Introduction to generative AI

Discussing artificial intelligence (AI) often reflects our human bias of interpreting intelligence through our own perception of it. While its boundaries remain imprecise, we commonly associate it with the ability to perform tasks with a certain level of proficiency. However, the term ‘artificial’ does not adequately define what AI is or what it can achieve, since its competencies are structured differently from our own.

 

Artificial intelligence refers to a field of computer science. The technologies it encompasses, often used in combination, enable computers to perform a range of specialized tasks. What distinguishes AI from conventional software is its autonomy: it is designed to learn and improve over time, capable of reasoning, perceiving external stimuli, and interacting with its environment. For these reasons, AI can undertake a variety of complex tasks that are typically associated with human capabilities and with our broader understanding of intelligence.

AI is not a recent development. The field encompasses a variety of technologies, some of which have existed for years or even decades (see accompanying timeline).

The term ‘generative’ refers to AI’s capacity to automatically create information based on large volumes of data on which it has been trained. AI does not simply copy and paste what it has analyzed; rather, it imitates, refines, and produces entirely new outputs, generated through a statistical recomposition of patterns and structures identified during training. These outputs, commonly referred to as content, may take the form of text, images, music, or computer code.

data.png

AI learns rules by analyzing thousands of texts, images, or pieces of music (datasets). It examines how they are structured and identifies recurring patterns, which it then uses to generate new content in response to a prompt.

Content generated by AI represents a statistical recomposition of everything it has learned. It is meaningful because it adheres to the grammatical, visual, or musical rules extracted from the training datasets. These rules are stored in the form of a Large Language Model (LLM, see below), enabling, among other things, the synthesis of information or the creation of original content.

Several companies are currently major players in the field of AI. Their tools, often freely accessible online, are among the most widely used today.

Microsoft Research
Founded in 1991 with a focus on computer computing, the subsidiary began investing in AI in the 2010s through various acquisitions and now holds 20% of AI-related patents. Innovations are regularly integrated into its products (e.g., Bing, HoloLens, Cortana, 365, Azure, etc.). In 2016, Microsoft Research AI was established, and LinkedIn was acquired to provide data for its models. The 2019 partnership with OpenAI, providing cloud resources, was crucial for the development of GPT models.

  • In 2023, Microsoft released Copilot, an AI assistant integrated across all its services.

IBM
A pioneer in the field, notably with DeepBlue, which defeated chess champion Kasparov in 1996. In 2011, the Watson program gained fame by winning the TV game show Jeopardy!, requiring it to understand questions, use a buzzer, and respond vocally. After expanding to Internet applications, its use was restricted due to observations that it often relied on Wikipedia and occasionally used the word “bullshit” in responses.

  • Watson was commercialized in 2012 for healthcare, finance, and research. Considered a failure by 2022, the program was sold to an investment fund.

DeepMind (Google)
Founded in 2010, DeepMind was acquired by Google in 2014 to become its AI subsidiary. Its models draw inspiration from neuroscience to develop learning algorithms, with a focus on systems capable of playing games. AlphaGo was the first to defeat a world champion in Go in 2017. Following the release of GPT-3, Google quickly developed its conversational AI prototype.

  • The tool Bard was initially launched with the LaMDA model, quickly replaced by PaLM2, and later succeeded by Gemini.

Meta AI
Established in 2013 as Facebook Artificial Intelligence Research (FAIR), the division released PyTorch in 2017, an open-source catalog of machine learning models used by companies such as Tesla and Uber. FAIR was renamed Meta AI after Facebook’s reorganization. In 2023, it released the Llama language model, initially accessible on request to researchers for non-commercial use, before being leaked a month later.

  • Llama 2 is an open-source family of language models. Meta does not currently offer publicly accessible AI tools.

OpenAI
Founded in 2015 as a non-profit, OpenAI became a for-profit company in 2019 and established a major partnership with Microsoft. It developed GPT-3, a language model, in 2020, and announced DALL-E for image generation in 2021. By 2022, GPT-3 was accessible online, marking a major milestone. With Microsoft’s investment, GPT-4 was released in 2023.

  • ChatGPT is built on GPT-4, a powerful model capable of processing text and images, as well as performing real-time web searches.

Copie_I-TENG_-_ENGLISH_-_Rapport_IA_Generatives_-_UNIGE_-_Fevrier_2024_-_vENG.png

An LLM is considered ‘large’ because it contains a vast number of parameters—on the order of billions—each representing a piece of information. It is a ‘model’ because it consists of a neural network trained on extensive text data to perform a range of non-specific tasks. It is ‘language-based’ because it replicates the syntax and semantics of human natural language by predicting the most probable continuation for a given input. This also enables it to possess a form of general ‘knowledge’ derived from the training texts.

the_sky.png

Generative AI tools are powered by vast amounts of data used to train the algorithms underlying their models. These data can originate from various sources:

  • They are collected from content available online.

  • User queries and their corresponding responses also feed the models and help improve them.

As users submit queries, these tools generate additional data from the initial dataset. Consequently, both the data and its quality are critical elements, serving as the raw material for the functioning and use of generative AI.

While data is highly valuable for these tools, it also carries potential risks.

  • Generated data is fallible, with implications for information verification and transparency.

  • Input data may be sensitive, raising concerns related to security, privacy, intellectual property, and more.

Generative AI gives the illusion of control. The tool retains mastery over countless parameters that we neither understand nor fully manage, making it difficult to institutionalize its use. A recent study (Dell’Acqua et al., 2023) describes a “jagged” or “irregular” boundary to refer to the unpredictable limits of generative AI. Non-linear and inconsistent, this boundary renders the tool ambivalent: it can either enhance or hinder task performance. For example, it can produce complex texts, such as poems, yet struggle to generate a list of words beginning with the same letter.

It would be a mistake to anthropomorphize generative AI as an assistant whose errors and quirks result from a lack of intelligence or atypical behavior.

Generative AI is an entirely new entity, and unlike individuals, we cannot extrapolate its general competence from a few tasks, as it does not rely on a correlated or transversal set of cognitive resources. We lack the perspective necessary to define its boundaries.

A particular challenge for the University of Geneva (UNIGE) relates to personal rights, notably regarding personal data, which is not addressed by University regulations or directives and cannot be fully governed due to its complexity. The decision to provide recommendations is based on the potential harm to both the institution and individuals of failing to guide the use of generative AI, whose tools—approved or not—remain accessible.

Suivant

Risks