AI and ethics最新文献

ChatGPT, a chatbot using the GPT-n series large language model, has surged in popularity by providing conversation, assistance, and entertainment. This has raised questions about its agency and resulting implications on trust and blame, particularly when concerning its portrayal on social media platforms like Twitter. Understanding trust and blame is crucial for gauging public perception, reliance on, and adoption of AI-driven tools like ChatGPT. To explore ChatGPT’s perceived status as an algorithmic social actor and uncover implications for trust and blame through agency and transitivity, we examined 88,058 tweets about ChatGPT, published in a ‘hype period’ between November 2022 and March 2023, using Corpus Linguistics and Critical Discourse Analysis, underpinned by Social Actor Representation. Notably, ChatGPT was presented in tweets as a social actor on 87% of occasions, using personalisation and agency metaphor to emphasise its role in content creation, information dissemination, and influence. However, a dynamic presentation, oscillating between a creative social actor and an information source, reflected users’ uncertainty regarding its capabilities and, thus, blame attribution occurred. On 13% of occasions, ChatGPT was presented passively through backgrounding and exclusion. Here, the emphasis on ChatGPT’s role in informing and influencing underscores interactors’ reliance on it for information, bearing implications for information dissemination and trust in AI-generated content. Therefore, this study contributes to understanding the perceived social agency of decision-making algorithms and their implications on trust and blame, valuable to AI developers and policymakers and relevant in comprehending and dealing with power dynamics in today’s age of AI.

We present an application of conformal prediction, a form of uncertainty quantification with guarantees, to the detection of railway signals. State-of-the-art architectures are tested and the most promising one undergoes the process of conformalization, where a correction is applied to the predicted bounding boxes (i.e., to their height and width) such that they comply with a predefined probability of success. We work with a novel exploratory dataset of images taken from the perspective of a train operator, as a first step to build and validate future trustworthy machine learning models for the detection of railway signals.

With the recent advancements in the usage of Artificial Intelligence (AI)-based systems in the healthcare and medical domain, it has become necessary to monitor whether these systems make predictions using the correct features or not. For this purpose, many different types of model interpretability and explainability methods are proposed in the literature. However, with the rising number of adversarial attacks against these AI-based systems, it also becomes necessary to make those systems more robust to adversarial attacks and validate the correctness of the generated model explainability. In this work, we first demonstrate how an adversarial attack can affect the model explainability even after robust training. Along with this, we present two different types of attack classifiers: one that can detect whether the given input is benign or adversarial and the other classifier that can identify the type of attack. We also identify the regions affected by the adversarial attack using model explainability. Finally, we demonstrate how the correctness of the generated explainability can be verified using model interpretability methods.

A growing awareness of bias in artificial intelligence (AI) systems has recently emerged, leading to an increased number of publications discussing ethics in AI. Nevertheless, the specific issue of gender bias remains under-discussed. How can design contribute to preventing the emergence of gender bias in AI-driven systems? To answer this question, we investigated the current state of AI ethical guidelines within the European Union. The results revealed that most guidelines do not acknowledge gender bias but address discrimination. This raised our concerns, as addressing multiple biases simultaneously might not effectively mitigate any of them due to their often-unconscious nature. Furthermore, our results revealed a lack of quantitative evidence supporting the effectiveness of bias prevention implementation methods and solutions. In conclusion, based on our analysis, we propose four recommendations for designing effective guidelines to tackle gender biases in AI. Moreover, we stress the central role of diversity in embedding the gender perspective from the beginning in any design activity.

Improving agent capabilities and increasing availability of computing platforms and internet connectivity allows more effective and diverse collaboration between human users and automated agents. To increase the viability and effectiveness of human–agent collaborative teams, there is a pressing need for research enabling such teams to maximally leverage relative strengths of human and automated reasoners. We study virtual and ad hoc teams, comprising a human and an agent, collaborating over a few episodes where each episode requires them to complete a set of tasks chosen from given task types. Team members are initially unaware of their partner’s capabilities, and the agent, acting as the task allocator, must adapt the allocation process to maximize team performance. The focus of the current paper is on analyzing how allocation decision explanations can affect both user performance and the human workers’ outlook, including factors, such as motivation and satisfaction. We investigate the effect of explanations provided by the agent allocator to the human on performance and key factors reported by the human teammate on surveys. Survey factors include the effect of explanations on motivation, explanatory power, and understandability, as well as satisfaction with and trust/confidence in the teammate. We evaluated a set of hypotheses on these factors related to positive, negative, and no-explanation scenarios through experiments conducted with MTurk workers.

Skills-based matching promises mobility of workers between different sectors and occupations in the labor market. In this case, job seekers can look for jobs they do not yet have experience in, but for which they do have relevant skills. Currently, there are multiple occupations with a skewed gender distribution. For skills-based matching, it is unclear if and how a shift in the gender distribution, which we call gender mobility, between occupations will be effected. It is expected that the skills-based matching approach will likely be data-driven, including computational language models and supervised learning methods. This work, first, shows the presence of gender segregation in language model-based skills representation of occupations. Second, we assess the use of these representations in a potential application based on simulated data, and show that the gender segregation is propagated by various data-driven skills-based matching models. These models are based on different language representations (bag of words, word2vec, and BERT), and distance metrics (static and machine learning-based). Accordingly, we show how skills-based matching approaches can be evaluated and compared on matching performance as well as on the risk of gender segregation. Making the gender segregation bias of models more explicit can help in generating healthy trust in the use of these models in practice.

As opaque decision systems are being increasingly adopted in almost any application field, issues about their lack of transparency and human readability are a concrete concern for end-users. Amongst existing proposals to associate human-interpretable knowledge with accurate predictions provided by opaque models, there are rule extraction techniques, capable of extracting symbolic knowledge out of opaque models. The quantitative assessment of the extracted knowledge’s quality is still an open issue. For this reason, we provide here a first approach to measure the knowledge quality, encompassing several indicators and providing a compact score reflecting readability, completeness and predictive performance associated with a symbolic knowledge representation. We also discuss the main criticalities behind our proposal, related to the readability assessment and evaluation, to push future research efforts towards a more robust score formulation.

When deployed, machine-learning (ML) adoption depends on its ability to actually deliver the expected service safely, and to meet user expectations in terms of quality and continuity of service. For instance, the users expect that the technology will not do something it is not supposed to do, e.g., performing actions without informing users. Thus, the use of Artificial Intelligence (AI) in safety-critical systems such as in avionics, mobility, defense, and healthcare requires proving their trustworthiness through out its overall lifecycle (from design to deployment). Based on surveys on quality measures, characteristics and sub-characteristics of AI systems, the Confiance.ai program (www.confiance.ai) aims to identify the relevant trustworthiness attributes and their associated key performance indicators (KPI) or their associated methods for assessing the induced level of trust.

With the advancement of DNNs into safety-critical applications, testing approaches for such models have gained more attention. A current direction is the search for and identification of systematic weaknesses that put safety assumptions based on average performance values at risk. Such weaknesses can take on the form of (semantically coherent) subsets or areas in the input space where a DNN performs systematically worse than its expected average. However, it is non-trivial to attribute the reason for such observed low performances to the specific semantic features that describe the subset. For instance, inhomogeneities within the data w.r.t. other (non-considered) attributes might distort results. However, taking into account all (available) attributes and their interaction is often computationally highly expensive. Inspired by counterfactual explanations, we propose an effective and computationally cheap algorithm to validate the semantic attribution of existing subsets, i.e., to check whether the identified attribute is likely to have caused the degraded performance. We demonstrate this approach on an example from the autonomous driving domain using highly annotated simulated data, where we show for a semantic segmentation model that (i) performance differences among the different pedestrian assets exist, but (ii) only in some cases is the asset type itself the reason for this reduction in the performance.

The importance of high data quality is increasing with the growing impact and distribution of ML systems and big data. Also, the planned AI Act from the European commission defines challenging legal requirements for data quality especially for the market introduction of safety relevant ML systems. In this paper, we introduce a novel approach that supports the data quality assurance process of multiple data quality aspects. This approach enables the verification of quantitative data quality requirements. The concept and benefits are introduced and explained on small example data sets. How the method is applied is demonstrated on the well-known MNIST data set based an handwritten digits.