Science

Deepfake, racism, autonomous weapons: how to give ethics to AI

On November 24, UNESCO adopted a draft recommendation on the ethics of artificial intelligence (AI) with the aim of providing universal ethical foundations to its member states. This announcement, important as we will see, can generate questions at first sight. Indeed, how can ethics be related to AI? Isn’t this concept reserved for individuals? To answer this question, let’s first start by going back to the basics of what ethics is, as well as understanding what AI is.

Ethics, prescriber of values ​​and laws

Remember that ethics is synonymous with morality, as well as current, meaning both “customs”. There are mainly three philosophical doctrines that allow us to define and think about ethics. Briefly summarizing, we can say that the ethics of the ancient Greek philosopher Aristotle consists of perfecting the virtues, combined with a way of being prudent, having the concern to respect the customs of the city. For the Prussian philosopher Immanuel Kant, on the other hand, ethics is a duty, that of the categorical imperative, which consists in acting only if the action at stake is universalizable, that is, it can serve as a model for other men. Finally, in utilitarian thought, whose precursors are the British thinkers Jeremy Bentham and John Stuart Mill, ethics consists of choosing actions according to their consequences, good or bad, seeking the happiness of the greatest number.

These three historical doctrines, developed by a series of more current philosophers (J. Habermas, H. Jonas, J. Derrida, P. Ricœur, etc.) have shaped the contemporary meaning of the concept of ethics as it is used in the debate. public. Today, the concept of ethics is generally associated with the opinions issued by advisory ethics committees on moral dilemmas, at the end of more or less inclusive and democratic deliberative processes.

These opinions can define values, principles or even norms of behavior, avoiding harm to others and aiming at human happiness in a just society. Norms, in particular, have a practical vocation through their natural transposition into laws, regulations or even codes of ethics, by defining the rights and duties of citizens, that is, by establishing limits to their behavior.

AI, science without self-awareness

Artificial intelligence, on the other hand, is a protean technology intended to reproduce or simulate human intelligence. At the crossroads between mathematics and computer science, this discipline is experiencing spectacular growth thanks to the arrival of big data and the dizzying calculation capabilities of microprocessors. In particular, the AI ​​allows to perform homogeneous groupings of sets of abstract objects on the basis of training data. This functionality, a priori simple, paradoxically finds surprising applications for which more complex mechanisms could be imagined.

Thus, grouping homogeneous images, for example photos of cats, allows image recognition. Similarly, the grouping of sounds (phonemes) that make up words allows voice recognition. Finally, associating objects based on a given criterion allows machines to reason and make decisions, for example grouping candidates suitable for hiring, that is, distinguishing them from those whose application is rejected.

However, regardless of the sophistication of the AI ​​and the impressiveness of its performance, the current AI is said to be “weak”, because it “only” allows the execution of a single specific task: playing the game of Go, identifying a tumor in an x-ray, detect fraudulent bank transactions, note the borrower’s risk of default, etc. A general artificial intelligence, capable of carrying out all the cognitive tasks typical of humans, even limited to the capacities of a two-year-old child, is not on the agenda, while the creation of an AI equipped with a form of self-control consciousness and thought remains a sweet dream with respect to current knowledge in neuroscience.

In short, an AI is nothing more than an adaptive computer program (which adapts to its training data in Machine Learning or to its context in reinforcement learning), without conscience, without values, without thoughts, without free will, incapable of discern right from wrong and therefore act voluntarily and consciously in accordance with morality or law. In addition, if an AI can mechanically produce legal results, that is, in accordance with the law, talking about an ethical behavior of an AI does not make sense, since acting ethically requires the possibility of a choice in conscience, made possible by Free Will.

AI is therefore amoral, like all technologies. But in the same way that mastery of electricity has led to applications that have become essential for humanity (heating, lighting, etc.) as well as harmful (gegene has also been used for torture purposes), AI offers benefits but also potentially harmful. uses for mankind.

AlphaGO is a weak AI: it knows how to play Go… but it can’t recognize the image of a cat.
Prachatai/Flickr, CC BY-NC-ND

In fact, although the beneficial applications of AI are indisputable (improved medical diagnoses, automatic emergency braking of vehicles, predictive maintenance, automatic translation, etc.), the risks associated with it do exist. Let’s think first of the democratic risk, with the information bubbles of social networks that polarize the debate, or with deepfakes, those falsified videos that allow anyone to say anything to anyone in a very realistic way. AI then enables pervasive surveillance to emerge, with facial recognition, including analysis of our private conversations. We can also think of all the decisions that AIs make automatically without it being possible to explain why. The latter may also be biased with, in particular, a proven risk of discrimination. Work done in the United States illustrates this last point and shows that African-American populations in the study were more penalized by jurisdictions using predictive justice AI. We can add the issue of lethal autonomous weapons, the externalities of AI in terms of carbon footprint or even its impact on employment and inequalities…

Monitor the use of AI

As we can see, the risks associated with AI applications are numerous, obvious, and serious. But if we think about it, it is not so much AI, as a technology, that is or is not ethical, but much more its uses, imagined and developed by humans, because only the latter have freedom of choice in conscience.

Alerted to these problems, manufacturers and the scientific community have recently embarked on research projects in ethical AI (Fair Machine Learning), aimed in particular at correcting the discrimination biases of the models but also at making the results of the algorithms. Explainability aims to make model results intelligible to users. However, although any algorithm is understood by its developer, the results of certain AI cannot be explained and justified in a reasoned way with respect to the complexity of the trained model. So we can talk about a black box, unacceptable for decisions that have an impact on people’s lives (hiring, credit, etc.). This question refers particularly to neural networks, models inspired by biological neurons and their most popular and powerful avatars, deep learning (or Deep Learning) and convolutional networks for which the number of parameters can be counted in thousands of millions. Explainability is therefore a necessary condition for the social acceptance of decisions based on AI models. In addition, its arrival would reinforce the confidence of citizens in this technology.

But if these projects go in the direction of greater justice and trust, this is not enough to cover the scope of the risks linked to AI, but only those related to the legality of the models. Thus, the issues, already mentioned, related to democratic risk, generalized surveillance or even lethal autonomous weapons escape these technical solutions because they are of a different nature, ethical (is it moral to authorize the application in question?), legal (is it legal? what are the rights and duties associated with the application, once it has been authorized), or even political (how to manage the economic, social and environmental consequences of the applications in question?).

Specifically, the draft recommendation for an ethics of AI, adopted by the 193 members of UNESCO on November 24, establishes a normative framework of values ​​and principles that constitutes a common base, prior to the formulation of laws and regulations. by the member states. . For example, one specific recommendation relates to the issue of discrimination:

“Member states must ensure that gender stereotypes and discriminatory biases are not transposed into AI systems, but are proactively identified and corrected. “

In addition, this text advocates for safeguards in all stages of the AI ​​life cycle, empowering stakeholders, individuals and organizations (and not AIs as we have seen), concerned about the legality, design, development and operation of applications based on this technology.

Therefore, this event represents a useful and important step forward, but two caveats should be mentioned:

  • First of all, these recommendations are not legally binding insofar as the Member States undertake, by virtue of the UNESCO Constitution (Art. IV, B, §4 and §6), to submit only the recommendations to the competent national authorities and then to report on them. the follow-up given to them through the presentation of a subsequent report.

  • Next, AI is seen as the instrument of the next economic revolution, making it an important geostrategic issue, both economically and militarily.

However, in an international context of exacerbated economic competition and tensions between the great powers (we are thinking first of all of Sino-US relations), one may wonder about the weight of these recommendations…

PHP Script, Elementor Pro Weadown, WordPress Theme, Fs Poster Plugin Nulled, Newspaper – News & WooCommerce WordPress Theme, Wordfence Premium Nulled, Dokan Pro Nulled, Plugins, Elementor Pro Weadown, Astra Pro Nulled, Premium Addons for Elementor, Yoast Nulled, Flatsome Nulled, Woocommerce Custom Product Ad, Wpml Nulled,Woodmart Theme Nulled, PW WooCommerce Gift Cards Pro Nulled, Avada 7.4 Nulled, Newspaper 11.2, Jannah Nulled, Jnews 8.1.0 Nulled, WP Reset Pro, Woodmart Theme Nulled, Business Consulting Nulled, Rank Math Seo Pro Weadown, Slider Revolution Nulled, Consulting 6.1.4 Nulled, WeaPlay, Nulledfire

Back to top button