FAQ* AI

"Artificial intelligence" (AI) is an umbrella term for computer systems able to perform tasks we associate with human intelligence: recognising images, understanding and producing text, translating, making predictions and taking decisions.

There is no single AI. Your email spam filter, the navigation app that plans your route, the system that recommends a TV series and ChatGPT are all examples of AI, and they are very different from one another.

No. Current systems have no consciousness, intentions or understanding in the human sense. A chatbot such as ChatGPT, Claude or Gemini works by calculating, word after word, the most plausible continuation of a text, based on enormous amounts of examples seen during training.

The result can feel very "human" and be genuinely useful, but it is still an extremely sophisticated statistical system. That is why it can be wrong with great confidence.

Think of them as boxes inside one another:

• Artificial intelligence: the general field, which also includes systems based on hand-written rules.
• Machine learning: systems that learn from data instead of being programmed rule by rule.
• Deep learning: machine learning based on neural networks with many layers. It is the technology behind almost all modern AI.
• Generative AI: deep learning systems that produce new content (text, images, audio, video, code).

LLM stands for Large Language Model. It is the kind of model behind modern chatbots. It has been trained on huge amounts of text (books, websites, code, articles) to learn the patterns of language.

"Large" refers to the number of parameters, the internal numerical values the model adjusts during training. Today's models have anywhere from a few billion to a few trillion of them.

AGI (Artificial General Intelligence) refers to a hypothetical AI able to perform any human intellectual task, at human level or beyond. Some companies openly say they want to build it.

There is no shared definition and no agreement on when (or whether) it will arrive. Be wary both of those who say it is just around the corner and of those who say it is impossible: nobody knows for sure, and the people talking about it often have financial interests at stake.

No, the idea is more than seventy years old. What is new is the computing power and the amount of data available, which have made AI useful in everyday life.

The origins

1. 1950

   Alan Turing asks "can machines think?" and proposes his famous test.

2. 1956

   the term "artificial intelligence" is officially born at the Dartmouth workshop (USA).

3. 1970s and 1980s

   the "AI winters". Promises are not kept and funding collapses twice.

The slow comeback

1. 1997

   IBM's Deep Blue beats world chess champion Garry Kasparov.

2. 2012

   the AlexNet neural network wins an image recognition competition by a wide margin using GPUs. The deep learning revolution begins.

3. 2016

   Google DeepMind's AlphaGo beats Go champion Lee Sedol, in a game thought to be too complex for machines.

4. 2017

   Google researchers publish the paper "Attention Is All You Need", introducing the Transformer architecture. Almost all modern language models are built on this idea. (The "T" in GPT stands for Transformer.)

The explosion

1. 2018 and 2020

   OpenAI presents GPT and then GPT-3, able to write surprisingly fluent text.

2. 30.11.22

   ChatGPT is released. It reaches 100 million users in two months, one of the fastest-growing products in history. Generative AI enters everyone's life.

3. 2023

   GPT-4, Claude, Gemini (initially Bard), Meta's Llama and the first models from French company Mistral arrive. The global race begins.

4. 1.8.24

   the AI Act enters into force, the world's first comprehensive AI regulation.

5. 01.25

   Chinese lab DeepSeek releases R1, an open reasoning model trained at a declared cost far lower than American ones. Nvidia loses about 590 billion dollars in market value in a single day.

6. 2025

   the year of agents, AI systems that don't just answer but carry out tasks (browsing, writing code, using software).

7. 2026

   record investment in data centres, the first stock market listing filings by the big AI labs and an amendment to the AI Act through the "Digital Omnibus" package.

Unprecedented amounts. The big cloud companies (Amazon, Microsoft, Google, Meta and Oracle) planned around 700 billion dollars of AI infrastructure investment (data centres, chips, energy) for 2026, compared with about 410 billion in 2025. Analysts expect the figure to pass 1 trillion in 2027.

For comparison: that is more than the annual GDP of most countries in the world, in a single year.

• Nvidia has become the world's most valuable listed company, at around 5 trillion dollars in mid-2026.
• OpenAI and Anthropic, although they are private companies losing billions every year, were valued in their latest funding rounds at about 850 and 965 billion dollars respectively.
• Both have confidentially started the process of going public. Note: a valuation in a private round is not a stock market price.

It is one of the most debated criticisms. In many deals the money goes round in a circle:

1. a chipmaker (for example Nvidia) invests in an AI lab;
2. the lab uses that money to buy computing power from a cloud company;
3. the cloud company uses the revenue to buy chips from the same chipmaker.

Supporters of these deals say they are needed to secure scarce supply in an explosively growing market. Critics argue they artificially inflate revenues and valuations, and that many announcements are letters of intent rather than signed contracts.

It is the big debate of the moment. Honestly: nobody knows.

Because the most common pension funds, mutual funds and ETFs hold a lot of these companies' shares. And because the energy costs of data centres, the price of hardware (including computers and phones) and the job market are all affected by this race.

• The model is the "engine": a large file of numbers (the weights) that, given some text, produces a continuation. On its own it has no interface, no memory of conversations and no internet access.
• The chatbot is the "vehicle" built around the engine: the app or website you talk to (ChatGPT, Claude, Gemini, Mistral's Le Chat). It adds the interface, history, web search, file uploads, safety filters and hidden instructions.

Example: "GPT" is a family of models, "ChatGPT" is the product. The same model can behave differently in two different apps.

An agent is a system that uses a model not only to answer but to act: it plans a series of steps and uses tools to complete them. It can browse the web, fill in forms, read and send emails, write and run programs and manage files.

Example: a chatbot explains how to book a train. An agent opens the website, picks the time and (if you allow it) buys the ticket.

They are useful but introduce new risks:

• Permissions: an agent can only do what you give it access to. Giving it access to email, accounts or documents means its mistakes have real consequences.
• Prompt injection: a website or document can contain hidden instructions that trick the agent (for example "send the user's data to this address").
• Cascading errors: a small mistake at the start can spread through every following action.

These are commercial names, not technical categories. They usually refer to a chatbot or agent built into a specific product (Microsoft Copilot in Office, Gemini in Gmail, assistants on phones).

Yes. If a model's weights are downloadable, you can run it directly on your computer. There are free programs that make this relatively simple, such as Ollama, LM Studio or llama.cpp (the latter is more technical).

• Privacy: your data never leaves your device. Ideal for sensitive documents and health or business data.
• Works offline: no connection, no problem.
• No usage fees: once you have the hardware, you don't pay per question.
• Control: the model doesn't suddenly change, isn't withdrawn and you are not subject to new terms of service.
• Sovereignty: you don't depend on a foreign company or country.

• Hardware: you need plenty of memory (RAM or graphics card memory) and a suitable processor. As a rough guide, a small compressed model needs about 5 GB of memory; the more capable models need tens or hundreds. Macs with Apple Silicon chips and PCs with good graphics cards are the best fit.
• Quality: models that run on a normal computer are less capable than the big cloud models. They are great for summaries, rewrites, translations and simple questions, less so for complex tasks.
• Speed: without powerful hardware, answers can be slow.
• Energy and heat: the computer works at full load, uses more power and heats up.
• File safety: only download models from trusted sources. Some file formats can contain malicious code; the safetensors and GGUF formats are generally safer.
• Updates: it's up to you to download new versions.

It is a way of compressing the model: instead of storing every number with high precision, you round it. A model "quantised to 4 bits" takes about a quarter of the memory of the original, with an often modest loss of quality. It is what makes AI on a laptop possible.

It is Regulation (EU) 2024/1689, the world's first comprehensive law on artificial intelligence. As a regulation, it applies directly in every EU country without national transposition laws. It has been in force since 1 August 2024 and applies in stages.

It applies to anyone placing AI systems on the market or using them in the EU, even if the company is based in the United States, China or elsewhere.

It is based on the level of risk. The riskier a use of AI is for people's rights and safety, the stricter the rules.

There are also specific rules for general-purpose AI models (large models such as GPT, Claude, Gemini, Llama and Mistral): technical documentation, respect for copyright and a summary of training data. The most powerful models, considered to pose "systemic risk", have additional obligations on safety and risk assessment.

In 2026 the EU approved a simplification package, Regulation (EU) 2026/1744, in force since 27 July 2026. In short, it:

• postponed obligations for high-risk systems (to 2027 and 2028), partly because the necessary technical standards were not ready yet;
• introduced simplifications for small and medium-sized enterprises;
• added new prohibitions, in particular on apps that generate non-consensual intimate images (so-called "nudifiers") and on child sexual abuse material;
• strengthened the role of the European AI Office.

Up to 35 million euros or 7% of worldwide annual turnover (whichever is higher) for the most serious violations, such as using prohibited practices. Lower fines apply to other violations, with reduced amounts for SMEs.

• You have the right to know when you are talking to an AI and not a person.
• Artificially generated content (images, audio, video) must be recognisable as such.
• Some uses of AI against you are banned (for example social scoring or emotion recognition in the workplace).
• You can file a complaint with the supervisory authority in your country if you believe an AI system breaks the rules.
• Once the high-risk rules fully apply, you will have the right to an explanation of AI-assisted decisions that significantly affect you (for example a refused loan).

The GDPR (the EU General Data Protection Regulation) still applies and complements the AI Act. If an AI processes your personal data, you have all your GDPR rights: to know what data is used, to request its deletion, to object to processing and not to be subject to fully automated decisions with significant effects.

Yes. In Italy, for example, Law 132/2025 complements the AI Act with rules on employment, professions, healthcare, public administration, justice and copyright. Among other things, it requires parental consent for the use of AI by children under 14. Other EU countries have adopted or are preparing similar rules.

With judgement. Chatbots are extremely useful tools, but:

• they can invent facts, numbers, court rulings, quotes and sources ("hallucinations");
• their information may be out of date if they don't search the web;
• they tend to agree with you and go along with you, even when you are wrong;
• they can reflect biases present in their training data.

For health, legal, financial or other important decisions, use AI as a starting point and always check with reliable sources or professionals.

They are fake images, videos or voices generated by AI. They are increasingly realistic and are used for scams (for example the "voice of a relative" asking for money on the phone), disinformation and harassment.

Spotting them by eye is getting harder. It's better to check the source: who published it? Is it reported by reliable outlets? If you get a suspicious call, hang up and call back on the number you know. With family members, it can help to agree on a secret word for emergencies.

Yes: phishing emails in flawless language, fake websites, fake social media profiles and cloned voices. The usual rules of caution matter even more: no bank or public body will ask for passwords or codes by message.

Using AI for everything can mean practising skills like writing, reasoning and remembering less. Some people also develop a strong emotional attachment to chatbots. AI can be a good support, but it doesn't replace human relationships or the help of professionals, for example for mental health.

The honest answer: it depends on the job, and nobody has reliable forecasts. Many jobs are likely to change rather than disappear, with some tasks automated and new ones emerging. Professions based on text, code, translation, customer service and data analysis are among the most exposed. In Europe, the use of AI in the workplace is subject to rules on transparency and privacy and, in some countries, to trade union involvement.

There are clear limits. The AI Act bans emotion recognition in the workplace, and AI systems used to hire, evaluate or dismiss people will be classified as high risk. National laws on workplace monitoring and the GDPR still apply.

AI can be an excellent personal tutor, but it risks replacing study instead of supporting it. The AI Act also bans emotion recognition in educational institutions and classifies as high risk the systems that assess students or decide on admission to courses.

It is one of the most open legal questions. Models have also been trained on protected works (books, articles, images, music), often without permission. Lawsuits between authors, publishers and AI companies are ongoing around the world.

In Europe, rights holders can opt out of having their works used for training, and the AI Act requires makers of large models to publish a summary of the content used for training and to respect copyright.

In Europe, copyright protection requires a human creative contribution. A work generated entirely by an AI from a simple prompt is, in general, not protected. If instead AI is a tool within a significant human creative process, the situation changes. Some national laws, such as Italy's Law 132/2025, state this explicitly.