Social Impact of New Technologies
Technologies and Their Social Context
Every significant technology gives rise to questions that extend far beyond the realm of engineering. The printing press forced us to rethink authorship and censorship. Photography raised concerns about authenticity and the right to one's image. The Internet brought up issues of privacy and the distribution of power over information. Artificial intelligence systems are no exception: they confront us with problems that cannot be solved by simply updating code or passing a single law.
It is crucial to understand the nature of these challenges. They are not «defects» of the technology in a technical sense: the systems are functioning exactly as they were designed. Ethical questions arise not because AI is inherently «good» or «bad», but because the technology actively interacts with society, institutions, economic interests, and deeply rooted practices. It is this very interaction that creates tension.
In this article, we will examine four areas where this tension is felt most acutely: authorship, privacy, data bias, and the distribution of responsibility. Our goal is not to provide ready-made answers, but to help form a robust framework for the independent analysis of these issues.
AI Authorship and Copyright Challenges
Authorship: Who Stands Behind the Result
When a language model generates text or a diffusion model creates an image, a question arises that seems simple only at first glance: who is the author of the resulting output?
The traditional concept of authorship assumes a subject who makes decisions, possesses a creative intent, and is responsible for the outcome of the work. It emerged in a specific historical context as a legal and cultural tool establishing the link between the creator and the creation. Generative systems break this link from several sides at once.
First, the model itself is trained on colossal datasets of works created by humans. Its ability to generate content is a direct consequence of analyzing works, each of which has an author. The model does not invent a style from scratch: it reproduces, combines, and interpolates patterns extracted from real works. To what extent this can be considered the system's «own» creativity is a question that currently has no settled solution.
Second, the user composing the text request (the prompt) is also not an author in the classical sense. They set the direction but do not control every nuance of the result – much like a director does not personally draw every frame of an animation, even though their artistic vision determines the outcome. The degree of human creative involvement in the generation process can vary from minimal to quite significant.
Third, the organization or developer who created the model shapes its «taste» and limitations through architecture, data selection, fine-tuning, and filters. This is also a form of influence on the final result.
Current legal systems are adapting to this new reality at varying speeds. In some jurisdictions, works created without direct human involvement cannot be protected by copyright. In others, criteria for «substantial creative contribution» are being discussed. None of the existing concepts yet provide an exhaustive answer, as such questions have been formulated for the first time in history.
AI Data Privacy and User Consent
Privacy: Data as Raw Material
Modern large-scale models are trained on massive datasets collected from open sources: texts, images, software code, and dialogues. This poses an important question: to what extent did the people who created this content consent to its use as training material?
Historically, publishing something on the Internet implied a willingness for public reading and distribution. However, using publicly available text to train a commercial model is an action different in nature. Here, it is not just a matter of familiarization, but the extraction of patterns that are then reproduced in other contexts. While the model does not «remember» specific texts in the conventional sense, its capabilities are shaped precisely by them.
A separate aspect is personal data. If names, addresses, or medical information end up in the training corpus, the model may accidentally reproduce them in response to certain queries. This is not a deliberate choice by the system, but a consequence of training on insufficiently cleaned data. Nevertheless, the consequences of such a process are quite real.
No less significant is the other side of privacy – the information that users provide to systems during a dialogue. Every query to a language model carries information about the person: their interests, professional activities, and personal circumstances. How this data is stored and protected is determined by the policies of specific organizations, rather than the properties of the technology itself.
The problem of privacy in the context of AI is thus twofold: it concerns both those whose data was used for training and those who interact with the system today.
Algorithmic Bias in Artificial Intelligence
Bias: Data Reflects the World as It Was
One of the most studied and, at the same time, subtle issues is the bias of AI systems. It is often called «algorithmic bias», which is not entirely accurate. An algorithm in itself is neutral: it performs exactly the tasks it is optimized for. Bias is generated by data that reflects the real world with its historical inequality, structural imbalances, and cultural stereotypes.
If a system was trained on texts where certain professions are predominantly associated with one gender, it will reproduce that association. Not out of a desire to discriminate, but because that is the statistical structure of the data. If certain populations are underrepresented in medical databases, a diagnostic system will work less accurately for them, even if its overall efficiency is high.
This gives rise to several interconnected problems:
First is the difficulty of detection. Bias often remains invisible until it manifests in specific incidents. Systems can demonstrate high average accuracy while hiding systematic errors regarding specific groups.
Second is scale. If a human making a biased decision affects a limited circle of people, an automated system replicates the same error millions of times. Scale turns a statistical artifact into a systemic practice.
Third is the feedback loop. Data regarding the system's own decisions can become training material for future generations of models, thereby entrenching and amplifying the original patterns.
At the same time, «cleaning» data of bias is not as trivial a task as it seems. The decision of what to consider bias and what to consider a true reflection of reality is itself a value-based choice. Different communities and cultures may answer this question differently.
Accountability and Responsibility in AI Decisions
Responsibility: Who Is Accountable for the System's Decision
When a human makes a decision – denies a loan, makes a diagnosis, or passes a sentence – it is clear who bears responsibility. When a similar decision is made or supported by an automated system, the situation becomes significantly more complex.
A typical process involves several participants (actors). The model developer created the architecture but did not know all the scenarios of its use. The organization that implemented the system chose the tool but did not develop the algorithm. The user interacts with the result but does not manage it directly. And the person whose life is affected by the decision may not have been involved in any of the stages at all.
Responsibility in such a situation does not vanish; it becomes «diluted».This phenomenon is sometimes called the «problem of many hands»: when many participants are involved, each of whom may have acted in good faith, establishing ultimate responsibility becomes extremely difficult.
It is also significant that the level of autonomy in systems varies. It is one thing to have a recommendation system where the final choice remains with a human. It is quite another to have a system whose decisions are executed automatically. The higher the level of autonomy, the more acute the question of responsibility for an error becomes.
Another important aspect is explainability. Many modern models, especially neural networks, operate on the «black box» principle: their conclusions are statistically justified but difficult to interpret in terms understandable to humans. If a system cannot explain the reasons for its decision, the opportunities to contest it remain limited.
This does not mean that responsibility cannot be distributed in principle. But it indicates that legal and social mechanisms created for situations with a clearly defined subject require serious rethinking.
Ethics as a Space for Open Questions
All four aspects considered share one thing in common: they are not purely technical problems. These challenges arise at the intersection of technology, society, law, economics, and culture, and that is precisely why they do not have universal, once-and-for-all settled answers.
This is not a cause for alarm or an argument against progress. It is a characteristic of any significant social transformation: questions always outpace answers. What matters is not the presence of ready-made recipes, but the ability to correctly formulate the questions themselves.
What does «correctly» mean? It means not attributing intentions to systems that they do not have. Not placing responsibility on technology that belongs to people and organizations. Not searching for a single culprit where systemic effects are at play. And not demanding a final solution in a situation that continues to evolve dynamically.
Ethical questions of AI are not properties of the «machine itself».They are the result of the interaction between algorithms, data, economic incentives, and social institutions. The ability to see this interaction is one of the most practically important skills for everyone living and working in the era of the mass proliferation of intelligent systems.