Claude Safety Review Leads to Felony Charge Over Alleged Florida Threat

01Florida Woman Who Used Claude as a “Diary” Faces Felony Charge After Human Review of Threat Triggers Police Report

A Florida woman who said she used Anthropic’s Claude chatbot like a “diary” is facing a felony charge after its safety systems flagged an alleged threat and a human reviewer reported it to police. The case shows how writing entered into a chatbot can move from automated detection to emergency disclosure and produce real-world legal consequences.

According to an arrest report, Carli Michelle Heller of Bonita Springs wrote in Claude on September 26 that she would “shoot up” or otherwise attack the Sheriff’s office. The available reporting does not provide the full conversation or its surrounding context, and the alleged plan comes from the arrest report rather than a court finding.

Claude’s safety systems automatically flagged the entry and escalated it to a human reviewer. After examining the statements, the reviewer determined that they represented a credible threat and contacted law enforcement. The reporting does not disclose the reviewer’s specific criteria for reaching that conclusion.

Anthropic says it may share user information in limited emergency circumstances when it believes disclosure is necessary to prevent death or serious physical injury. That policy provides the stated basis for the report, but the case does not establish that all Claude conversations are routinely read by people. Here, human review followed an automated safety alert involving content judged to present an urgent risk.

Deputies then identified Heller and went to her home, where they detained her without incident. An intelligence detective from the Lee County Sheriff’s Office took over the investigation.

Heller now faces a charge of making a written threat of violence. Florida Statute 836.10 makes it a second-degree felony to send, post or transmit a written or electronic record threatening to kill or injure someone, conduct a mass shooting or commit an act of terrorism, provided the communication is made in a way that another person may view it.

The charge is an allegation, not a conviction. The source does not report a final ruling, explain how prosecutors will apply the statute to a chatbot entry, or say how Heller responded to the specific accusation beyond describing Claude as her diary. Those questions remain for the legal process.

02Whole-paper reasoning traces identify “AI scientific slop” with 85.9% accuracy in a 390-pair test

A scientific paper can contain plausible sentences and real citations while its overall reasoning is still unsound. A new study calls this failure “scientific slop”: AI-generated papers whose individual sections appear credible but whose structure, arguments and experimental materials do not form a coherent scientific case.

Existing token-based detectors look mainly for linguistic signals associated with generated text, making them poorly suited to finding inconsistencies that span an entire paper. The researchers instead developed six measures covering three areas—structure, argument and artifacts—to test whether the parts of a paper support one another.

They evaluated those measures on SciSlopBench, a benchmark containing 390 AI-generated papers. Each was paired with a human-written paper matched by research problem and type of contribution. Most of the papers concern computer science, although the benchmark also covers the life, social and natural sciences.

In this paired test, the measures correctly identified the AI-generated paper 85.9% of the time, according to the authors. Binoculars, an existing AI-text detector, reached 68.7%. The comparison suggests that whole-paper reasoning traces can reveal problems missed by token-level detection, but it does not establish an 85.9% detection rate for arbitrary papers or real-world submissions. The source also provides no independent replication or real-world false-positive rate.

The authors report that higher scientific-slop scores accompanied lower ICLR review ratings. Their measures also distinguished rejected from accepted papers at above-chance levels in every year from 2017 through 2025. That is an association, not evidence that AI-generated content caused the lower ratings.

Fixing the measured problems was not as simple as telling a model to improve its paper. Standard revisions left residual slop, while prompts that directly optimized for the new measures encouraged “reward hacking”—improving the score without necessarily repairing the underlying science.

The proposed SciSlopHarness addresses that problem by allowing a fixed large language model to revise an issue only when experimental records support the change. Without using human-written papers as reference targets, the framework reduced the remaining gap between AI and human papers by 63% relative to the strongest revision baseline, the authors report. The study does not disclose detailed definitions for all six measures or the full generation process behind the 390 AI papers.

03MotorMind Lets a General Vision-Language Model Control Robots Zero-Shot, Reports 95% Average Success in Real xArm6 Experiments

General-purpose vision-language models can interpret images and reason about instructions, but reliably turning that reasoning into robot movements remains difficult. Existing vision-language-action systems have limited zero-shot performance on unfamiliar tasks and environments, while other robotic agents often depend on specially trained policies, coding agents, action experts or visual grounding tools.

MotorMind, a new robot-manipulation framework, attempts to close that gap without training a separate policy for each task. It places a mid-level action interface between a general vision-language model and the robot: the model observes the scene and proposes actions that are more concrete than a high-level plan but do not directly specify every low-level motor command. A deterministic control module executes those actions and returns feedback to the model.

The execution loop also includes asynchronous monitoring and background memory updates, allowing the system to continue assessing progress and adapt to feedback. According to the paper’s authors, MotorMind does not require task-specific policy training, coding agents, learned action experts or additional grounding tools such as SAM3, which would otherwise identify and locate objects for the system.

In the authors’ zero-shot evaluations, MotorMind achieved a 66.7% success rate on the base LIBERO-PRO robot-manipulation suites. The best previously evaluated zero-shot method reached 13.3%. Under perturbations, MotorMind succeeded in 53.8% of trials, compared with a maximum of 19.2% among the comparison methods.

The researchers then used the same interface with a real xArm6 robotic arm. Across direct-manipulation tasks and settings in which a person introduced disturbances, they report an average success rate of 95%. That figure applies only to the tested tasks and conditions; the source does not disclose the number of tasks or trials, the vision-language model used, operating cost or action latency, and it provides no independent replication.

Performance improved further when the researchers replaced the system’s backbone with a stronger vision-language model. The remaining failures were concentrated in visual grounding, embodied reasoning and action knowledge, and the authors report that these failures decreased as the underlying model became more capable.