Artificial intelligence: definition, types, uses and dilemmas

La Artificial Intelligence AI has crept into our daily lives almost without us noticing: mobile phones, cars, platforms streamingMedical diagnoses, chatbots… all of this works thanks to a set of techniques that allow machines to imitate certain capabilities that we have always associated with human beings, such as learning, reasoning or understanding language.

When we talk about AI, we're not referring to a single magical technology, but to a large umbrella that encompasses algorithms, mathematical models, hardware and dataUnder this umbrella coexist things as diverse as deep neural networks, symbolic expert systems, autonomous robots, conversational assistants, or generative models capable of creating texts, images, music, or code from scratch.

What exactly is artificial intelligence?

UNESCO, through its World Commission on the Ethics of Scientific Knowledge and Technology (COMEST), described AI in 2019 as the field that deals with machines capable of imitating functions of human intelligenceincluding perception, learning, reasoning, problem-solving, linguistic interaction, and even the creation of original works.

In everyday language we often use "artificial intelligence" to talk about any system that imitates human cognitive processesTo perceive, learn, reason, or make decisions in the face of more or less complex problems. Some authors, such as Andreas Kaplan and Michael Haenlein, focus on a system's capacity to correctly interpret external dataLearn from them and use what you have learned to achieve specific goals by adapting flexibly.

As technology advances, many tools once considered AI are no longer seen as "intelligent" and are now simply considered ordinary software. A very clear example is... optical character recognition (OCR): Today it seems like a standard function, but decades ago it was at the forefront of artificial intelligence research.

In essence, we can understand AI as the ability of certain machines or programs to Use algorithms, learn from the data, and apply that learning when they make decisions in a way similar to how a person would. From this perspective, both symbolic approaches (based on explicit rules) and the numerical and statistical methods that dominate modern AI fit the bill.

Some researchers, such as Takeyas, frame it directly as a branch of computer science that studies computational models capable of carrying out typically human activities based primarily on reasoning and observable behavior.

Types of artificial intelligence systems

Stuart J. Russell and Peter Norvig, two of the most influential voices in the field, propose a widely used classification that distinguishes four major families of systems based on how they behave and whether they focus on thinking or acting:

• Systems that think like humansThey attempt to replicate human mental processes. This includes, for example, many neural networks inspired by the biological brain and models that seek to mimic the human way of solving problems, making decisions, or learning.
• Systems that act like humansThey focus on resembling us in their observable behavior. Humanoid robotics and many conversational agents fall into this category because the goal is for interaction with them to be perceived as natural and human.
• Systems that think rationallyThey emphasize the logic and internal consistency of the reasoning, regardless of whether it resembles actual human thought. Classical, rule-based expert systems are a typical example of this approach.
• Systems that act rationallyThese are the so-called rational agents, designed to choose at any given moment the action that maximizes a certain notion of utility or benefit, given the circumstances and the information available to them.

Another very common classification distinguishes AI based on its scope and power. The so-called Narrow or weak AI It is the one that dominates today: systems designed for specific tasks (recommending content, recognizing faces, translating text, driving a vehicle) that do them very well, but do not go beyond that scope.

At the opposite extreme is the hypothetical general artificial intelligence Or a strong one: a machine capable of performing any intellectual task we currently associate with humans or even animals, learning autonomously and adapting to radically different contexts. Beyond that, there is still speculation about a super intelligent AIwhich would far exceed human cognitive capacity in virtually all relevant domains.

Schools of AI: symbolic and computational

Historically, there are usually two main approaches or “schools” within AI. On the one hand, there is the conventional or symbolic artificial intelligence, also called deductive, which bases its strength on explicitly representing knowledge through symbols, logical rules, ontologies or structured probabilities.

Along these lines we find, for example, the case-based reasoning (solving current problems by seeking analogies with cases already solved), rule-based expert systems, Bayesian networks that model uncertainty with probability, or approaches based on autonomous behaviors regulated by logics and rules.

Facing her, the so-called computational intelligence (or sub-symbolic, inductive) relies on models that learn from empirical data, incrementally adjusting their internal parameters. Neural networks, evolutionary algorithms, deep learning, and swarm techniques are part of this approach.

Computational intelligence pursues a dual objective: on the one hand, to understand the principles that enable intelligent behavior in natural or artificial systems; on the other hand, to translate that understanding into concrete procedures to design systems that act with a certain degree of autonomy and adaptation.

Machine learning and deep learning

The heart of modern AI is the automatic learning (Machine learning), which investigates algorithms capable of improving their performance with experience. Instead of programming step by step how to solve a problem, a model is trained with examples so that it discovers the relevant patterns on its own.

Within machine learning, a distinction is usually made between supervised learning (the data comes labeled and the system learns to predict those labels), unsupervised (look for structures and regularities without prior labels, such as groupings) and intermediate approaches such as semi-supervised or reinforcement learning.

Supervised learning is used both for sort out (deciding which category something belongs to: spam or non-spam email, customer who will unsubscribe or not) as for regression (predicting continuous numerical values). Unsupervised, on the other hand, allows for the discovery of customer segments, size reduction, or the detection of anomalies without any human having to indicate what is relevant beforehand.

A particularly powerful subset is the deep learningBased on artificial neural networks with many stacked layers, these networks are capable of automatically extracting high-level representations from raw data, making them very effective with complex information such as images, audio, video, or free text; and they allow tasks such as improve the quality of an online image.

Thanks to these in-depth techniques, applications that not so long ago seemed like science fiction have flourished: reliable voice recognitionHigh-performance computer vision, reasonable machine translation between dozens of languages, or language models that generate coherent paragraphs without manual programming of linguistic rules.

Generative AI, prompts, and multimodal models

One of the branches that has been generating the most noise lately is the generative artificial intelligenceThis type of system learns the statistical structure of large datasets (text, images, code, audio, etc.) and is then able to produce new content which, without being a literal copy, maintains similar features and patterns.

The most well-known examples are the large language models (LLM) such as GPT-3, GPT-4, Gemini or Claude, which allow you to converse, write documents, summarize information, generate code or answer complex questions from a simple textual instruction.

That instruction is known as promptA prompt, question, or set of guidelines given to the model to help them understand the task and the tone or format in which it should be performed. The quality of the result depends heavily on the prompt qualityIt depends on how well we define the context and how clear our request is.

Something similar has happened with image generation tools like Stable Diffusion, Midjourney, or DALL-E, capable of producing hyperrealistic illustrations or photographs from a natural language description, and even tools that They turn photos into drawingsToday, generative AI is not limited to text and images: it can also create video, music, synthetic voices, and sound effects with an increasing level of realism.

The most advanced models are already multimodalIn other words, they can receive and combine data from different sources (text, image, audio, video) to better understand the situation and generate richer responses. This approach is inspired by how we perceive the world using multiple senses simultaneously.

Branches and advanced variants of AI

In addition to the distinction by power levels (narrow, general, superintelligent), several terms have been coined in the literature to refer to specific areas. The so-called Explainable AI (XAI) encompasses methods and tools designed to help people understand why a model has arrived at a certain prediction or decision.

In parallel, discussions are taking place about the Friendly AI, a theoretical form of strong AI explicitly focused on having a beneficial and safe impact on humanity, especially thinking about recursive self-improvement systems that could grow in capacity at high speed.

La Quantum AI It is an interdisciplinary field that seeks to leverage quantum algorithms to accelerate typical artificial intelligence tasks, especially in machine learning and optimization. Some studies point to quadratic performance gains for certain key operations.

Debates about strong artificial intelligence (IGA) and its possible arrival. Some authors see it as feasible in a matter of decades; others believe it will take much longer, and there are those who think that perhaps a true equivalence with general human intelligence will never be achieved.

From the first automata to giant models

Although the term “artificial intelligence” was formally coined in 1955–1956, the ideas underlying the discipline go back much further. Philosophers such as Aristotle They had already reflected on logical rules to reach rational conclusions, and throughout the centuries automata and mechanical devices appeared capable of behaviors that were surprising for their time.

The 20th century saw the great foundational milestones. Alan Turing formulated his famous universal machine and poses the well-known “Turing test” to evaluate whether a machine can be considered intelligent in human eyes. Ada Lovelace, even earlier, had intuited that machines could go beyond simple numerical calculations.

In 1956 John McCarthy led the famous Dartmouth Conferencewhere the term "artificial intelligence" was consolidated and very optimistic ambitions were set for the following decade, which were never fully realized and led to the so-called "AI winters", periods of disillusionment and funding cuts.

During the following decades, advances such as the following occurred expert systems (capable of reasoning with domain knowledge), the first mobile robots equipped with perception, AI-oriented programming languages ​​such as LISP or PROLOG, early models of neural networks such as the perceptron, and later the rise and fall of these networks after criticisms such as those of Minsky and Papert.

In the 80s and 1990s, neural networks experienced a strong resurgence thanks to the backpropagation algorithm, while the 90s brought highly publicized milestones, such as the victory of Deep Blue about Garry Kasparov in chess. From 2010 onwards, with the rise of big data and computing power, came the boom of deep learning and the results in vision, voice, and natural language skyrocket.

Current applications of artificial intelligence

Today, AI is routinely used in fields as diverse as economics, medicine, engineering, transport, communications, or defenseMany video games incorporate AI components to manage the difficulty or behavior of non-human opponents.

In the business environment, AI helps to automate repetitive tasks and tedious tasks (invoice processing, inventory control, document classification), while driving strategic decisions through advanced analytics and predictive models that detect patterns difficult to see at first glance.

In the healthcare sector, algorithms can help in the diagnostic imagingIn the early detection of diseases, the design of personalized treatments, and the monitoring of chronic patients, they are used in finance. In finance, they enable fraud detection systems, credit risk assessment, and real-time monitoring of suspicious transactions.

There are also applications directly visible to the end user: digital assistants on mobile phones, smart speakers, automatic translators, content recommendation platforms, spam filters, navigation and driving assistance systems, or productivity tools that summarize meetings and emails, and apps to improve the quality of photos.

In the workplace, platforms based entirely on AI are even emerging, capable of automate mass applications and simplify the job search by sending resumes to multiple offers in parallel without constant manual intervention from the user.

Natural language processing and perception

El natural language processing Natural Language Processing (NLP) aims to enable machines to read, interpret, and generate human language functionally. Some of its most visible applications include automatic question answering systems, information extraction from large volumes of text, opinion mining, and translation between languages.

Traditional approaches relied heavily on word frequencies, grammatical rules, and relatively simple statistical methods. Today, language models based on architectures of Transformers They have taken NLP to another level, allowing the generation of long and coherent texts and obtaining very rich semantic representations of sentences.

Meanwhile, the so-called machine perception It encompasses the ability of artificial systems to interpret data from sensors: cameras, microphones, LiDAR, radar, ultrasound, touch sensors, or even radio frequency signals. Computer vision addresses challenges such as object detection and segmentation, facial recognition, pose estimation, and the understanding of entire scenes.

These systems must deal with the inherent ambiguity of sensory information. For example, the same pixel pattern in an image could correspond to objects of very different sizes depending on their distance from the camera, and the algorithm must rely on plausible models of the world to decide which interpretation is most reasonable.

AI in education and public policy

Education is one of the fields where AI has the greatest transformative potential and, at the same time, the most risks and open debates. It is being used to personalize content, support assessment, generate teaching materials or to help students with special needs through adaptive interfaces.

Organizations such as UNESCO, the OECD, the European Commission, UNICEF, and the World Economic Forum have developed recommendations to ensure that AI in the classroom is used in a way that is appropriate and effective. ethical, transparent, fair and respectful of privacyIt is emphasized that it should serve as a complement to the teaching staff, not as a substitute for creativity and human judgment.

In recent years, proposals for educational regulations have emerged that establish principles such as algorithmic transparency, protection of student data, non-discrimination, the responsibility of teachers and institutions, the need for specific teacher training, and the continuous evaluation of the real impact of these technologies.

Among the benefits are more dynamic classes, early detection of learning difficulties, and support for inclusion. Among the risks mentioned are... bias in data, the danger of excessive dependence on students and teachers themselves, the possible erosion of written and oral skills, or the lack of interest in autonomous research if automatic assistants are overused.

In parallel, many governments and international organizations are developing national AI strategies and regulatory frameworks, such as the proposed AI Law in the European Union, which aims to harmonize rules on the use of high-risk systems and set obligations regarding security, transparency and data governance.

Ethics, risks and social debates surrounding AI

The massive deployment of AI has profound implications in areas such as employment, privacy, security, equal opportunities, and even the concept of rights and responsibilities. This is why an entire field dedicated to AI has flourished. ethics of artificial intelligence, with sub-branches such as roboethics (the relationship of people with robots) or machine ethics (the behavior of systems towards humans).

One of the recurring fears is the technological unemploymentThe automation of cognitive tasks could render many professionals obsolete. Past experiences in industrial automation show that this process can generate significant tensions if it is not accompanied by proactive training policies and a redistribution of benefits.

There are also concerns about the potential malicious uses of AI, from campaigns of disinformation and manipulation politics to the generation of hyperrealistic fake content (such as nudes fabricated from real photos of minors or images of public figures in fake situations) or the creation of autonomous weapons that make attack decisions without direct human supervision.

Prominent figures such as Stephen Hawking, as well as entrepreneurs and technology experts, have warned about the existential risks of advanced AI that is misaligned with human values. In 2023, several open letters signed by researchers, executives of major companies, and public thinkers called for action. pauses in the development of very powerful models and stricter global regulatory frameworks.

In the area of ​​privacy, machine learning systems require enormous volumes of data, which has led to large-scale collection and surveillance practices that many consider unacceptable. Approaches such as anonymization, differential privacy, and federated learning are being explored to mitigate these risks, although the debate remains very much alive.

Intellectual property, copyright, and AI creations

The emergence of generative AI has set off alarm bells in the field of intellectual propertyOn the one hand, many models are trained using copyrighted works without explicit agreements with the copyright holders, relying on doctrines such as "fair use" in some countries. On the other hand, the question arises as to who is the author of the generated content: the user, the company training the model, or the system itself?

Organizations such as the World Intellectual Property Organization (WIPO) have begun to thoroughly analyze these issues, but for now there is no clear consensus. There is debate about the extent to which a work generated by AI can be considered a “creation of the mind” in the traditional legal sense, and whether it would make sense to recognize any kind of limited legal personality to certain highly advanced systems so that they assume fiscal or civil responsibilities linked to their economic activity.

Lawsuits from authors and artists against AI companies for using their work without permission to train business models are piling up in the courts. At the same time, a segment of the creative community is calling for safeguards to ensure that AI does not completely erode people's ability to earn a living from their artistic and intellectual work.

The solutions being explored range from new licenses and remuneration agreements to specific regulatory frameworks for synthetic content. Meanwhile, warnings are issued about the danger of trying to solve all AI problems by simply expanding or stretching the scope of copyright without restraint, with the risk of blocking innovation that same right was intended to protect.

Overall, artificial intelligence has gone from being an academic promise to a ubiquitous technology with a direct impact on the economy, culture, politics, and daily life in just a few decades. Understanding its foundations, applications, limitations, and risks is key to harnessing its advantages while remaining mindful of the need for robust ethical and legal frameworks to sustain this powerful tool. at the service of the people and not the other way around.