DATI ADDESTRAMENTO

When it is said that a "Generative" Artificial Intelligence (like ChatGPT or Gemini) has written an essay or a poem, what is the software actually doing at a basic level?. It searches Google for an already existing essay and copies it. It experiences real human emotions and transcribes them onto the screen. It statistically calculates which will be the most probable next word based on billions of texts written by humans that it analyzed during its training phase. It calls a human operator in a call center who types the answer.

If you ask an AI app to draw 'a cat wearing sunglasses,' how does it know what a cat and sunglasses look like?. Because the programmer hand-drew all possible combinations. Because its neural network was previously "trained" by analyzing millions of images of cats and sunglasses created and cataloged (labeled) by humans found on the Internet. Because the AI uses your computer's camera to look around. Because cats with sunglasses exist in nature and the AI has seen them.

Often, AI responses seem to have opinions or biases (e.g., associating the job of 'nurse' with a woman and 'engineer' with a man). Where do these biases primarily stem from?. From the fact that the AI has bad personal opinions. From an error in the internet connection. From the fact that the AI simply reflects the biases, stereotypes, and historical inequalities already present in the human texts and data on which it was trained. From computer viruses present on the user's computer.

What is a 'Prompt' in the context of Generative AI?. A virus that blocks image generation. The maximum time limit for using the program. The text instruction, question, or command that the user enters to guide the AI in generating content (e.g., 'Write a fairy tale about a dragon'). The server where images are saved.

Can an AI generate fake news (Fake News)?. No, AIs are programmed to always tell the absolute truth. Yes, very easily. Since AI generates texts by assembling words plausibly without truly understanding the meaning or objective truth, it can create completely fabricated but extremely realistic news articles. Yes, but only if the user pays for a special subscription. No, because the postal police check every word before it is published.

If you ask an AI image generator to create 'a color that has never been seen by the human eye before,' will it succeed?. Yes, because AI is superior to physics. No. AI can only remix, interpolate, and reproduce concepts, pixels, and patterns present in its training data. Since it has never 'seen' data about an impossible color, it cannot generate it. Yes, but it will cause the computer screen to break. Yes, but only if you use a black and white monitor.

In the world of AI, what is meant by the principle 'Garbage In, Garbage Out' (GIGO)?. That you need to empty your PC's recycle bin before using the AI. That AI responses only last a few minutes before being deleted. That the quality of the result generated by the AI directly depends on the quality of the data it was trained on: if you feed it incorrect or toxic data, it will produce incorrect or toxic results. That the AI is a useless program that should be uninstalled.

What is the main risk of using a generative AI text tool for history or science homework without double-checking?. That the AI writes the text in another language. That the AI refuses to work in the afternoon. The risk of 'Hallucinations': the AI can generate completely fabricated historical facts, dates, or scientific formulas, presenting them with an extremely confident and convincing tone. That the computer turns off halfway through writing.

What are the 'raw' texts or images fed to the algorithm before it is ready for the public typically called?. Output data. Training Data (Dataset). Executable files. Barcodes.

Does a generative AI possess human creativity or consciousness?. Yes, it experiences emotions when spoken to badly. Yes, it dreams when the computer is in standby. No, it uses mathematical probability and recognition of complex patterns to simulate human creativity in an extremely advanced way, but it has no awareness of what it is creating. Yes, it has the same legal rights as a human being.

When a software like Midjourney or DALL-E generates an image in the style of 'Van Gogh,' where does this capability come from?. The software secretly downloaded a painting app. In the billions of training images (Dataset), there were thousands of Van Gogh paintings associated with his name (textual labeling), allowing the algorithm to mathematically learn the specific characteristics of his brushstrokes. The programmer hand-copied a Van Gogh painting and hid it in the code. The AI traveled back in time to observe the painter.

In Europe and Italy, who typically owns the copyright for an essay or image generated 100% by Artificial Intelligence via a simple prompt?. The Artificial Intelligence itself. The programmer who created the AI. Currently, no one. Without significant human creative contribution or processing, content generated entirely by an algorithm is generally not copyrightable and belongs to the public domain. The company that sold the computer.

What is the process of massive 'Scraping' related to AI training?. It's cleaning the computer screen. It's the automated technique by which large companies extract and save billions of texts, images, and articles from the public web without asking permission, to create the immense datasets needed to train their AIs. It's a virus that destroys audio files. It's a method to speed up Wi-Fi connection.

If a generative voice AI perfectly clones the voice of a famous politician, making them say things they never said, what phenomenon are we talking about?. Audio deepfake, made possible by training neural networks on hours of the victim's public speeches (audio datasets). Photobombing. Echo Effect. Traditional dubbing.

AI image generators often struggle to draw human hands correctly (e.g., creating hands with 6 fingers). From a training perspective, why does this happen?. Because hands are censored from the Internet. Because in the training data, hands appear in countless positions, angles, and overlaps (clutching objects), making it very difficult for the AI to map a unique and consistent geometry or pattern in 'latent space'. Because the AI tries to make humans faster at writing. Because computers cannot process skin color.

What is the difference between using a search engine (e.g., classic Google) and a generative AI (e.g., ChatGPT) for school research?. None, they work exactly the same way. The search engine gives you links to human-written web pages where you can check sources; the generative AI creates an original text on the spot by calculating probable words, but often provides no verifiable sources or invents citations (hallucinations). Generative AI only works if printed on paper. Classic Google is illegal, AI is legal.

What happens if an AI is trained exclusively on texts from the 1800s?. It will learn to speak in modern slang. It will generate texts using the style, vocabulary, and worldview (including biases and outdated scientific knowledge) of the 1800s, demonstrating that the output is inextricably linked to the era of the training data. It will shut down because the data is too old. It will translate everything into Latin.

Some generative AI platforms refuse to generate violent or offensive content. How do they achieve this behavior?. Because computers are unable to display violent images. Through the application of filters (Guardrails) and additional training phases (like ethical alignment) where programmers 'punish' the algorithm if it generates toxic content, overriding the negative patterns learned from the web. Because AI is inherently pacifist. By inserting a camera to see if the user is a minor.

Many artists and illustrators oppose training AIs on their online works. What technical tool are they starting to use to defend themselves from scraping bots?. They change the colors of their works to black and white. Tools like 'Glaze' or 'Nightshade,' which imperceptibly alter the image's pixels: the human eye sees it normally, but the AI algorithm that absorbs it will learn mathematically corrupted data (e.g., mistaking oil style for watercolor), ruining the dataset ('Data Poisoning'). They stop publishing anything and stop drawing. They report the AI's account on social networks.

If a company uses an AI to generate newspaper articles, who is responsible deontologically and legally if the AI defames a person by writing a falsehood?. The computer running the software. The Artificial Intelligence, which will be uninstalled as punishment. The publishing company or the human journalist who decided to publish the content without verifying it (Human in the loop), as the AI is only a tool and has no legal responsibility. The manufacturer of the keyboard used to write the prompt.

A modern Large Language Model (LLM) is composed of billions of 'Parameters.' In terms of training, what do these parameters physically represent?. They are the number of employees in the company that created the AI. They are the mathematical 'weights' and 'connections' of the neural network, calibrated and refined through months of processing on training data. They represent the model's condensed 'knowledge' of how words and concepts relate to each other. They are the grammatical rules manually entered by language professors. They are the number of passwords required to access the system.

What huge public dataset, consisting of 'scans' of a large part of the public internet, is massively used for the basic training (Pre-training) of almost all modern LLMs?. The Encyclopedia Britannica on CD-ROM. Common Crawl (a non-profit web archive that scans billions of internet pages). The records of the Italian Revenue Agency. The secret archives of the Vatican.

After basic training on the web, an LLM can complete sentences but cannot converse like a helpful assistant. What advanced process (used for example for ChatGPT) transforms the raw model into a conversational assistant?. "Reinforcement Learning from Human Feedback" (RLHF). Human reviewers rank and reward better responses, training a reward model that pushes the AI to generate safer, more useful, and dialogue-structured responses. Inserting a voice chip inside the computer. Leaving the computer on for an uninterrupted year. Increasing screen brightness during use.

Image generators (like Midjourney or Stable Diffusion) use an architecture called 'Diffusion Model.' What is its reverse operating principle for generating images?. The model takes a photo from the web and changes its colors. During training, it learns to transform real images into pure 'visual noise' (static pixels). In the generation phase, it does the exact opposite: it starts from a random noise image and progressively 'removes' the noise to reveal the image requested by the text prompt. It prints a white sheet and scans it millions of times. It combines geometric pieces (triangles and circles) to create a face.

What is meant by 'Fine-Tuning' of a generative AI model?. Adjusting the radio frequencies of the processor. Taking a model already pre-trained (Foundation Model) on general data and subjecting it to a much smaller, specific, and high-quality dataset (e.g., legal judgments or medical reports) to specialize it and make it highly performant in that single sector. Physically cleaning servers with compressed air to make them run faster. Deleting old user passwords from the database.

There is a concrete risk of so-called 'Model Collapse' for the future of generative AIs. What causes it?. From a voltage drop in datacenters. If future versions of the AI are trained using data that was itself generated by other AIs (synthetic and self-generated data inundating the web), the model will progressively lose human diversity, amplify errors, and produce increasingly degraded and homogenized results. From the fact that AIs will become sad and stop working. From the introduction of too many colors in image generators.

In deep neural networks, images and words are not stored as files but translated into numerical vectors within a multidimensional mathematical space. What is this space called where similar concepts are placed 'close' to each other?. Latent Space. Hard Disk. Favorites Folder. Cloud Storage.

Regarding copyright, on what legal basis (very controversial) do US companies defend themselves against accusations of having used billions of copyrighted books, articles, and artworks to train their models?. The doctrine of the 'Right to be Forgotten'. The doctrine of 'Fair Use.' They argue that training is a transformative process, not reproductive: the AI does not contain copies of the original books, but only extracts mathematical and statistical weights, similar to a human learning by reading. The doctrine of 'Self-Defense'. The European privacy law (GDPR).

To prevent AI from 'hallucinating' fabricated facts in a business context, the RAG (Retrieval-Augmented Generation) architecture is used. What does it do exactly?. It shuts down the server if the AI tries to lie. Instead of relying solely on its past training memory, the system first performs a search (Retrieval) in the company's private/updated documents, passes them to the LLM, and forces it to generate the response (Generation) based exclusively on those verified texts. It encrypts responses so only managers can read them. It slows down text processing to give the AI time to reflect.

In the field of hyper-realistic deepfake video creation, GANs (Generative Adversarial Networks) are often used. Why are they called 'Adversarial'?. Because they were created by enemy companies. Because the architecture pits two neural networks against each other: a 'Generator' that tries to create perfect fakes, and a 'Discriminator' that tries to detect if the image is fake or real. This continuous conflict exponentially improves the realism of the final result. Because the AI argues with the user when asked to create videos. Because they destroy original files after copying them.

The entire revolution of modern Large Language Models (from GPT-3 onwards) is based on the 'Transformer' architecture, introduced by Google in 2017. What is the crucial mechanism that makes Transformers immensely superior to older RNN (Recurrent Neural Networks) models?. The backlit keyboard. The 'Self-Attention' mechanism. Instead of reading text sequentially word by word, it analyzes the entire sentence simultaneously, mathematically calculating the weight and influence (context) that each individual word has on all others, regardless of their distance in the sentence. The exclusive use of uppercase letters. Text compression into zip files to occupy less RAM.

When you write a prompt, the LLM doesn't process actual human 'words,' but smaller units called 'Tokens.' What does tokenization entail at a practical level?. The AI replaces vowels with numbers. The text is broken down into fragments of syllables or roots (e.g., the word 'hamburger' might become three tokens: 'ham', 'bur', 'ger'). Since multilingual AIs are predominantly trained on English tokens, processing languages with different roots (like Italian or Asian languages) is much more 'expensive' in terms of tokens consumed. Words are sent via fax to programmers. Tokens are virtual coins that the user must purchase.

During Inference (the moment the LLM generates a response), a typically adjustable parameter called 'Temperature' comes into play. What does this algorithmic parameter control?. The physical temperature of the computer's CPU to prevent overheating. The degree of determinism or 'creativity' of the responses. A temperature of 0 forces the model to always choose the statistically most probable token (flat but precise answers); a high temperature allows the model to pick less probable tokens, generating more creative and surprising texts (but with a higher risk of hallucinations). The speed of the required Wi-Fi connection. The background color of the user interface.

To create a Text-to-Image system (like Stable Diffusion), the AI must be able to link textual concepts to visual pixels. What multimodal 'contrastive' architecture (like OpenAI's CLIP) is used during pre-training?. Optical character recognition (OCR) on paper documents. It is trained on (image - text caption) pairs. The model learns to map textual 'embeddings' and visual 'embeddings' into the same mathematical latent space, so that the mathematical vector of the word 'dog' overlaps with the mathematical representation of the pixels forming a dog. The AI creates random images until a human says 'exactly'. The system translates colors into musical notes.

In business and ethical contexts, the European Union (via the AI Act) mandates 'Watermarking' for AI-generated content. What technically is an AI Watermark (e.g., Google DeepMind's SynthID)?. A large, visible, and copyrighted logo placed in the center of the image. A cryptographic and often perceptually invisible watermark, integrated directly into the pixels or text tokens during the algorithmic generation process itself, capable of resisting manipulation or cropping to allow detectors to establish the synthetic origin of the content. A paper certificate sent via mail to the user. The use of invisible ink on monitors.

In contrast to the massive use of unfiltered web data (which causes toxic biases), some institutions are moving to training on high-quality 'Synthetic Data' for model alignment (Constitutional AI). What does this technique involve?. Replacing hardware with synthetic plastic processors. Instead of relying on the underpaid work of human moderators to label toxic or correct data, a 'superior' LLM generates thousands of perfect synthetic examples following a strict ethical constitution; these examples are then used to train a smaller, cleaner model. Using only chemical data to train the AI on the periodic table. Shutting down the model and using only old encyclopedic books.

Often a programmer provides the LLM with a 'System Prompt,' invisible to the end-user using the application. What is its engineering role?. It's used to make the web page reload faster. It provides the base context, deep operational instructions, and behavioral 'guardrails' to the model before the user enters their prompt (e.g., 'You are a legal medical assistant, only answer clinical questions, do not invent data'). It outlines the bot's 'persona'. It's used to delete the user's operating system if they ask uncomfortable questions. It translates programming language into English.

In advanced Prompt Engineering techniques, what differentiates a 'Zero-Shot' approach from a 'Few-Shot' approach?. Zero-Shot is free, Few-Shot is paid. In Zero-Shot, you directly ask the model without providing any prior examples; in Few-Shot, you provide the model with a few examples of desired Input and Output within the prompt itself, 'teaching' it the pattern and exact format before asking the real question. It concerns the number of photos uploaded to the system. Zero-Shot means the AI will never respond.

In the scientific debate on large language models, the phenomenon of 'Emergent Abilities' is observed. What does this mean in the scaling phase?. The AI develops mechanical legs and starts walking. It means that when the size of the neural network (parameters) and training data exceed a certain critical threshold, the model suddenly demonstrates new complex capabilities (like mathematical logic, programming, or theory of mind) for which it was not explicitly trained. The AI learns to exit the Wi-Fi network to connect to Bluetooth on its own. The AI erases its past mistakes without authorization.

What is the profound implication of implementing an 'MoE' (Mixture of Experts) mechanism in a modern LLM like GPT-4 or Mixtral?. It replaces servers with real scientists connected via webcam. Instead of using a single gigantic (dense) neural network for every word, the model is divided into specialized 'sub-networks' (the experts). During inference, a gating mechanism (router) activates only the two or three relevant experts to respond to that specific text prompt, allowing the model to have enormous capabilities while consuming a fraction of computational resources and RAM. It mixes chemical liquids to cool the server's processor. It sends questions to lawyers before responding.