Generative AI Literacy Training Improves Intelligence Analysts' Discrimination of Real and AI-Generated Images

Across social and online platforms, people are increasingly exposed to AI-generated images. As a consequence, the task of distinguishing AI-generated from authentic images is becoming a central challenge for information ecosystems. While humans perform better than chance, accuracy falls short of many operational needs. Initial evidence shows that visually oriented training can improve deepfake detection but does not improve participants' ability to identify real images as real. Here, we investigate the efficacy of a brief training intervention for intelligence analysts employed by the United States government in 2024. We conducted a counterbalanced within-subject randomized experiment in which we showed participants real and AI-generated images varying in pose complexity and scene context and asked them whether each image was real or AI-generated, both before and after an expert delivered a 30-minute training that pointed out patterns in seven real and 50 AI-generated images. We collected 2,544 image-level judgments from 32 intelligence analysts. We find training increased overall accuracy by 9 percentage points (95% CI: [2.7, 15.4]) from a baseline of 72%. We find the improvement is driven by a 14.2 percentage point increase in accuracy for real images (95% CI: [0.7, 27.7]). Through a careful experimental setup that curated matched pairs of real and AI-generated images across pose complexity categories, we reveal how these trainings influence people with different levels of digital forensics and generative AI experience and identify the kind of image-based content where this training intervention appears to be most effective. Ultimately, these results provide causal evidence that a brief, structured training can improve human judgment across a diverse array of real and AI-generated images, informing organizational responses to AI-generated visual misinformation.

Localizing RL-Induced Tool Use to a Single Crosscoder Feature

Fine-tuning through RL reshapes the internal representations of language models to enable agentic behaviors such as tool use, yet the mechanistic basis of these changes remains poorly understood. While RL substantially improves structured tool-call generation, it is unclear which features emerge, which are preserved, and whether identified features can be leveraged for retraining-free behavioral control. In this work, we show that $\textit{Dedicated Feature Crosscoders (DFC)}$ isolate a compact set of RL-specific features that mediate tool-calling capability in $\texttt{Qwen2.5-3B}$. Across a $48$-crosscoder hyperparameter sweep, encode-decode reconstruction improves the RL model's tool correctness by $+31.1 \pm {9.7}$ pp and passively transfers tool-calling ability to the frozen base model by $+6.8 \pm 5.0$ pp which we call a $\textit{capability spillover}$. Our findings show that DFC partitioning concentrates RL-introduced capability into a minimal, steerable feature set that enables runtime behavioral control of agentic LLMs.

Learning from the Right Rollouts: Data Attribution for PPO-based LLM Post-Training

Traditional RL algorithms like Proximal Policy Optimization (PPO) typically train on the entire rollout buffer, operating under the assumption that all generated episodes provide a beneficial optimization signal. However, these episodes frequently contain noisy or unfaithful reasoning, which can degrade model performance and slow down training. In this paper, we propose \textbf{Influence-Guided PPO (I-PPO)}, a novel framework that integrates data attribution into the RL post-training loop. By calculating an influence score for each episode using a gradient-based approximation, I-PPO identifies and eliminates episodes that are anti-aligned with a validation gradient. Our experiments demonstrate that I-PPO consistently outperforms SFT and PPO baselines. We show that our filtering process acts as an intrinsic early stopping mechanism, accelerating training efficiency while effectively reducing unfaithful CoT reasoning.

ComplLLM: Fine-tuning LLMs to Discover Complementary Signals for Decision-making

Multi-agent decision pipelines can outperform single agent workflows when complementarity holds, i.e., different agents bring unique information to the table to inform a final decision. We propose ComplLLM, a post-training framework based on decision theory that fine-tunes a decision-assistant LLM using complementary information as reward to output signals that complement existing agent decisions. We validate ComplLLM on synthetic and real-world tasks involving domain experts, demonstrating how the approach recovers known complementary information and produces plausible explanations of complementary signals to support downstream decision-makers.

This human study did not involve human subjects: Validating LLM simulations as behavioral evidence

A growing literature uses large language models (LLMs) as synthetic participants to generate cost-effective and nearly instantaneous responses in social science experiments. However, there is limited guidance on when such simulations support valid inference about human behavior. We contrast two strategies for obtaining valid estimates of causal effects and clarify the assumptions under which each is suitable for exploratory versus confirmatory research. Heuristic approaches seek to establish that simulated and observed human behavior are interchangeable through prompt engineering, model fine-tuning, and other repair strategies designed to reduce LLM-induced inaccuracies. While useful for many exploratory tasks, heuristic approaches lack the formal statistical guarantees typically required for confirmatory research. In contrast, statistical calibration combines auxiliary human data with statistical adjustments to account for discrepancies between observed and simulated responses. Under explicit assumptions, statistical calibration preserves validity and provides more precise estimates of causal effects at lower cost than experiments that rely solely on human participants. Yet the potential of both approaches depends on how well LLMs approximate the relevant populations. We consider what opportunities are overlooked when researchers focus myopically on substituting LLMs for human participants in a study.

Characterizing Photorealism and Artifacts in Diffusion Model-Generated Images

Diffusion model-generated images can appear indistinguishable from authentic photographs, but these images often contain artifacts and implausibilities that reveal their AI-generated provenance. Given the challenge to public trust in media posed by photorealistic AI-generated images, we conducted a large-scale experiment measuring human detection accuracy on 450 diffusion-model generated images and 149 real images. Based on collecting 749,828 observations and 34,675 comments from 50,444 participants, we find that scene complexity of an image, artifact types within an image, display time of an image, and human curation of AI-generated images all play significant roles in how accurately people distinguish real from AI-generated images. Additionally, we propose a taxonomy characterizing artifacts often appearing in images generated by diffusion models. Our empirical observations and taxonomy offer nuanced insights into the capabilities and limitations of diffusion models to generate photorealistic images in 2024.

The Value of Information in Human-AI Decision-making

Humans and AIs are often paired on decision tasks with the expectation of achieving complementary performance, where the combination of human and AI outperforms either one alone. However, how to improve performance of a human-AI team is often not clear without knowing more about what particular information and strategies each agent employs. We provide a decision-theoretic framework for characterizing the value of information -- and consequently, opportunities for agents to better exploit available information--in AI-assisted decision workflow. We demonstrate the use of the framework for model selection, empirical evaluation of human-AI performance, and explanation design. We propose a novel information-based instance-level explanation technique that adapts a conventional saliency-based explanation to explain information value in decision making.

A Computational Method for Measuring "Open Codes" in Qualitative Analysis

Qualitative analysis is critical to understanding human datasets in many social science disciplines. Open coding is an inductive qualitative process that identifies and interprets "open codes" from datasets. Yet, meeting methodological expectations (such as "as exhaustive as possible") can be challenging. While many machine learning (ML)/generative AI (GAI) studies have attempted to support open coding, few have systematically measured or evaluated GAI outcomes, increasing potential bias risks. Building on Grounded Theory and Thematic Analysis theories, we present a computational method to measure and identify potential biases from "open codes" systematically. Instead of operationalizing human expert results as the "ground truth," our method is built upon a team-based approach between human and machine coders. We experiment with two HCI datasets to establish this method's reliability by 1) comparing it with human analysis, and 2) analyzing its output stability. We present evidence-based suggestions and example workflows for ML/GAI to support open coding.

Measuring Free-Form Decision-Making Inconsistency of Language Models in Military Crisis Simulations

There is an increasing interest in using language models (LMs) for automated decision-making, with multiple countries actively testing LMs to aid in military crisis decision-making. To scrutinize relying on LM decision-making in high-stakes settings, we examine the inconsistency of responses in a crisis simulation ("wargame"), similar to reported tests conducted by the US military. Prior work illustrated escalatory tendencies and varying levels of aggression among LMs but were constrained to simulations with pre-defined actions. This was due to the challenges associated with quantitatively measuring semantic differences and evaluating natural language decision-making without relying on pre-defined actions. In this work, we query LMs for free form responses and use a metric based on BERTScore to measure response inconsistency quantitatively. Leveraging the benefits of BERTScore, we show that the inconsistency metric is robust to linguistic variations that preserve semantic meaning in a question-answering setting across text lengths. We show that all five tested LMs exhibit levels of inconsistency that indicate semantic differences, even when adjusting the wargame setting, anonymizing involved conflict countries, or adjusting the sampling temperature parameter $T$. Further qualitative evaluation shows that models recommend courses of action that share few to no similarities. We also study the impact of different prompt sensitivity variations on inconsistency at temperature $T = 0$. We find that inconsistency due to semantically equivalent prompt variations can exceed response inconsistency from temperature sampling for most studied models across different levels of ablations. Given the high-stakes nature of military deployment, we recommend further consideration be taken before using LMs to inform military decisions or other cases of high-stakes decision-making.

A Conceptual Framework for Ethical Evaluation of Machine Learning Systems

Research in Responsible AI has developed a range of principles and practices to ensure that machine learning systems are used in a manner that is ethical and aligned with human values. However, a critical yet often neglected aspect of ethical ML is the ethical implications that appear when designing evaluations of ML systems. For instance, teams may have to balance a trade-off between highly informative tests to ensure downstream product safety, with potential fairness harms inherent to the implemented testing procedures. We conceptualize ethics-related concerns in standard ML evaluation techniques. Specifically, we present a utility framework, characterizing the key trade-off in ethical evaluation as balancing information gain against potential ethical harms. The framework is then a tool for characterizing challenges teams face, and systematically disentangling competing considerations that teams seek to balance. Differentiating between different types of issues encountered in evaluation allows us to highlight best practices from analogous domains, such as clinical trials and automotive crash testing, which navigate these issues in ways that can offer inspiration to improve evaluation processes in ML. Our analysis underscores the critical need for development teams to deliberately assess and manage ethical complexities that arise during the evaluation of ML systems, and for the industry to move towards designing institutional policies to support ethical evaluations.