawal, and that the paper says the patient also received “topical anti-inflammatory therapy” without saying what treatment was used, making it harder to attribute the improvement specifically to dupilumab.
