AI on a Leash? Reflections on Machine Control

Explore reflections on Artificial Intelligence control: how can we ensure machines operate for our benefit while guaranteeing safety, ethics, and responsible algorithmic governance.

Artificial intelligence is now a constant presence in our lives, often silent but increasingly pervasive. From the suggestions offered by our smartphones to medical diagnoses assisted by sophisticated algorithms, from self-driving cars to systems managing critical infrastructure, intelligent machines are weaving a dense fabric that envelops our daily lives. This omnipresence raises a fundamental question, an inquiry that goes far beyond simple technological curiosity: who really has control over this unstoppable force? Who holds the reins of these artificial minds that shape our present and, in all likelihood, will define our future?

Who Really Has Control Over AI?

The answer to this question is neither simple nor straightforward. At first, we might be tempted to point to the developers, the engineers who design and program these complex systems. Certainly, their technical expertise is indispensable in the creation of AI. As explored in our article on what artificial intelligence really is, they are the ones writing the lines of code, feeding the algorithms with enormous amounts of data, defining the neural architectures that allow machines to learn and evolve.

However, once an artificial intelligence system is released into the world, the dynamics of control become much more nuanced and intricate.

The Problem of Algorithmic Bias: When AI Inherits Our Prejudices

Consider, for example, the large language models that power chatbots and virtual assistants. They are trained on colossal amounts of text and code from the internet, a veritable mine of heterogeneous and often unfiltered information. In this machine learning process, the algorithm identifies patterns, establishes connections, and develops its own "understanding" of language. But this understanding is inevitably influenced by the data on which it was trained.

If this data contains distortions, implicit or explicit biases, the AI model could reproduce and even amplify them. As highlighted in this article on algorithmic bias, this phenomenon represents a crucial challenge for AI control. It is not a matter of malicious intent on the part of programmers, but a pitfall inherent in the process of learning from imperfect data.

The issue of bias in data is particularly sensitive in critical fields such as facial recognition or risk assessment systems used in judicial or credit contexts. If the training data is not representative of all segments of the population, the algorithm may show significantly different performance depending on ethnicity, gender, or other protected characteristics, leading to forms of algorithmic discrimination that are entirely unintentional but no less harmful. Fondazione Patrizio Paoletti and Agenda Digitale

A prime example is COMPAS, an algorithm used in some US judicial systems to predict a defendant's likelihood of recidivism. The problem with COMPAS was that this software exhibited a strong bias, leading to twice as many false positives for recidivism among Black defendants (with a rate of approximately 45%) compared to Caucasian defendants (where the system recorded a rate of 23%). Algorithmic bias: artificial intelligence also stumbles over prejudices

The Role of Major Technology Companies

Another fundamental aspect to consider is the role of the companies and organizations that develop and implement artificial intelligence. They are the ones who define the objectives, choose the training data, and decide how and where these technologies will be used. Market logic, economic interests, and corporate strategies play a decisive role in shaping the development and diffusion of AI.

As discussed in our in-depth article on AI and surveillance, this concentration of power raises important questions about the transparency and accountability of AI systems. How can we ensure that decisions made by increasingly sophisticated algorithms are fair, ethical, and in line with democratic values? Who is responsible when an artificial intelligence system makes a mistake or causes harm?

The "Black Box" Challenge

The intrinsic complexity of many AI models, the so-called "black boxes" whose internal workings are difficult to interpret even for experts, makes it even harder to assign responsibility and exercise effective control. This algorithmic opacity is one of the central issues in AI ethics, where transparency and understandability become fundamental requirements for responsible technology.

The Regulatory Framework: Towards Regulatory Control

The issue of AI control does not only concern developers or large tech companies. It is a challenge that concerns the entire society and has led legislators worldwide to intervene with increasingly structured regulatory frameworks.

The European AI Act: A Regulatory Model

The AI Regulation (EU Regulation 2024/1689) represents the world's first comprehensive legal framework on AI. The goal of the rules is to promote trustworthy AI in Europe. AI Law | Shaping Europe's digital future This legislative instrument establishes a clear set of risk-based rules for AI developers and deployers regarding specific uses of AI.

As of February 2, 2025, the provisions of the AI Act concerning systems posing unacceptable risks and digital literacy are in force. AI Act: in force from February 2 for high-risk systems and training The European regulation represents a concrete attempt to "put AI on a leash" through a risk-based approach, ranging from completely banned systems to those subject to rigorous controls.

Penalties and Liability

The effectiveness of this regulatory control also relies on a system of significant penalties. Fines can range from 7.5 million EUR, or 1.5% of global annual turnover, to 35 million EUR, or 7% of worldwide annual turnover, depending on the type of compliance violation. What is the EU AI Act? | IBM

Self-Regulation and Transparency Initiatives

Alongside regulatory efforts, self-regulation initiatives have emerged from the industry and the scientific community. Partnership on AI (PAI) is an independent 501(c)(3) non-profit organization originally established by a coalition of representatives from tech companies, civil society organizations, and academic institutions, supported by multi-year grants from Apple, Amazon, Meta, Google/DeepMind, IBM, and Microsoft. Partnership on AIWikipedia

PAI develops tools, recommendations, and other resources by inviting voices from across the AI community and beyond to share insights that can be synthesized into actionable guidelines. It then works to promote their adoption in practice, inform public policy, and advance public understanding. About – Partnership on AI

Current Challenges of Control

Speed of Development vs. Capacity for Control

One of the main problems in AI control is the time gap between technological development and the capacity to understand and regulate it. As highlighted in our article on ChatGPT and the Future of Communication, the evolution of AI systems proceeds at a dizzying pace, often outstripping institutions' ability to adapt regulatory and control frameworks.

The Globalization of AI

The global nature of AI development poses further challenges to control. While Europe develops the AI Act, other countries and regions adopt different approaches, creating potential regulatory conflicts and opportunities for "regulatory arbitrage" for tech companies.

Towards More Effective Control

Investments in Interpretable AI

It is crucial to invest in the research and development of "interpretable" and "transparent" AI—systems where decision-making is not an inscrutable mystery but can be understood and verified. Only through greater comprehensibility can we exercise more effective control and build solid trust in these technologies.

Education and Public Awareness

As highlighted in our article on 5 AI Tools for Beginners, it is essential to promote a culture of algorithmic responsibility, where those who design and use AI are aware of its potential ethical and social implications.

Multistakeholder Collaboration

Without intentional coordination, we risk creating a fragmented landscape where AI developers and deployers are unclear on the best practices for safe and responsible AI. New Report from Partnership on AI Aims to Advance Global Policy Alignment on AI Transparency The challenge of AI control requires collaboration between governments, companies, researchers, and civil society.

Frequently Asked Questions

Who really controls artificial intelligence today?

Control of AI is distributed among different actors: developers and tech companies that create the systems, governments that regulate them, and users who utilize them. There is no single entity that completely controls AI, which makes governance particularly complex.

Is the European AI Act sufficient to control AI?

The AI Act represents an important but not definitive step. It is the world's first comprehensive regulatory framework, but its effectiveness will depend on implementation and future technological evolution. It could serve as a model for other global regulations.

How can we prevent bias in artificial intelligence?

Preventing bias requires a multidimensional approach: more representative training data, diverse development teams, rigorous testing, and tools for continuous monitoring of systems in production. Transparency in algorithms is fundamental.

What happens if an AI system causes harm?

Liability varies depending on the jurisdiction and the type of system. The European AI Act establishes specific responsibilities for providers and deployers of AI systems, with penalties that can reach up to 7% of annual global turnover.

Is it possible to have total control over AI?

Total control is probably impossible and may not even be desirable, as it could stifle innovation. The goal should be effective control that balances safety, ethics, and technological progress.

Conclusions: A Delicate Balance

The power of artificial intelligence is undeniable and its potential to improve our lives is immense. However, this power comes with great responsibility. The question of who controls intelligent machines is not just a technical issue, but an ethical, social, and political challenge that defines the kind of future we want to build.

The answer does not lie in completely shackling AI, nor in letting it develop without controls. As highlighted by regulatory initiatives like the AI Act and collaborations like the Partnership on AI, the best path forward appears to be one of distributed, transparent, and adaptive control, involving all stakeholders in society.

Ensuring this power is exercised responsibly, transparently, and for the common good is an urgent and indispensable task for all of humanity. It is not about stifling progress, but about guiding it with wisdom and foresight, keeping the compass of fundamental human values firmly in our hands.

The challenge of AI control will continue to evolve alongside the technology itself. What remains constant is the need for vigilance, dialogue, and collective commitment to ensure that artificial intelligence remains a tool in the service of humanity, and not the other way around.