IntroductionIn an era increasingly dominated by artificial intelligence, the matter of informed consent has never been more crucial. A study conducted by Ipsos indicates a significant 11-point drop in internet trust since 2019[1]. Particularly in the European Union, there is growing concern about the handling of personal data. This article aims to shed light on the ways in which AI platforms, like OpenAI’s ChatGPT, fall short of meeting key guidelines established by the European Union’s General Data Protection Regulation (GDPR).Utilizing data from Enforcement Tracker , we present a graph illustrating the distribution of GDPR fines across various sectors. The Media, Telecoms, and Broadcasting sectors are particularly noteworthy, both for the number of violations and the scale of the fines imposed, signaling serious continuos non-compliance .

Abstract: One of the main challenges of current machine learning models is achieving greater control in synthetic content generation tasks. In this paper, we focus on a methodology for working with Stable Diffusion that allows us to observe the conceptualization that the model has regarding historical temporal representation given a seed. We show how prompt engineering can be used to obtain images of historical recreations in order to evaluate the representation that these models can make of historical evolution. Keywords: prompt engineering, machine learning, artificial intelligence, historical recreations, stable diffusion. INTRODUCCIÓN In the field of synthetic artistic image production, we encounter generative models that solve image synthesis tasks by learning implicit statistical distribution either through Generative Adversarial Networks GANs  \cite{Goodfellow_2020} or diffusion models \cite{NEURIPS2020_4c5bcfec} \cite{Geroski_2000}.  The latter have been gaining ground over adversarial networks because they solve the problem of training convergence and allow for good results with guidance. STABLE DIFFUSION One of the main open-source diffusion generative models is Stable Diffusion. It is an open-source Text2IM model that converts text to an image. Created with the support of companies such as CompVis, Stability AI, and Runway ML on LAION datasets, the model is available for fine-tuning and can be implemented on an advanced GUI (AUTOMATIC1111) with low computational requirements (4GB VRAM) in local mode or in the cloud through collaborative tools such as Google Colab. It has a large and active community of collaborators who are increasingly enthusiastic  \cite{art} OpenArt AI (discord.com) .In general, from a UI user perspective, there are two fundamental aspects for generating synthetic content: a text input (prompt) and the configuration of certain parameters to enable the inference of probabilistic prediction and image generation.   We have conducted this work to the best of our knowledge, given the limited scientific literature available on these recently developed products.  DREAM STUDIO on-line generator One way to make the use of generative models more accessible is through online generators. Dream Studio is an online generator for Stable Diffusion, a web application that facilitates the generation process for the different versions of the model through an interface from which we can configure the output image size, number of images, Cfg scale (or how much the output is adjusted to the text), steps (properly diffusion of noise that increases the level of detail of the image also increasing the probability of artifacts), sampler selection, use of CLIP or choose to generate a random image or a single seed. The latter parameter is a controller of the generated noise so that with fixed parameters, varying only the seed, we obtain different images, while if we keep it fixed, we will get the same image.1.2.1 The importance of the prompt message. The other key aspect to consider is the so-called Prompt Engineering (PE) . \cite{arta}  \cite{prompting}. This term refers to the art of creating tokens with keywords that have some persistence in the model, allowing us to approximate the output to the specific details we require. However, as we will see below, several problems can arise between the configuration and the prompt.  \cite{artb} 2. Methodology   Stable Diffusion guides suggest a logical structure in prompting that starts by describing the type of image desired, followed by the subject matter, details to add, and finally, the style and other formal elements, along with the so-called Magic words - words like 64k UHD HDR... that have persistence in the model and can increase the level of quality .  \cite{openart} We set our output to 704x512 with a Cfg Scale of 16 (as opposed to the default 7) and 15 steps (as opposed to the default 50), implemented for version 2.1-768, and used CLIP for seed generation. We force the Cfg Scale to ensure that the image matches the prompt. A) Creating an old view of a street in Rome. A standard prompt: "street Rome temple ancient". Due to the values of the Cfg Scale configuration, strongly contrasted images with saturated colors and artifacts are generated.