Matthew Groh

Ziv Epstein, Aaron Hertzmann, Laura Herman et al.

A new class of tools, colloquially called generative AI, can produce high-quality artistic media for visual arts, concept art, music, fiction, literature, video, and animation. The generative capabilities of these tools are likely to fundamentally alter the creative processes by which creators formulate ideas and put them into production. As creativity is reimagined, so too may be many sectors of society. Understanding the impact of generative AI - and making policy decisions around it - requires new interdisciplinary scientific inquiry into culture, economics, law, algorithms, and the interaction of technology and creativity. We argue that generative AI is not the harbinger of art's demise, but rather is a new medium with its own distinct affordances. In this vein, we consider the impacts of this new medium on creators across four themes: aesthetics and culture, legal questions of ownership and credit, the future of creative work, and impacts on the contemporary media ecosystem. Across these themes, we highlight key research questions and directions to inform policy and beneficial uses of the technology.

Luke W. Sagers, James A. Diao, Matthew Groh et al.

Dermatological classification algorithms developed without sufficiently diverse training data may generalize poorly across populations. While intentional data collection and annotation offer the best means for improving representation, new computational approaches for generating training data may also aid in mitigating the effects of sampling bias. In this paper, we show that DALL$\cdot$E 2, a large-scale text-to-image diffusion model, can produce photorealistic images of skin disease across skin types. Using the Fitzpatrick 17k dataset as a benchmark, we demonstrate that augmenting training data with DALL$\cdot$E 2-generated synthetic images improves classification of skin disease overall and especially for underrepresented groups.

Luke W. Sagers, James A. Diao, Luke Melas-Kyriazi et al.

While hundreds of artificial intelligence (AI) algorithms are now approved or cleared by the US Food and Drugs Administration (FDA), many studies have shown inconsistent generalization or latent bias, particularly for underrepresented populations. Some have proposed that generative AI could reduce the need for real data, but its utility in model development remains unclear. Skin disease serves as a useful case study in synthetic image generation due to the diversity of disease appearance, particularly across the protected attribute of skin tone. Here we show that latent diffusion models can scalably generate images of skin disease and that augmenting model training with these data improves performance in data-limited settings. These performance gains saturate at synthetic-to-real image ratios above 10:1 and are substantially smaller than the gains obtained from adding real images. As part of our analysis, we generate and analyze a new dataset of 458,920 synthetic images produced using several generation strategies. Our results suggest that synthetic data could serve as a force-multiplier for model development, but the collection of diverse real-world data remains the most important step to improve medical AI algorithms.

Matthew Groh, Caleb Harris, Roxana Daneshjou et al.

While artificial intelligence (AI) holds promise for supporting healthcare providers and improving the accuracy of medical diagnoses, a lack of transparency in the composition of datasets exposes AI models to the possibility of unintentional and avoidable mistakes. In particular, public and private image datasets of dermatological conditions rarely include information on skin color. As a start towards increasing transparency, AI researchers have appropriated the use of the Fitzpatrick skin type (FST) from a measure of patient photosensitivity to a measure for estimating skin tone in algorithmic audits of computer vision applications including facial recognition and dermatology diagnosis. In order to understand the variability of estimated FST annotations on images, we compare several FST annotation methods on a diverse set of 460 images of skin conditions from both textbooks and online dermatology atlases. We find the inter-rater reliability between three board-certified dermatologists is comparable to the inter-rater reliability between the board-certified dermatologists and two crowdsourcing methods. In contrast, we find that the Individual Typology Angle converted to FST (ITA-FST) method produces annotations that are significantly less correlated with the experts' annotations than the experts' annotations are correlated with each other. These results demonstrate that algorithms based on ITA-FST are not reliable for annotating large-scale image datasets, but human-centered, crowd-based protocols can reliably add skin type transparency to dermatology datasets. Furthermore, we introduce the concept of dynamic consensus protocols with tunable parameters including expert review that increase the visibility of crowdwork and provide guidance for future crowdsourced annotations of large image datasets.

Valdemar Danry, Pat Pataranutaporn, Matthew Groh et al.

Advanced Artificial Intelligence (AI) systems, specifically large language models (LLMs), have the capability to generate not just misinformation, but also deceptive explanations that can justify and propagate false information and erode trust in the truth. We examined the impact of deceptive AI generated explanations on individuals' beliefs in a pre-registered online experiment with 23,840 observations from 1,192 participants. We found that in addition to being more persuasive than accurate and honest explanations, AI-generated deceptive explanations can significantly amplify belief in false news headlines and undermine true ones as compared to AI systems that simply classify the headline incorrectly as being true/false. Moreover, our results show that personal factors such as cognitive reflection and trust in AI do not necessarily protect individuals from these effects caused by deceptive AI generated explanations. Instead, our results show that the logical validity of AI generated deceptive explanations, that is whether the explanation has a causal effect on the truthfulness of the AI's classification, plays a critical role in countering their persuasiveness - with logically invalid explanations being deemed less credible. This underscores the importance of teaching logical reasoning and critical thinking skills to identify logically invalid arguments, fostering greater resilience against advanced AI-driven misinformation.

Matthew Groh

Machine learning models are often brittle on production data despite achieving high accuracy on benchmark datasets. Benchmark datasets have traditionally served dual purposes: first, benchmarks offer a standard on which machine learning researchers can compare different methods, and second, benchmarks provide a model, albeit imperfect, of the real world. The incompleteness of test benchmarks (and the data upon which models are trained) hinder robustness in machine learning, enable shortcut learning, and leave models systematically prone to err on out-of-distribution and adversarially perturbed data. The mismatch between a single static benchmark dataset and a production dataset has traditionally been described as a dataset shift. In an effort to clarify how to address the mismatch between test benchmarks and production data, we introduce context shift to describe semantically meaningful changes in the underlying data generation process. Moreover, we identify three methods for addressing context shift that would otherwise lead to model prediction errors: first, we describe how human intuition and expert knowledge can identify semantically meaningful features upon which models systematically fail, second, we detail how dynamic benchmarking - with its focus on capturing the data generation process - can promote generalizability through corroboration, and third, we highlight that clarifying a model's limitations can reduce unexpected errors. Robust machine learning is focused on model performance beyond benchmarks, and as such, we consider three model organism domains - facial expression recognition, deepfake detection, and medical diagnosis - to highlight how implicit assumptions in benchmark tasks lead to errors in practice. By paying close attention to the role of context, researchers can design more comprehensive benchmarks, reduce context shift errors, and increase generalizability.

Matthew Groh, Craig Ferguson, Robert Lewis et al.

How does empathy influence creative problem solving? We introduce a computational empathy intervention based on context-specific affective mimicry and perspective taking by a virtual agent appearing in the form of a well-dressed polar bear. In an online experiment with 1,006 participants randomly assigned to an emotion elicitation intervention (with a control elicitation condition and anger elicitation condition) and a computational empathy intervention (with a control virtual agent and an empathic virtual agent), we examine how anger and empathy influence participants' performance in solving a word game based on Wordle. We find participants who are assigned to the anger elicitation condition perform significantly worse on multiple performance metrics than participants assigned to the control condition. However, we find the empathic virtual agent counteracts the drop in performance induced by the anger condition such that participants assigned to both the empathic virtual agent and the anger condition perform no differently than participants in the control elicitation condition and significantly better than participants assigned to the control virtual agent and the anger elicitation condition. While empathy reduces the negative effects of anger, we do not find evidence that the empathic virtual agent influences performance of participants who are assigned to the control elicitation condition. By introducing a framework for computational empathy interventions and conducting a two-by-two factorial design randomized experiment, we provide rigorous, empirical evidence that computational empathy can counteract the negative effects of anger on creative problem solving.

Fabian Gröger, Simone Lionetti, Philippe Gottfrois et al.

Benchmark datasets for digital dermatology unwittingly contain inaccuracies that reduce trust in model performance estimates. We propose a resource-efficient data-cleaning protocol to identify issues that escaped previous curation. The protocol leverages an existing algorithmic cleaning strategy and is followed by a confirmation process terminated by an intuitive stopping criterion. Based on confirmation by multiple dermatologists, we remove irrelevant samples and near duplicates and estimate the percentage of label errors in six dermatology image datasets for model evaluation promoted by the International Skin Imaging Collaboration. Along with this paper, we publish revised file lists for each dataset which should be used for model evaluation. Our work paves the way for more trustworthy performance assessment in digital dermatology.

Fabian Gröger, Simone Lionetti, Philippe Gottfrois et al.

The adoption of artificial intelligence in dermatology promises democratized access to healthcare, but model reliability depends on the quality and comprehensiveness of the data fueling these models. Despite rapid growth in publicly available dermatology images, the field lacks quantitative key performance indicators to measure whether new datasets expand clinical coverage or merely replicate what is already known. Here we present SkinMap, a multi-modal framework for the first comprehensive audit of the field's entire data basis. We unify the publicly available dermatology datasets into a single, queryable semantic atlas comprising more than 1.1 million images of skin conditions and quantify (i) informational novelty over time, (ii) dataset redundancy, and (iii) representation gaps across demographics and diagnoses. Despite exponential growth in dataset sizes, informational novelty across time has somewhat plateaued: Some clusters, such as common neoplasms on fair skin, are densely populated, while underrepresented skin types and many rare diseases remain unaddressed. We further identify structural gaps in coverage: Darker skin tones (Fitzpatrick V-VI) constitute only 5.8% of images and pediatric patients only 3.0%, while many rare diseases and phenotype combinations remain sparsely represented. SkinMap provides infrastructure to measure blind spots and steer strategic data acquisition toward undercovered regions of clinical space.

Fabian Gröger, Simone Lionetti, Philippe Gottfrois et al.

Benchmark datasets in computer vision often contain off-topic images, near duplicates, and label errors, leading to inaccurate estimates of model performance. In this paper, we revisit the task of data cleaning and formalize it as either a ranking problem, which significantly reduces human inspection effort, or a scoring problem, which allows for automated decisions based on score distributions. We find that a specific combination of context-aware self-supervised representation learning and distance-based indicators is effective in finding issues without annotation biases. This methodology, which we call SelfClean, surpasses state-of-the-art performance in detecting off-topic images, near duplicates, and label errors within widely-used image datasets, such as ImageNet-1k, Food-101N, and STL-10, both for synthetic issues and real contamination. We apply the detailed method to multiple image benchmarks, identify up to 16% of issues, and confirm an improvement in evaluation reliability upon cleaning. The official implementation can be found at: https://github.com/Digital-Dermatology/SelfClean.

Matthew Groh, Caleb Harris, Luis Soenksen et al.

How does the accuracy of deep neural network models trained to classify clinical images of skin conditions vary across skin color? While recent studies demonstrate computer vision models can serve as a useful decision support tool in healthcare and provide dermatologist-level classification on a number of specific tasks, darker skin is underrepresented in the data. Most publicly available data sets do not include Fitzpatrick skin type labels. We annotate 16,577 clinical images sourced from two dermatology atlases with Fitzpatrick skin type labels and open-source these annotations. Based on these labels, we find that there are significantly more images of light skin types than dark skin types in this dataset. We train a deep neural network model to classify 114 skin conditions and find that the model is most accurate on skin types similar to those it was trained on. In addition, we evaluate how an algorithmic approach to identifying skin tones, individual typology angle, compares with Fitzpatrick skin type labels annotated by a team of human labelers.

Negar Kamali, Candice Rockell Gerstner, Jessica Hullman et al.

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.

Negar Kamali, Karyn Nakamura, Aakriti Kumar et al.

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.

Aakriti Kumar, Nalin Poungpeth, Diyi Yang et al.

Large language models (LLMs) excel at generating empathic responses in text-based conversations. But, how reliably do they judge the nuances of empathic communication? We investigate this question by comparing how experts, crowdworkers, and LLMs annotate empathic communication across four evaluative frameworks drawn from psychology, natural language processing, and communications applied to 200 real-world conversations where one speaker shares a personal problem and the other offers support. Drawing on 3,150 expert annotations, 2,844 crowd annotations, and 3,150 LLM annotations, we assess inter-rater reliability between these three annotator groups. We find that expert agreement is high but varies across the frameworks' sub-components depending on their clarity, complexity, and subjectivity. We show that expert agreement offers a more informative benchmark for contextualizing LLM performance than standard classification metrics. Across all four frameworks, LLMs consistently approach this expert level benchmark and exceed the reliability of crowdworkers. These results demonstrate how LLMs, when validated on specific tasks with appropriate benchmarks, can support transparency and oversight in emotionally sensitive applications including their use as conversational companions.

Xuhai Xu, Haoyu Hu, Haoran Zhang et al.

Artificial intelligence (AI) is increasingly permeating healthcare, from physician assistants to consumer applications. Since AI algorithm's opacity challenges human interaction, explainable AI (XAI) addresses this by providing AI decision-making insight, but evidence suggests XAI can paradoxically induce over-reliance or bias. We present results from two large-scale experiments (623 lay people; 153 primary care physicians, PCPs) combining a fairness-based diagnosis AI model and different XAI explanations to examine how XAI assistance, particularly multimodal large language models (LLMs), influences diagnostic performance. AI assistance balanced across skin tones improved accuracy and reduced diagnostic disparities. However, LLM explanations yielded divergent effects: lay users showed higher automation bias - accuracy boosted when AI was correct, reduced when AI erred - while experienced PCPs remained resilient, benefiting irrespective of AI accuracy. Presenting AI suggestions first also led to worse outcomes when the AI was incorrect for both groups. These findings highlight XAI's varying impact based on expertise and timing, underscoring LLMs as a "double-edged sword" in medical AI and informing future human-AI collaborative system design.

Negar Kamali, Karyn Nakamura, Angelos Chatzimparmpas et al.

The high level of photorealism in state-of-the-art diffusion models like Midjourney, Stable Diffusion, and Firefly makes it difficult for untrained humans to distinguish between real photographs and AI-generated images. To address this problem, we designed a guide to help readers develop a more critical eye toward identifying artifacts, inconsistencies, and implausibilities that often appear in AI-generated images. The guide is organized into five categories of artifacts and implausibilities: anatomical, stylistic, functional, violations of physics, and sociocultural. For this guide, we generated 138 images with diffusion models, curated 9 images from social media, and curated 42 real photographs. These images showcase the kinds of cues that prompt suspicion towards the possibility an image is AI-generated and why it is often difficult to draw conclusions about an image's provenance without any context beyond the pixels in an image. Human-perceptible artifacts are not always present in AI-generated images, but this guide reveals artifacts and implausibilities that often emerge. By drawing attention to these kinds of artifacts and implausibilities, we aim to better equip people to distinguish AI-generated images from real photographs in the future.

Matthew Groh, Aruna Sankaranarayanan, Nikhil Singh et al.

Recent advances in technology for hyper-realistic visual and audio effects provoke the concern that deepfake videos of political speeches will soon be indistinguishable from authentic video recordings. The conventional wisdom in communication theory predicts people will fall for fake news more often when the same version of a story is presented as a video versus text. We conduct 5 pre-registered randomized experiments with 2,215 participants to evaluate how accurately humans distinguish real political speeches from fabrications across base rates of misinformation, audio sources, question framings, and media modalities. We find base rates of misinformation minimally influence discernment and deepfakes with audio produced by the state-of-the-art text-to-speech algorithms are harder to discern than the same deepfakes with voice actor audio. Moreover across all experiments, we find audio and visual information enables more accurate discernment than text alone: human discernment relies more on how something is said, the audio-visual cues, than what is said, the speech content.

Ziv Epstein, Matthew Groh, Abhimanyu Dubey et al.

How does the visual design of digital platforms impact user behavior and the resulting environment? A body of work suggests that introducing social signals to content can increase both the inequality and unpredictability of its success, but has only been shown in the context of music listening. To further examine the effect of social influence on media popularity, we extend this research to the context of algorithmically-generated images by re-adapting Salganik et al's Music Lab experiment. On a digital platform where participants discover and curate AI-generated hybrid animals, we randomly assign both the knowledge of other participants' behavior and the visual presentation of the information. We successfully replicate the Music Lab's findings in the context of images, whereby social influence leads to an unpredictable winner-take-all market. However, we also find that social influence can lead to the emergence of local cultural trends that diverge from the status quo and are ultimately more diverse. We discuss the implications of these results for platform designers and animal conservation efforts.

Matthew Groh, Ziv Epstein, Chaz Firestone et al.

The recent emergence of machine-manipulated media raises an important societal question: how can we know if a video that we watch is real or fake? In two online studies with 15,016 participants, we present authentic videos and deepfakes and ask participants to identify which is which. We compare the performance of ordinary human observers against the leading computer vision deepfake detection model and find them similarly accurate while making different kinds of mistakes. Together, participants with access to the model's prediction are more accurate than either alone, but inaccurate model predictions often decrease participants' accuracy. To probe the relative strengths and weaknesses of humans and machines as detectors of deepfakes, we examine human and machine performance across video-level features, and we evaluate the impact of pre-registered randomized interventions on deepfake detection. We find that manipulations designed to disrupt visual processing of faces hinder human participants' performance while mostly not affecting the model's performance, suggesting a role for specialized cognitive capacities in explaining human deepfake detection performance.

Ziv Epstein, Océane Boulais, Skylar Gordon et al.

The latent space modeled by generative adversarial networks (GANs) represents a large possibility space. By interpolating categories generated by GANs, it is possible to create novel hybrid images. We present "Meet the Ganimals," a casual creator built on interpolations of BigGAN that can generate novel, hybrid animals called ganimals by efficiently searching this possibility space. Like traditional casual creators, the system supports a simple creative flow that encourages rapid exploration of the possibility space. Users can discover new ganimals, create their own, and share their reactions to aesthetic, emotional, and morphological characteristics of the ganimals. As users provide input to the system, the system adapts and changes the distribution of categories upon which ganimals are generated. As one of the first GAN-based casual creators, Meet the Ganimals is an example how casual creators can leverage human curation and citizen science to discover novel artifacts within a large possibility space.

Matthew Groh, Ziv Epstein, Nick Obradovich et al.

Recent advances in neural networks for content generation enable artificial intelligence (AI) models to generate high-quality media manipulations. Here we report on a randomized experiment designed to study the effect of exposure to media manipulations on over 15,000 individuals' ability to discern machine-manipulated media. We engineer a neural network to plausibly and automatically remove objects from images, and we deploy this neural network online with a randomized experiment where participants can guess which image out of a pair of images has been manipulated. The system provides participants feedback on the accuracy of each guess. In the experiment, we randomize the order in which images are presented, allowing causal identification of the learning curve surrounding participants' ability to detect fake content. We find sizable and robust evidence that individuals learn to detect fake content through exposure to manipulated media when provided iterative feedback on their detection attempts. Over a succession of only ten images, participants increase their rating accuracy by over ten percentage points. Our study provides initial evidence that human ability to detect fake, machine-generated content may increase alongside the prevalence of such media online.

Shayne O'Brien, Matt Groh, Abhimanyu Dubey

The true distribution parameterizations of commonly used image datasets are inaccessible. Rather than designing metrics for feature spaces with unknown characteristics, we propose to measure GAN performance by evaluating on explicitly parameterized, synthetic data distributions. As a case study, we examine the performance of 16 GAN variants on six multivariate distributions of varying dimensionalities and training set sizes. In this learning environment, we observe that: GANs exhibit similar performance trends across dimensionalities; learning depends on the underlying distribution and its complexity; the number of training samples can have a large impact on performance; evaluation and relative comparisons are metric-dependent; diverse sets of hyperparameters can produce a "best" result; and some GANs are more robust to hyperparameter changes than others. These observations both corroborate findings of previous GAN evaluation studies and make novel contributions regarding the relationship between size, complexity, and GAN performance.

Ziv Epstein, Blakeley H. Payne, Judy Hanwen Shen et al.

AI researchers employ not only the scientific method, but also methodology from mathematics and engineering. However, the use of the scientific method - specifically hypothesis testing - in AI is typically conducted in service of engineering objectives. Growing interest in topics such as fairness and algorithmic bias show that engineering-focused questions only comprise a subset of the important questions about AI systems. This results in the AI Knowledge Gap: the number of unique AI systems grows faster than the number of studies that characterize these systems' behavior. To close this gap, we argue that the study of AI could benefit from the greater inclusion of researchers who are well positioned to formulate and test hypotheses about the behavior of AI systems. We examine the barriers preventing social and behavioral scientists from conducting such studies. Our diagnosis suggests that accelerating the scientific study of AI systems requires new incentives for academia and industry, mediated by new tools and institutions. To address these needs, we propose a two-sided marketplace called TuringBox. On one side, AI contributors upload existing and novel algorithms to be studied scientifically by others. On the other side, AI examiners develop and post machine intelligence tasks designed to evaluate and characterize algorithmic behavior. We discuss this market's potential to democratize the scientific study of AI behavior, and thus narrow the AI Knowledge Gap.