Some of the contractors hired to improve OpenAI's models have been using AI systems to train those same models, despite explicit instructions not to. 404 Media found multiple instances where contractors were fired for this practice, creating a feedback loop that researchers warn can degrade model performance over time through "model collapse."
The irony is sharp: the people tasked with preventing AI-generated training data from polluting future models are themselves using AI to generate training responses. This comes as the broader AI research community grapples with a flood of AI-authored papers threatening to overwhelm traditional peer review and scholarly reading practices.
AI agents design drug candidates without a lab
A research team deployed a "Virtual Biotech" - a system of as many as 37,000 AI agents - to identify candidate drugs for cancer therapy. The agents uncovered a molecular signal that could help predict clinical-trial success and, with human oversight, flagged a promising lung-cancer treatment. The system has not been tested in real-world drug discovery, and its predictions remain unvalidated by experiments or clinical trials.
Separately, Anthropic announced that its Claude AI autonomously discovered a novel enzyme system associated with an array of DNA repeats, a pattern reminiscent of CRISPR. The company said the system's characteristics have only been found together in a handful of other systems, all programmable for operations like cutting, copying, and pasting DNA. The function of the newly discovered system remains unknown.
Security flaws exposed in a patient-facing chatbot
A perspective article in NEJM AI describes how researchers, using only a standard web browser and a commercial AI assistant, retrieved a patient education chatbot's full system prompt, RAG pipeline configuration, internal vector database address, knowledge base documents, and the 1,000 most recent patient conversations - all without authentication. The authors argue that AI tools using retrieval-augmented generation must face the same rigorous pre-deployment security testing required of other health technologies.
Reading in the age of AI summaries
David Crotty, writing in Scholarly Kitchen, frames scholarly reading as split between quick-skimming "foxes" and deep-reading "hedgehogs." AI-generated summaries, he argues, place all the value on outputs rather than processes. "The processes of reading, analyzing, reflecting, and theorizing are important to how we learn to think," Crotty wrote. "The output doesn't matter, but the process of reaching that output does."
Giorgio Gilest pushes the point further in a separate essay, warning that free access to research is worthless when attention, not access, is the binding constraint - and attention is allocated by brand. "We will have open archives that nobody reads and closed reputational cartels that everybody does," Gilest said. "That is a worse settlement than the one we are leaving."
Skepticism toward superintelligence claims
Timnit Gebru and Emily M. Bender, writing in MIT Technology Review, offered a sharp critique of current AI hype and doomsaying. "Claims of incipient, dangerous superintelligence are not based in good scientific or engineering practice," they wrote. They questioned why computer programming and math receive so much attention as fields for applying large language models, noting that these domains involve problems where answers, once suggested, can be verified - making them easier targets for AI benchmarks than the messy, ambiguous work that defines most professional and scientific reasoning.
Early-life sugar exposure linked to later cancer risk
Researchers exploiting a natural experiment created by postwar sugar rationing in the United Kingdom found that people exposed to less sugar in early childhood were significantly less likely to develop several cancers later in life and showed signs of slower biological aging. Even in adulthood, they continued to consume less sugar and maintained healthier diets overall.
Meanwhile, a survey of 31 leading cancer centers found all reported shortages of at least one cancer drug. Ninety percent said treatment plans had to be modified because of shortages, creating additional insurance hurdles that may have delayed care. Nearly two-thirds resorted to mitigation strategies to keep patients on recommended treatments.
Why this matters for healthcare, science, and research professionals
The contractor story at OpenAI is a concrete warning: AI models trained on AI-generated data degrade, and the guardrails meant to prevent this are already failing. For researchers who rely on these models - or who review papers built with them - the quality of the underlying tools is not a given. The patient chatbot security test shows that RAG-based clinical tools can leak protected health information through shockingly basic exploits, reinforcing that pre-deployment security testing is not optional. And the sugar rationing study offers a rare, high-quality natural experiment linking early-life nutrition to cancer outcomes decades later - the kind of evidence that is difficult to generate any other way. For those building or evaluating AI tools in scientific contexts, courses on AI for Scientists and Generative AI Courses provide grounding in both the capabilities and the failure modes of current systems.
Your membership also unlocks:
