Generative AI: how it works and applications

Generative artificial intelligence (generative AI) is having an increasingly disruptive impact, both in the enterprise context and in the mainstream imagination, thanks to its ability to create high-quality multimedia content extremely rapidly, even for ordinary users without in-depth technical knowledge. When speaking of generative AI, we are talking about much more than a simple technological fact, as its applications promise to assist us constantly in both work and social life, offering possibilities that until recently were hardly imaginable.

As we will see, generative AI is not an absolute novelty in the field of artificial intelligence, since the first applications date back to chatbots from the 1960s. However, it is only in the past decade, with the introduction of GANs (generative adversarial networks), that a new era has truly begun, with the ability to create incredibly realistic images, video and audio.
Beyond purely creative capabilities, today generative AI can contribute with AI assistants able to generate enormous added value from company data. In any case, many of these applications are still in the early adoption phase and require significant expertise for implementation within processes and integration with existing systems.

Il potenziale della generative AI è enorme, ma occorre soprattutto saper riconoscere le tecnologie e le applicazioni in grado di portare davvero efficienza e innovazione nei processi, filtrando soluzioni esclusivamente viziate dall’hype mediatico, ma ancora troppo acerbe per produrre risultati apprezzabili.
Grazie alla comprovata esperienza di Moxoff, vediamo cos’è la generative AI, come funziona e quali sono le sue principali applicazioni oggi in azienda.

What is Generative AI

According to the definition provided by Gartner, generative AI is a: “Disruptive technology capable of generating artefacts that previously relied on human creativity, delivering innovative results free from the biases that are typical of human experience and thought processes.” Gartner also emphasises that: “IT leaders worldwide must use appropriate governance to harness its extraordinary creative potential.”
Generative AI is used both to quickly create new text, audio, and video content from a simple prompt, and to automate and make more efficient low-value-added tasks, so that humans can focus more specifically on activities that are more strategic for the business. This is particularly valuable when dealing with large volumes of data in repetitive procedures.

How Generative AI originated

The first episode that can be associated with generative artificial intelligence is most likely the early chatbot ELIZA, developed back in 1961 by Joseph Weizenbaum. This was a piece of software equipped with functionalities that allowed it to carry out simple interactions with humans based on natural language.

In the early Seventies, Seppo Linnainmaa introduced the concept of backpropagation, a process that would become fundamental in the training of deep neural networks (deep learning) on which many generative artificial intelligence applications are still based today.
In 1979, Kunihiko Fukushima published his studies on the Neocognitron, still recognised as the first deep neural network, used for the visual recognition of handwritten text, with the ability to adjust the weights between the various neural connections. In 1986, David Rumelhart revisited the concept of backpropagation to define a new training system for neural networks, ushering in a successful new era for deep learning techniques. In 1989, the renowned Yann LeCun used a neural network system based on a backpropagation algorithm to develop an application capable of recognising handwritten postcodes.
The history of artificial intelligence is a long alternation between summers and winters. Its potential has been known for decades, but it was only in the 2000s that we witnessed the first real mainstream breakthroughs. This happened thanks to advances in computing technologies and the advent of cloud computing, which made new development ecosystems available to a wide audience and, above all, the extraordinary computing power that AI systems require for training.

Generative AI as we understand it today is a relatively recent discipline, dating back only to 2014 with the advent of GANs (generative adversarial networks). A major boost to deep learning and generative AI in general came from the development of increasingly powerful GPUs (graphics processing units), which enabled the parallel computation required during algorithm training. Originally designed for PC gaming, GPUs became the go-to hardware for AI, joined more recently by NPUs (neural processing units), which ushered in the era of AI PCs by offloading CPUs and GPUs from the frequent, short calculations needed by AI-based applications.
In addition to GANs, transformers and other techniques capable of generative functionality for an ever-wider range of practical domains have been introduced. In November 2022, OpenAI introduced ChatGPT, the application that brought artificial intelligence to the centre of the world stage, thanks to its remarkable capabilities in conducting research and creating both textual and visual content rapidly and accessibly for all. The rest is history in the making.

How Generative AI works

Generative artificial intelligence applications largely fall under the so-called multimodal AI, as they use prompts that may contain text, images, and video, and respond with content that is also multimedia-based.
Generative AI has now become an integral part of many consumer applications that we use every day, often without realising it, for example on our smartphones. In the enterprise sphere, as we will see, generative AI can optimise and make business processes more efficient across a wide range of applications. Let’s look at some of the key elements that characterise how it works.

Algorithms and models

Generative Artificial Intelligence models combine various algorithms to represent and process available data. For content creation, for example, they use several natural language processing (natural language processing) techniques, which enable them to generate text that is convincing in both topic and narrative tone, often indistinguishable from that produced by humans.
From 2014 to today, various generative AI models have been developed for different application domains. The most significant are GANs, VAEs, and transformers.

GAN (generative adversarial network)

GANs use two neural networks in an adversarial manner: a discriminator, developed to distinguish synthetic data from real data, and a generator, which creates synthetic data attempting to emulate reality, until the discriminator is no longer able to detect any difference.

VAE (autoencoder variazionali)

Autoencoders are models that learn to encode data into a compressed format and then decode it to reconstruct the original input. Thanks to this feature, VAEs are widely used in applications such as image denoising, or generating images based on certain key characteristics of the images used during model training.

Transformer

Transformers are mainly used to create foundation models and large language models (LLMs). Among the first highly impactful examples are OpenAI’s GPT (Generative Pre-trained Transformer), which, at the time of writing, has evolved to version GPT-4o—the model behind ChatGPT—Meta Llama, Mistral 7B, and many others, including a wide range of open-source LLMs. These models use attention mechanisms to weigh the importance of content, enabling, for example, the generation of coherent and contextually appropriate text sequences.

Training process

The training process of a generative AI model generally falls within the context of deep learning, and involves a series of key stages:

1. 1. 1. 1. 1. Data collection
            2. Pre-processing
            3. Architecture selection
            4. Pre-training
            5. Tuning
            6. Optimisation
            7. Deployment

Optimisation and continuous improvement

Generative AI uses machine learning approaches to improve its knowledge and process results more efficiently over time, across various application contexts.
Among the most common optimisation techniques are fine-tuning, which involves refining the model by continuously providing labelled data relating to correct questions and answers, and reinforcement learning, where humans intervene by evaluating the model's results and inputting new values to achieve greater accuracy or relevance.
There are also techniques such as RAG (retrieval-augmented generation), which make it possible to include contextual data directly in the prompt to obtain coherent responses without necessarily having to retrain the model.

The advantages of Generative AI

A correct and informed implementation of generative AI in business processes enables a wide range of benefits, including a drastic increase in operational efficiency and the ability to personalise an incredible variety of content on a large scale.

Operational efficiency

Generative AI makes an essential contribution to the creation of new content in support of creative processes, especially by accelerating the conceptual phases of projects. Particularly useful is its ability to generate many draft versions from a single prompt.
Generative artificial intelligence is also highly effective for decision support. Its capability for predictive analysis on large datasets allows it to generate insights and recommendation hypotheses, enabling more informed and aware decisions based on objective contextual data.
AI applications are also reliable and operate 24/7, ensuring continuous availability—especially in applications such as chatbots for customer care.

Large-scale personalisation

Generative AI makes it possible to personalise the user experience in various applications, such as recommendation systems, which analyse user preferences and interactions to suggest purchase options in real time, creating optimal engagement in the buying process. Similar approaches are used in digital marketing to offer highly personalised experiences to customers, generating tailored product configurations based on their preferences, or sending specific messages to each target audience.
Beyond marketing and sales, generative AI also provides crucial support in customer care, thanks to chatbots capable of handling customer service requests more naturally and effectively than traditional systems, delivering personalised responses and real-time assistance without the need for constant human intervention.

Multiple use cases

Generative AI is successfully used to create a wide variety of multimedia content.

1. 1. 1. 1. 1. Text: generative models based on transformers are capable of generating instructions, technical documents, brochures, emails, and creating content for websites, blogs, and other digital marketing tools. Generative AI can automate repetitive, low-value writing tasks such as summarisation and meta descriptions for web pages.
            2. Images and video: applications such as Midjourney, DALL-E and Stable Diffusion have demonstrated the incredible creative potential of generative AI for visual production, with the ability to generate realistic, surreal, or conceptual images in a specific style, thus enabling rapid concept creation from a simple text prompt.
            3. Voice, sound and music: generative AI excels in synthesising vocal content and sounds, enabling the creation of incredibly realistic applications for chatbots and virtual assistants, as well as narrating audiobooks and composing music based on the samples used for model training.
            4. Software code: generative AI is used by developers to write code, translate content between various programming languages, and summarise code functionality, with the aim of drastically reducing application time-to-market.
            5. Synthetic data: generative AI models can generate synthetic data based on training with real or synthetic datasets. In the pharmaceutical field, this ability allows drug designers to create molecular structures with properties specified via prompt.

The challenges of Generative AI

The disruptive potential of generative artificial intelligence inevitably involves a range of challenges to be addressed at ethical, technological, and regulatory levels.

Ethical issues (deepfakes, fake news, etc.)

GANs make it possible to generate deepfake content, which can be used by cybercriminals to fraudulently impersonate real people, involving them in illegal activities. If used with malicious intent, generative AI can have decidedly negative impacts on society. Therefore, it is essential to recognise and, above all, mitigate the harmful effects of this new generation of cyber threats.
It is also crucial to supervise the operation of AI models to prevent discriminatory, racist, or otherwise human rights-violating results. To avoid models producing distorted outputs, it is necessary to ensure diverse training data, as well as to establish policies and guidelines based on the concrete application of responsible AI principles.

Data quality and bias

When training data is of poor quality, not relevant to the context, or insufficient to describe that context, generative AI applications can produce results that seem plausible but are actually incorrect, misleading, or otherwise incoherent. This issue is known as hallucinations.
To mitigate the harmful effects of hallucinations, developers tend to include a series of “guardrails” in the model, guiding it to search only within data considered sufficiently relevant or reliable. Essential in this context are the optimisation and continuous improvement phases of the model, such as fine-tuning and reinforcement learning.

Regulations and compliance

Generative AI models present the typical black box issues of artificial intelligence, with all the resulting challenges regarding explainability of results. Therefore, it is necessary to implement Explainable AI processes capable of ensuring the required level of trust in predictive systems.
Another challenge concerns the intellectual property rights of the data used during training, as demonstrated by the recent copyright infringement lawsuit brought by the New York Times against OpenAI.

It is also necessary to pay attention to proprietary data shared during the use of public LLMs, as such data may be reused to make the model itself more efficient, to the advantage of competitors. This could represent a violation of privacy regulations, industry-specific regulations, or private agreements.
In addition to GDPR, which has been in force since 2018, organisations will soon also have to comply with the provisions of the AI ACT, the European Regulation for Artificial Intelligence, which mandates specific transparency procedures and measures for every risk level.

Applications of Generative AI

L’intelligenza artificiale generativa viene oggi impiegata con successo in vari contesti applicativi.

Creazione di contenuti

Generative AI supports content creation in various ways:

1. 1. 1. 1. 1. Automatic content generation: LLMs and multimodal models can automatically create content for professional use, optimising the resources needed for their production.
            2. High quality: The knowledge gained through training on vast amounts of data enables LLMs to generate highly relevant results on the chosen topic, associating information from far more sources than a single person could consult.
            3. Personalisation of communication: Cognitive technologies enable generative AI models to personalise the tone and style of communication, adapting to various application contexts—ranging from highly formal to more informal content.

Design and planning

Generative artificial intelligence forms the basis of generative design, a new design paradigm that enables results to be generated according to objectives, rather than merely by form as in traditional systems. Based on training data containing all relevant characteristics and constraints (such as materials, regulations, dimensions, etc.), AI-enabled software starts from a known shape and proposes a series of design hypotheses. This is the case with topology optimisation, which allows the production of an object with similar mechanical properties using less material compared to traditional systems.

Industrial sector

Generative AI is used for a wide range of applications in manufacturing, to reduce time-to-market, enable digital twin systems, predictive maintenance applications, decision support, and other functionalities capable of optimising and making production processes more efficient. Generative AI is also employed to optimise logistics and improve the efficiency of the supply chain.

Human resources management

Generative AI supports various processes within HR. Its algorithms can speed up and make the recruitment process more efficient by searching, analysing, and categorising large volumes of CVs from various sources, identifying candidates with the qualifications and skills required for a given job application.
Generative AI also provides valuable support in performance evaluation and in developing personalised training paths based on each employee’s objectives, thereby contributing to improved employee satisfaction.

Technologies complementary to Generative AI

Generative AI works in a synergistic and complementary way with other emerging technologies, such as machine learning, cloud computing, and computer vision.

Machine Learning

Machine learning and generative AI are very often complementary. For example, an AI assistant can automate the drafting of replies to incoming emails. However, it would have nothing to act upon if there were not, upstream, a tool for classifying documents based on machine learning.

Cloud Computing

Generative AI often requires modernisation of the IT infrastructure. This need facilitates the adoption of cloud technologies, both for AI-based applications and, more broadly, to meet all anticipated workloads. Companies are therefore developing a cloud native approach, changing how they develop and deploy software.
Cloud computing also allows the use of leading multimodal AI applications in the public cloud. According to a recent study by Maximize Market Research, the cloud native market is expected to generate a global turnover of approximately 2.3 trillion dollars by 2029.

Computer Vision

Among the artificial intelligence techniques used synergistically with generative AI, computer vision certainly stands out. It uses pattern recognition functions to identify, for example, the objects present in images. Computer vision underpins applications such as facial recognition, product quality control, medical diagnostics, and moderation of non-compliant content (e.g. copyright violations).

Moxoff and Generative AI

Moxoff uses generative artificial intelligence models to learn patterns and relationships within data, employing supervised learning to generate new outputs.
Beyond the technologies and technical expertise that companies need to implement generative AI in their processes, Moxoff is able to guarantee full compliance with all applicable regulations, mitigating the risks arising from potential copyright violations or other infractions resulting from the improper use of data acquired for training generative AI models.

Organisations must also prepare for the introduction of the AI ACT, which sets out new obligations based on the level of risk related to their activities.
Moxoff’s contribution is also crucial when companies choose to adopt private AI solutions, often implemented starting from an open source LLM, which requires a high level of customisation in order to generate a usable asset capable of increasing its efficiency and value over time.