For decades, artificial intelligence (AI) seemed like a field of academic or scientific interest, but far from the needs of many companies. In market applications its use was mainly limited to highly controlled environments or with close human supervision. This meant that, in many cases, it was not practical for much of the business world.
In the last decade everything seems to have changed: Interest in AI has risen dramatically. There is not a day that goes by that we don’t hear about AI. in the media, on the street, at work, etc.; of the multiple uses given to AI, sometimes with an overuse of the term “AI”. But regardless of the possible overuse of the term, it is clear that something has changed so that this term, which a few years ago seemed to be something only in the scientific or academic field, now has endless potential applications.
It seems quite clear that the great accuracy achieved by AI systems in recent years is behind this great leap in their possibilities of use. Tasks that years ago required constant correction by a human, today may only require occasional interventions or even allow for the total absence of intervention. This has meant that, from a business point of view, AI has become an interesting area for companies that can help improve their internal processes and productivity.
But every technology tends to have a positive and a negative side, and AI could not be less so. Behind the great advances and increased applicability of AI, some concerns related to issues have emerged. of opacity, privacy, bias, deepfakes, etc. which the emergence of ChatGPT has made even more evident. Although it was ChatGPT that brought this issue to the attention of the general public, concerns about issues related to ethics in AI began much earlier, with the use of data from multiple sources. The massive use of data led to concerns about the biases contained in this data, which machine learning (ML) algorithms further amplified.. The use of deep learning (DL) algorithms in areas such as justice, banking or medicine highlighted their lack of transparency regarding how and why they make certain decisions. But the matter does not end here: Regarding explainability, there is a paradox that, in general, the more precise an algorithm achieves, the less transparent it is, so DL tends to achieve more precision than ML, but the way in which it has made decisions is less explainable, to the point that in many cases they behave like black boxes that make it impossible to know how or why certain decisions have been made.
That is why, now more than ever, the regulation of AI has become a cause for concern for governments, companies, individuals, etc. Even the European Commission itself has been working for years on developing a regulatory framework for AI that ensures that Europeans can have more confidence in it.
The CCMA, as a company belonging to the public sector, is not exempt from this concern. Quite the opposite: To all this concern is added the responsibility it has as a public medium, both from the point of view of social responsibility, transparency, sustainability, etc., and from the point of view of credibility and prestige. Currently, the CCMA has content recommendation systems for users, facial recognition systems, logo recognition to classify and search its content, audio-to-text transcription systems or a semi-automatic subtitling system. In addition, work is also being done on the possible use of automatic summary generation systems in Catalan, fully automatic subtitling, genre detection, detection of sustainable development goal topics in its content, automatic generation of meteorological information through avatars, etc.
All these tasks require ML and DL algorithms such as eXtreme Gradient Boosting (XGBoost), Light Gradient-Boosting Machine (LightGBM), Hidden Markov Models (HMM), Convolutional Neural Networks (CNN), Transformers, etc., which can generate or amplify biases when making decisions or generating new content. Although, in most cases, behind the use of these algorithms we find humans supervising the results, this is not always the case. The clearest case is that of content recommenders for users. It is evident that a human will not analyze the thousands of recommendations that are generated daily; Moreover, even if a human wanted to have some clues about the criteria followed by the algorithm to generate the recommendations, in many cases it would not be possible given the same opacity of the algorithm, which makes it difficult to know whether these recommendations are in line with the objectives and principles of the CCMA’s function as a public medium. This lack of transparency of recommendation algorithms It poses a great challenge to public media with respect to their independence, responsibility, justice, etc.
Recommendation is not the only challenge that public media faces with AI. The biases introduced or amplified by ML/DL systems can also generate problems in the field of documentation in the selection of audiovisual material from searches or automatic classification during program production. Algorithms that have been trained with data external to the CCMA may contain biases of gender, ethnicity, ideology, etc. which require constant supervision by a human. But those trained with internal CCMA data are also not free from bias if this data has not been selected following a correctly defined protocol.
The supervision carried out by the CCMA to minimize the biases of its content will make perfect sense as long as the discriminatory aspects that could be generated by the same AI tools used both in the content production process and in the rest of internal processes are taken into account.
That is why public media are faced with the great challenge of using AI while respecting the principles that give meaning to being a public service, which It will force compromises to be made between using high-performance, non-transparent AI systems or lower-performance, but more transparent systems. At the same time, the adoption of this commitment should not mean a loss of competitiveness with respect to the competition.