AI Daily (21 Nov) – Students sue after AI gun alert; Google adds Gemini tool retrieval

•

Today’s top line: A false AI gun-detection alert led to an armed police response and a lawsuit; Google previewed Tool Retrieval for Gemini agents to scale large tool libraries and cut token and latency costs.

Students sue after AI gun detector triggered armed police response

Summary: Four current and former students filed a lawsuit after a school’s AI gun-detection system, Omnilert, generated a false alert that staff briefly cancelled but then led to police being called and arriving with at least six patrol cars. The complaint says officers drew guns, handcuffed and searched the students, and that the plaintiffs seek an injunction barring the district from using Omnilert until protocols prevent police responses to false alerts.

Digest: The case highlights how automated safety systems can trigger real-world force when human review and response protocols fail, raising legal and policy questions for schools, vendors and police departments.

Source: Ars Technica

Google previews Tool Retrieval for Gemini API to scale agents

Summary: Google launched Tool Retrieval in preview for the Gemini Interactions API so agents can keep large tool libraries on standby and load parameter schemas only when needed. Developers enable it by adding a tool_search tool and setting defer_loading: true; the model still sees tool names and descriptions and requests schemas when required, all inside one API call.

Digest: Tool Retrieval reduces input tokens, latency and tool-selection errors as agents scale to dozens or hundreds of tools, making production agent systems more efficient and easier to debug.

Source: Google

LLMs find most clinical prediction studies do not share reusable code

Summary: An international team led by MIT used large language models to screen nearly 4,000 clinical prediction studies and then inspected hundreds of linked code repositories. They found code sharing remains uncommon and that making code public often does not include the documentation and supporting information needed for reuse; the results inform a new reporting guideline, TRIPOD-Code, published in Nature Medicine.

Digest: Poor code sharing undermines reproducibility for models that can influence patient care; hospitals, journals and regulators should watch TRIPOD-Code adoption to improve transparency and safety.

Source: MIT News

Authorization-bounded agent closed analysis-to-design gap in self-driving lab demo

Summary: An arXiv paper presents an architecture that lets an AI agent plan measurements and interpret results inside the Design-Make-Test-Analyze (DMTA) cycle while keeping state-changing operations authorised, validated and auditable. Demonstrated on a liquid chromatography-mass spectrometry (LC-MS) campaign, the agent submitted jobs, interpreted mass spectrometry and diode-array data, flagged failures and produced an oxidant ranking that matched scripted peak integration; in a replayed-data simulation it found the highest-yielding condition at about its seventh injection versus 14.5 on average for random order.

Digest: The work shows agents can accelerate experimental optimisation while enforcing human-set limits and traceability, a key step toward safer autonomous laboratories and faster discovery.

Source: arXiv

LLM agent framework automates industrial maintenance work orders

Summary: A new arXiv paper introduces CodeReAct, an agentic framework that interprets natural-language instructions, extracts knowledge, summarises asset status and recommends work orders by standardising industrial data into business objects. The agent’s Action phase generates executable code, replacing manual coding; a chiller maintenance case study showed the system can analyse years of work orders, alerts and anomalies and reduce the need for data scientist involvement.

Digest: Firms running large maintenance fleets could cut analysis time and staffing needs, but deployments will need strong validation and integration with existing operational systems.

Source: arXiv

Hugging Face releases harness adapters to improve model performance and cost

Summary: Hugging Face published a recipe and artifacts for RL-trained LoRA harness adapters that tune a frozen base model to a specific agent harness and task. For Qwen3.8-27B under the pi harness, an adapter improved Terminal-Bench 2.1 pass@1 from 39% to 55% and reduced average token usage from 2.9K to 1.5K; the team trained one adapter per harness (Pi, Hermes Agent, OpenCode) and merged them into a single super-adapter, with code and models available on the Hub.

Digest: Organisations can boost accuracy and cut inference costs without retraining full models by training lightweight adapters inside their deployment harness; watch for adapter sharing and security reviews.

Source: Hugging Face

Quick hits