Artificial intelligence is shaping decisions in policy, business, education, and media, and it is changing the conditions we all live and work under.
Like other advanced technologies, it has the capacity to both improve people’s lives and to cause harm.
At Sandbox Labs, we are committed to helping citizens understand AI technologies. It is the first step to foster critical and thoughtful engagement with these tools and to build support for models of AI development that are human-centered and serve the public interest.
This page is that first step: a plain-language guide to what AI is, how it works, and the risks worth knowing about.
There are many definitions of artificial intelligence.
In general, AI is understood as a category of technologies that let computers do tasks usually associated with human intelligence, such as recognizing patterns, making predictions and generating text, images, audio and video. It is an umbrella term for many tools with different capabilities.
Today, when people say “AI”, they often mean generative AI. The differences between the types of AI are worth knowing, because each one works differently and plays a different role in our lives.
Rules written by people
Learns from data
Follows rules written by people.
In practice
IBM Deep Blue
In 1997 Deep Blue beat world chess champion Garry Kasparov. Its programmers told it what each piece is worth, and it scored 200 million possible positions a second using those rules.
Learns patterns from data.
In practice
Gmail’s spam filter
Nobody wrote Gmail’s spam rules. It learned what spam looks like from millions of emails people flagged, and Google says it blocks spam with over 99.9% accuracy.
Creates new content from learned patterns.
In practice
Image generators like DALL·E and Stable Diffusion
You describe a picture in words. The tool starts from random noise and shapes it, step by step, into a new image that matches. This is called diffusion.
Generative AI trained on human language.
In practice
ChatGPT, Gemini and Claude
They write an answer one word at a time, each time choosing the most likely next word. Because every word is a prediction, an answer can sound confident and still be wrong.
Decades of research, breakthroughs and setbacks shaped the AI we use today.
The field of AI
The mathematician proposes the imitation game, later known as the Turing test: can a machine hold a conversation you can’t tell apart from a person’s?
The field of AI
A summer workshop at Dartmouth College brings researchers together around a new term: “artificial intelligence”.
Symbolic AI
Often called the first AI program, it follows logic rules written by Allen Newell, Herbert Simon and Cliff Shaw, and proves 38 theorems from a classic maths book.
Machine learning
IBM’s Arthur Samuel builds a checkers program that gets better by playing against itself, and calls the idea machine learning.
Symbolic AI
Joseph Weizenbaum’s ELIZA imitates a therapist by matching keywords to scripted replies. Many people who used it still felt understood.
The field of AI
A UK government report finds that AI hasn’t delivered what it promised. Funding is cut at most British universities, and the field slows for years.
Symbolic AI
Stanford’s MYCIN suggests treatments for blood infections using hundreds of if-then rules written with doctors.
Machine learning
Rumelhart, Hinton and Williams show how a neural network can learn by correcting its own errors. The method, backpropagation, still underpins deep learning.
Symbolic AI
The market for specialized AI computers collapses and expert systems prove costly to maintain. Funding dries up into the early 1990s.
Symbolic AI
IBM’s Deep Blue defeats Garry Kasparov by scoring 200 million positions a second with chess knowledge written in by its programmers.
Machine learning
A neural network called AlexNet recognizes objects in photos far better than any rival, and deep learning becomes the leading approach.
Generative AI
Generative adversarial networks (GANs) set two neural networks against each other: one creates images, the other tries to spot the fakes.
Machine learning
DeepMind’s AlphaGo beats champion Lee Sedol 4–1. Unlike Deep Blue, it learned its own strategy: first from expert games, then by playing itself.
LLMs
Google researchers publish “Attention Is All You Need”, introducing the transformer: the design behind today’s large language models.
LLMs
OpenAI trains a transformer on thousands of books to predict the next word. It is the first “generative pre-trained transformer”.
LLMs
OpenAI’s GPT-3 scales up to 175 billion parameters and learns to handle new language tasks from a few examples. Its fluency shows how far large-scale training can take a general-purpose model.
Generative AI
Image generators such as Stable Diffusion and DALL·E 2 open to the public. Anyone can type a description and get a new picture in seconds.
LLMs
OpenAI releases ChatGPT. It reaches an estimated 100 million users in about two months, and generative AI goes mainstream.
LLMs
GPT-4 arrives as Anthropic’s Claude, Meta’s Llama and Google’s Gemini bring more powerful models from competing labs. AI becomes a fast-moving global race, with different companies and research communities shaping its future.
Reasoning models
Reasoning models such as OpenAI’s o1 are trained to work through a problem step by step before they reply.
Reasoning models
China’s DeepSeek releases R1, an open reasoning model that rivals OpenAI’s o1 at a fraction of the cost. Nvidia loses nearly $600 billion in market value in a single day, and the AI race goes global.
Generative AI
Google’s Veo 3 turns a written description into a video clip with matching sound: dialogue, sound effects and music, generated together.
Agentic AI
Coding agents such as Anthropic’s Claude Code take on whole tasks: they read a project, write and test changes, and keep going until the job is done, while people review the result.
The field of AI
AI is now part of daily work for millions of people. Governments, companies and communities are still deciding how it should be built, used and governed.
Most modern generative AI systems follow the same basic process.
An imperfect but useful way to view language models is to understand them as prediction machines. When building a sentence, the system predicts what the next word, and the one after that, is most likely to be, based on patterns learned from an enormous amount of training text.
Phase 1:Building the model
01
AI developers feed vast amounts of text, images, audio or code into the model: a neural network that acts as a digital pattern-recognition engine.
02
The model analyzes the data and finds recurring relationships and structures in it.
03
What comes out of training is a web of mathematical patterns and probabilities, not a searchable database of facts. The model uses these patterns to turn raw probabilities into natural, human-like language.
The trained model is ready to use
Phase 2:Using the model
04
People give the model a prompt: a question, an instruction, an audio recording, a photo and so on.
05
The model uses the learned patterns to generate a likely output. The response is built one small piece (a token) at a time, with each piece shaping what comes next, until it is complete.
06
The result is newly generated content, also known as synthetic content: an answer, an image, a video, synthesized speech and so on, responding to the original prompt.
A closer look at step 5
1
Current input
2
Model predicts next token
3
Token chosen
4
Process repeats
Press Play or Next step to watch the model build a sentence one token at a time.
Token 0 of 4
LLMs recognize and reproduce patterns very well, but they don't possess understanding or comprehension the way humans do.
That distinction explains what these systems are good at and where they go wrong. An appropriate way to treat an LLM output is as a first draft from a fast, well-read assistant who is sometimes wrong. It can act as a useful starting point that needs to be reviewed and assessed by a human.
Having learned from enormous amounts of text, an LLM can spot themes, repetitions and connections across far more material than a person could read in the same time.
It can turn rough notes into a first draft, condense a long report into key points, or rewrite the same content for a different audience or format, such as an email, a list or a table.
Because it has learned from text in many languages, it can translate quickly and often keep the tone as well as the meaning. Results are weaker for languages with less training data.
Tasks like sorting feedback, tagging documents or pulling details out of forms can be done across thousands of items in minutes. A person still needs to check the results.
LLMs sometimes “hallucinate”: they produce false information that reads as entirely plausible. The system has no reliable way to signal when it reaches a gap in its knowledge.
Generative AI works by recognizing patterns in existing data. What it produces is built from things people have already made, and it has no lived experience to check its ideas against.
Unless it’s connected to live information, a model only knows what was in its training data. The world keeps changing, so an answer based on outdated data can be wrong or irrelevant.
The same request phrased differently can produce a different answer.
Building on generative AI, reasoning models and agentic AI are the latest developments in the field.
Reasoning models are LLMs trained to work through a problem in steps before they answer: breaking it down, checking their work and trying another route when something doesn’t add up. This makes them better at maths, coding and planning, but slower and more costly to run. Their “thinking” is still prediction, so it can still be wrong.
Examples
Agentic AI systems, or AI agents, are given a goal rather than a single question. They plan the steps, use tools such as web search, code or other apps, check the results and keep going until the task is done. Because they act on someone’s behalf, their mistakes can have real consequences, so people need to set limits and review what they do.
Examples
As AI becomes more widespread, there's growing attention to how the technology carries both individual and systemic risks, affecting specific people and societies at large.
The choices that drive AI technologies, including how models are built, what data they are trained on, and where they are deployed, reflect the priorities of the organizations that build them.
These choices impact everyone who uses the technology or is affected by it, most of whom had no involvement in the making of these decisions.
Public-interest AI provides an alternate pathway for AI development and deployment that centres the needs of communities. It envisions AI technologies that are designed, built, and used to support outcomes that serve the long-term wellbeing of the public, and safeguard people’s democratic rights.
The concept of Public AI, supported by a wide community of experts and practitioners, supports AI systems that incorporate public control as much as possible, in as many as possible components of the AI stack, which includes compute power, data, to the models themselves, as well as the apps that make AI accessible to the public. Public AI calls for accessible, interoperable, and transparent AI systems that are accountable to the public and governed according to democratic principles.
Sandbox Labs is part of a global movement supporting Public-Interest AI and Public AI. In collaboration with the UPF-Barcelona School of Management and hosted at Norrsken House Barcelona, we bring together our expertise in digital technology, education, and policy to create the Public-Interest AI Accelerator, a new initiative aiming at skilling up today’s and tomorrow’s leaders to ensure the future of AI benefits people and societies at large.
Our pilot edition of the Accelerator took place in 2026. Over nine weeks, participants engaged with experts in law and digital rights, health, policy, computer science, and neuroscience to explore AI’s role in geopolitics and supply chains, the legal and human-rights questions it raises, what ethical governance actually looks like, and how public-interest AI gets built and implemented in the real world.
The Public-Interest AI Accelerator is designed to help participants engage critically with AI technologies and answer key questions such as who benefits? Who carries the risks? What trade-offs must be transparent? How can public policy leaders make sound decisions in an AI-world?