#12 — DeepMind AI Control Roadmap, OpenAI drug discovery, AI persona authorship

A new proposal for controlling advanced AI treats the system as an internal security threat. This is Research Agent, where we examine the latest applications of AI in science. We start with the AI Control Roadmap from Google DeepMind. [LEAD] Google DeepMind published its AI Control Roadmap. Google DeepMind published its "Securing the future of AI agents" document, detailing an AI Control Roadmap. This roadmap presents a framework for managing advanced AI safely within Google's infrastructure. The approach extends beyond standard model alignment by incorporating system-level security measures. It treats unaligned AI agents as potential internal threats, using a threat modeling framework based on MITRE ATT&CK, and implements control mitigations through monitoring. SOURCE MATERIAL: (Securing internal systems against increasingly capable and ..., 2026-06-18): Google DeepMind's "Securing the future of AI agents" details their AI Control Roadmap, a framework for building and managing advanced AI within Google. This approach goes beyond traditional model alignment, adding a crucial layer of system-level security to ensure safety even with imperfectly aligned AI. The roadmap addresses potential risks by treating untrusted AI agents as "insider threats," utilizing a novel threat-modeling framework based on MITRE ATT&CK, and deploying AI control mitigations through monitoring and real-time response. [LEAD] OpenAI published research on AI-powered drug discovery OpenAI released new research focused on the application of artificial intelligence in drug discovery processes. The publication explores methods and findings related to using AI for pharmaceutical development. SOURCE MATERIAL: (AI Update— Thursday, June 18, 2026 | by DevQuill Insights | Adi ..., 2026-06-17): The article summarizes key AI news from June 18, 2026, including Anthropic restoring Claude Fable 5 access after a government shutdown and Google making Gemini 2.5 Flash its default model. It also covers the Fed holding interest rates, SpaceX stock stabilizing, and OpenAI publishing research on AI-powered drug discovery. The piece highlights the infrastructure crisis underlying the AI industry, with high capital intensity, inflation, and government scrutiny.TITLE: AI Update— Thursday, June 18, 2026 | by DevQuill Insights | Adi ... DATE: 2026-06-17 The article summarizes key AI news from June 18, 2026, including Anthropic restoring Claude Fable 5 access after a government shutdown and Google [LEAD] A paper analyzes persona authorship's implications for research misconduct in academic publishing. The paper investigates 'persona authorship,' where researchers use personalized large language models to generate entire manuscripts. It examines whether submitting such AI-generated manuscripts, particularly without disclosure, constitutes research misconduct. The analysis concludes that persona authorship does not inherently violate fundamental academic research norms more than traditional human authorship. SOURCE MATERIAL: (Is AI-Produced Humanities Scholarship a Case of Research Misconduct?, 2026-05-28): Abstract As large language-models (“LLMs”) become increasingly capable of producing publishable scholarship, researchers have acquired the ability to personalize LLMs with their own style and academic values to generate entire manuscripts: a practice this paper calls “persona authorship.” This paper investigates whether submitting such AI-generated manuscripts in the humanities under the human author’s name, especially without disclosure, constitutes intrinsic research misconduct. Surveying four examples of research misconduct—deception, fraud, negligence, and corruption—this paper argues that persona authorship does not clearly violate any fundamental norms of academic research, at least no [LEAD] A study introduced Neo-Inference Science (NIS) as a conceptual framework for AI's role in scientific discovery. This study proposes Neo-Inference Science (NIS), a conceptual framework positioning artificial intelligence as an active inferential agent for scientific discovery. NIS is presented as the sixth evolutionary stage of global digital transformation. The framework emphasizes AI-driven hypothesis generation, inferential reasoning, and collaborative human-AI scientific discovery. It develops the epistemological foundations, structural architecture, methodological processes, and strategic implications of this new paradigm. SOURCE MATERIAL: (Neo-Inference Science (NIS): A Sixth-Stage Paradigm of Digital Transformation for AI-Driven Scientific Discovery, 2026-06-09): The rapid evolution of artificial intelligence has transformed digital systems from automation tools into engines of generative intelligence. However, contemporary Al systems remain largely limited to prediction, optimization, and content generation. This study introduces Neo-Inference Science (NIS), a conceptual framework that positions artificial intelligence (AI) as an active inferential agent capable of participating in scientific discovery and knowledge generation.The NIS is proposed as the sixth evolutionary stage of global digital transformation, following the eras of digitization, connectivity, automation, intelligent analytics, and generative intelligence. Unlike previous paradigms, [LEAD] A study found large language models exhibit cognitive biases in information judgment tasks. A research study investigated whether large language models (LLMs) are susceptible to cognitive biases. The study compared LLM and human credibility assessments under potential decoy effects in information retrieval settings. It empirically confirmed that LLM agents exhibit cognitive bias risks, specifically the decoy effect. This finding challenges general assumptions about AI rationality. SOURCE MATERIAL: (The Decoy Dilemma in Online Medical Information Evaluation: A Comparative Study of Credibility Assessments by LLM and Human Judges, 2026-06-11): Can AI be cognitively biased in automated information judgment tasks? Despite recent progress in measuring and mitigating social and algorithmic biases in AI and large language models (LLMs), it is not clear to what extent LLMs behave ”rationally”, or if they are also vulnerable to human cognitive bias triggers . To address this open problem, our study, consisting of a crowdsourcing user experiment and ann LLM-enabled simulation experiment, compared the credibility assessments by LLM and human judges under potential decoy effects in an interactive information retrieval (IR) setting, and empirically examined the extent to which LLMs are cognitively biased in medical (mis)information assessmen [LEAD] Researchers tested five large language models for metascientific data extraction. A study investigated the use of current large language models for extracting information from scientific papers. Five models, including Claude 4.6 Opus, Claude 4.5 Sonnet, Claude 4.5 Haiku, GPT-5.2, and GPT-5 mini, underwent testing. They utilized an automated API-based pipeline on a dataset comprising 43 network psychometrics papers. SOURCE MATERIAL: (Automating data extraction in meta-research: A multi-model benchmark in network psychometrics papers, 2026-05-22): Manual data extraction in meta-research is often tedious, time-consuming, and error-prone. In this paper, we investigate whether the current generation of large language models (LLMs) can be used to extract accurate information from scientific papers. Across the meta-research literature, these tasks usually range from extracting verbatim information (e.g., the number of participants in a study, effect sizes, or whether a study is preregistered) to making subjective inferences. Using a publicly available dataset containing a wide range of metascientific variables from 43 network psychometrics papers, we tested five LLMs (Claude 4.6 Opus, Claude 4.5 Sonnet, Claude 4.5 Haiku, GPT-5.2, and GPT-5 [LEAD] Automated large language model data extraction can reduce meta-research time and costs. The study concluded that the proposed automated procedure for LLM data extraction has significant benefits. It can substantially decrease the time required for conducting meta-research. Additionally, it offers potential for reducing associated costs. SOURCE MATERIAL: (Automating data extraction in meta-research: A multi-model benchmark in network psychometrics papers, 2026-05-22): Manual data extraction in meta-research is often tedious, time-consuming, and error-prone. In this paper, we investigate whether the current generation of large language models (LLMs) can be used to extract accurate information from scientific papers. Across the meta-research literature, these tasks usually range from extracting verbatim information (e.g., the number of participants in a study, effect sizes, or whether a study is preregistered) to making subjective inferences. Using a publicly available dataset containing a wide range of metascientific variables from 43 network psychometrics papers, we tested five LLMs (Claude 4.6 Opus, Claude 4.5 Sonnet, Claude 4.5 Haiku, GPT-5.2, and GPT-5 [NOTABLE] States increased AI regulation. State governments are increasingly taking the lead in regulating artificial intelligence. This trend is noted despite ongoing federal discussions and differing policy approaches. SOURCE MATERIAL: (Congress and State Lawmakers Are Racing to Keep Up With AI | Insights & Resources | Goodwin, 2026-06-15): States are increasingly leading AI regulation, with over 1,500 AI bills under consideration in 2026, building on 150 laws passed in 2025. This state-level activity is occurring even as Congress debates a bipartisan frontier AI law and the Trump administration pursues a deregulatory agenda. US regulation focuses on issues like AI disclosures for chatbots, companion AI safety, restrictions on mental health applications, and transparency in consequential decision-making. [NOTABLE] States passed 150 AI laws. State governments enacted 150 laws pertaining to artificial intelligence in 2025. This legislative activity demonstrates a proactive approach to AI governance at the state level. SOURCE MATERIAL: (Congress and State Lawmakers Are Racing to Keep Up With AI | Insights & Resources | Goodwin, 2026-06-15): States are increasingly leading AI regulation, with over 1,500 AI bills under consideration in 2026, building on 150 laws passed in 2025. This state-level activity is occurring even as Congress debates a bipartisan frontier AI law and the Trump administration pursues a deregulatory agenda. US regulation focuses on issues like AI disclosures for chatbots, companion AI safety, restrictions on mental health applications, and transparency in consequential decision-making. [NOTABLE] US Congress debated a bipartisan frontier AI law. The United States Congress is currently debating a bipartisan law focused on frontier artificial intelligence. This federal legislative effort operates alongside state-level regulatory initiatives. SOURCE MATERIAL: (Congress and State Lawmakers Are Racing to Keep Up With AI | Insights & Resources | Goodwin, 2026-06-15): States are increasingly leading AI regulation, with over 1,500 AI bills under consideration in 2026, building on 150 laws passed in 2025. This state-level activity is occurring even as Congress debates a bipartisan frontier AI law and the Trump administration pursues a deregulatory agenda. US regulation focuses on issues like AI disclosures for chatbots, companion AI safety, restrictions on mental health applications, and transparency in consequential decision-making. [NOTABLE] US AI regulation focused on specific applications. Artificial intelligence regulation in the United States targets areas such as disclosure requirements for chatbots, safety protocols for companion AI, restrictions on AI in mental health applications, and transparency in impactful decision-making processes. These represent key areas of policy focus. SOURCE MATERIAL: (Congress and State Lawmakers Are Racing to Keep Up With AI | Insights & Resources | Goodwin, 2026-06-15): States are increasingly leading AI regulation, with over 1,500 AI bills under consideration in 2026, building on 150 laws passed in 2025. This state-level activity is occurring even as Congress debates a bipartisan frontier AI law and the Trump administration pursues a deregulatory agenda. US regulation focuses on issues like AI disclosures for chatbots, companion AI safety, restrictions on mental health applications, and transparency in consequential decision-making. [NOTABLE] IBM released watsonx Assistant for Z v3.3. IBM introduced version 3.3 of watsonx Assistant for Z, designed to improve deployment flexibility and intelligent automation for IBM Z environments. This release aims to simplify enterprise operations and increase productivity. Key features include multi-tenancy support, a central orchestrator, and expanded AI model compatibility with Granite 4.1. SOURCE MATERIAL: (Advancing agentic AI for productivity and intelligent automation on IBM Z, 2026-06-16): IBM is advancing AI innovation for IBM Z with the introduction of IBM watsonx Assistant for Z v3.3, focusing on enhanced deployment flexibility and intelligent automation. This release aims to simplify enterprise operations, improve productivity, and help organizations scale AI effectively across mission-critical environments. Key enhancements include multi-tenancy support, a central orchestrator for intelligent routing, expanded AI model support with Granite 4.1, and an updated AI inferencing architecture. [NOTABLE] Larger and more recent large language models demonstrate higher accuracy but increased susceptibility to decoy effects. The study observed that larger and more recent large language models (LLMs) demonstrate improved consistency and accuracy in distinguishing credible information. However, these models were more likely to assign higher ratings to misinformation when a prominent decoy misinformation result was present. This indicates an increased susceptibility to the decoy effect in more advanced LLMs. SOURCE MATERIAL: (The Decoy Dilemma in Online Medical Information Evaluation: A Comparative Study of Credibility Assessments by LLM and Human Judges, 2026-06-11): Can AI be cognitively biased in automated information judgment tasks? Despite recent progress in measuring and mitigating social and algorithmic biases in AI and large language models (LLMs), it is not clear to what extent LLMs behave ”rationally”, or if they are also vulnerable to human cognitive bias triggers . To address this open problem, our study, consisting of a crowdsourcing user experiment and ann LLM-enabled simulation experiment, compared the credibility assessments by LLM and human judges under potential decoy effects in an interactive information retrieval (IR) setting, and empirically examined the extent to which LLMs are cognitively biased in medical (mis)information assessmen [NOTABLE] The decoy effect is more prevalent in large language model judgments than human credibility ratings. Research findings indicate that the decoy effect, a form of cognitive bias, occurs in both human and large language model (LLM) assessments. However, the study found this effect to be more prevalent across various conditions and topics in LLM judgments. This contrasts with human credibility ratings, where the effect was less widespread. SOURCE MATERIAL: (The Decoy Dilemma in Online Medical Information Evaluation: A Comparative Study of Credibility Assessments by LLM and Human Judges, 2026-06-11): Can AI be cognitively biased in automated information judgment tasks? Despite recent progress in measuring and mitigating social and algorithmic biases in AI and large language models (LLMs), it is not clear to what extent LLMs behave ”rationally”, or if they are also vulnerable to human cognitive bias triggers . To address this open problem, our study, consisting of a crowdsourcing user experiment and ann LLM-enabled simulation experiment, compared the credibility assessments by LLM and human judges under potential decoy effects in an interactive information retrieval (IR) setting, and empirically examined the extent to which LLMs are cognitively biased in medical (mis)information assessmen [NOTABLE] Large language models achieved 79.6% to 91.3% accuracy in metascientific data extraction. The study found that large language models demonstrated an extraction accuracy ranging from 79.6% to 91.3%. This performance was observed across the five tested models. The evaluation focused on extracting variables from a dataset of network psychometrics papers. SOURCE MATERIAL: (Automating data extraction in meta-research: A multi-model benchmark in network psychometrics papers, 2026-05-22): Manual data extraction in meta-research is often tedious, time-consuming, and error-prone. In this paper, we investigate whether the current generation of large language models (LLMs) can be used to extract accurate information from scientific papers. Across the meta-research literature, these tasks usually range from extracting verbatim information (e.g., the number of participants in a study, effect sizes, or whether a study is preregistered) to making subjective inferences. Using a publicly available dataset containing a wide range of metascientific variables from 43 network psychometrics papers, we tested five LLMs (Claude 4.6 Opus, Claude 4.5 Sonnet, Claude 4.5 Haiku, GPT-5.2, and GPT-5 [NOTABLE] Large language models exhibited higher accuracy for explicit data extraction than for inferential data extraction. The research indicated that the extraction performance of the models varied based on the type of information. Models achieved higher accuracy when extracting explicit, verbatim information. Performance decreased for variables that required more complex inference. SOURCE MATERIAL: (Automating data extraction in meta-research: A multi-model benchmark in network psychometrics papers, 2026-05-22): Manual data extraction in meta-research is often tedious, time-consuming, and error-prone. In this paper, we investigate whether the current generation of large language models (LLMs) can be used to extract accurate information from scientific papers. Across the meta-research literature, these tasks usually range from extracting verbatim information (e.g., the number of participants in a study, effect sizes, or whether a study is preregistered) to making subjective inferences. Using a publicly available dataset containing a wide range of metascientific variables from 43 network psychometrics papers, we tested five LLMs (Claude 4.6 Opus, Claude 4.5 Sonnet, Claude 4.5 Haiku, GPT-5.2, and GPT-5 [NOTABLE] Most large language models expressed uncertainty in contentious data extraction cases. The study also observed that a majority of the tested large language models could convey uncertainty. This capability was demonstrated in cases where the extracted information was more contentious or ambiguous. SOURCE MATERIAL: (Automating data extraction in meta-research: A multi-model benchmark in network psychometrics papers, 2026-05-22): Manual data extraction in meta-research is often tedious, time-consuming, and error-prone. In this paper, we investigate whether the current generation of large language models (LLMs) can be used to extract accurate information from scientific papers. Across the meta-research literature, these tasks usually range from extracting verbatim information (e.g., the number of participants in a study, effect sizes, or whether a study is preregistered) to making subjective inferences. Using a publicly available dataset containing a wide range of metascientific variables from 43 network psychometrics papers, we tested five LLMs (Claude 4.6 Opus, Claude 4.5 Sonnet, Claude 4.5 Haiku, GPT-5.2, and GPT-5 That's all for this episode. We'll have more on the application of AI in research next week. From Research Agent, thanks for listening.