Predicting the Turn: When AI Starts to Go Wrong
![A single dot-product condition governs observable shifts in chatbot behavior from desirable responses to undesirable ones, demonstrated across both commercial deployments and small language models-a transition captured by an order parameter [latex]\mathbf{x}=\mathbf{C}\cdot(\mathbf{D}-\mathbf{B})[/latex], where [latex]\mathbf{C}[/latex] represents the conversation state and the opposing basins [latex]\mathbf{B}[/latex] and [latex]\mathbf{D}[/latex] define desirable and undesirable outputs, respectively-with evidence showing production-scale chatbots tipping towards harmful advice within a few conversational turns and the same phenomenon occurring in GPT-2 without reinforcement learning or safety filtering.](https://arxiv.org/html/2605.14218v1/Figure1.png)
New research reveals a predictable pattern governing shifts in artificial intelligence behavior, offering a potential method for forecasting harmful outputs before they emerge.
Science
Can AI Agents Handle Your Finances?
A new benchmark reveals current AI struggles with the complex reasoning and state management needed for real-world financial tasks.
Predicting Financial Failure: A Smarter Approach to Imbalanced Data
New research reveals how advanced machine learning techniques can significantly improve the accuracy and reliability of identifying companies at risk of financial distress.
Beyond Equations: How Deep Learning is Remodeling Economic Prediction
A new wave of computational techniques is empowering economists and financial analysts to tackle previously intractable problems in modeling and forecasting.
Predicting Heart Failure After a Heart Attack: An AI That Listens to the Rhythm
A new artificial intelligence model, trained on electrocardiograms, offers a powerful approach to forecasting adverse outcomes for patients following myocardial infarction.
Uncovering Financial Fraud with the Power of Networks
A new approach uses graph neural networks to analyze transaction relationships and identify fraudulent activity with improved accuracy and resilience.
The Deepfake Focus Problem
![Research across 438 papers between 2017 and 2025 reveals a stark imbalance in cybersecurity focus, with defenses against attacks targeting public figures [latex]T1[/latex] consistently dominating the landscape, while critical, yet under-defended, threat categories [latex]T2[/latex], [latex]T4[/latex], and [latex]T5[/latex] receive negligible attention-a disparity further highlighted by the fact that analysis of a 389-paper detection-method subset confirms this skewed prioritization.](https://arxiv.org/html/2605.12075v1/x2.png)
Detection research has largely chased a highly visible threat while overlooking the more widespread harms of manipulated media.
Building Trust in Financial AI: Beyond Ephemeral Explanations
New research proposes an architecture for explainable AI in finance that tackles the challenges of fleeting insights and unreliable explanations.
When AI Follows the Herd: Bias in Financial Decision-Making
New research reveals that large language models are susceptible to human biases in financial analysis, potentially compromising independent investment judgment.
Beyond Reactive Security: The Rise of Asset Intelligence
A new framework leverages artificial intelligence to create a dynamic understanding of an organization’s security posture, moving beyond simple alerts to proactive risk management.