By akademiotoelektronik, 11/11/2022

Artificial intelligence: the keys to guarantee a framework of confidence

Establishing the necessary confidence in artificial intelligence systems (AI), in order to be able to withdraw only safety and profits, is fundamental.Now today, he observes, artificial intelligence challenges subjects of transparency, traceability, control of learning and developments by data, reliability, bias, security and confidentiality.Consequently, how to have an objective assessment of the design, the realization and, periodically, the operation, the enrichment and the maintenance of these systems?This is the question posed by the working group responsible for the AI theme and confidence in the Academy of Accounting and Financial Sciences and Techniques.

Launched in 2017, this group consists of figures, lawyers, teachers, consultants, organizations, associations and stakeholder institutions.His goal ?Think about regulations, challenges, risks, and the way of auditing an algorithm to be able to certify it "compliant".As such, they have studied European regulation projects as well as the work carried out and repository available, internationally.They analyzed the texts and situations identified to detect the limits, the biases, the risks, the responsibilities and thus identify the good practices and the framework necessary for auditability and certification.Bookie n ° 38 “Artificial intelligence and confidence: regulations, issues, risks, audit and certification“ is the fruit of this work.Among the experts invited to present their contribution during this webinarian devoted the advent of a confidence, Camille Rosenthal-Sabroux, professor emeritus at Paris Dauphine-Psl University (Laboratory of Analysis and Modeling of Systems for Helpto the decision) and Alain Bensoussan, lawyer at Lexing Alain Bensoussan, presented the human and legal aspects.

The human dimension

For Camille Rosenthal-Sabroux, professor emeritus at Paris Dauphine-Psl University (analysis laboratory and modeling of systems for decision-making), taking into account the technical dimension by intelligent robots must absolutely notmake the human dimension disappear, but on the contrary to amplify it.Indeed, by releasing more time for humans, it thus increases its availability and its potential importance."You have to be aware of what AI-based systems are able to do, what are their limits, how can they help man.For this, the latter must remain at the heart of the reflection, ”she explains.Man must understand and master AI -based systems.The professor defends to do this a multidisciplinary approach.Also, the audit of algorithms by an expert is a key factor in confidence and it is an aid to understand these systems.

According to the professor, we must at the same time improve our knowledge of AI systems.It is also necessary to inform and train employers, executives, decision -makers in companies, in order to give confidence to users.It is finally necessary to apply good practices in terms of design and exploitation of AI -based systems.

The point of view of a specialist lawyer

"We have exceeded, for a good part of the countries, ethical reflection, to be in legal action", explains Alain Bensoussan, of Lexing Alain Bensoussan Avocats.who specifies that nearly a hundred regulations exist in the world on law and AI, while an equally important regulation project as the GDPR is in preparation at European level.

The risk approach (with a level that varies according to the sectors concerned), prevails in most countries of the world and the EU is part of this.Therefore, there is no ban, except in special cases."We are experimenting on one side and we certify on the other.All our work relates to this need to certify.Indeed, in view of today's technologies, the AI are autonomous, they escape their creator, "said Alain Bensoussan.This approach by freedom therefore aims to find a balance between the need to innovate and that of guarding against the dangers conveyed by AI.

True concentration of the experience of existing ethics in the world, this European regulation aims to be minimalist.It strives to organize certain prohibitions in a limited manner, to supervise high -risk systems, on a principle of AI freedom.With all the same a strong requirement."The AI must be designed, carried out, set up, and maintained in accordance with fundamental rights," recalls Alain Bensoussan, who specifies that it is a question of building through human-robot interaction and not one againstthe other.

The text first plans to prohibit using AI to alter the confidence of people "We cannot go below the floor of the intimacy of people in order to discover things that they would not want to makeapparent, ”underlines the lawyer.Using AI on vulnerable people is also prohibited.Indeed, these do not have enough discernment to oppose the "power" of this AI.Another prohibited, in particular made famous by its treatment in a series, concerns social rating.Biometrics at a distance in real time in public space should also be prohibited, except for the purposes of protecting victims or the fight against serious offenses against terrorism."These prohibitions will create an AI" design by dignity ", under the terms of article 1 of the EU fundamental rights charter," explains Alain Bensoussan.

A second category concerns AI located at the heart of safety organization systems.In this case, it is necessary to set up a safety plan and a compliance plan.When AI interacts with humans, it is important that humans understand that they are in front of an AI, for Alain Bensoussan.There is therefore an obligation to give them this information.

Finally, concerning high-risk AI, strict requirements are also laid and in particular the obligation to keep certain elements (archives of the data set used to cause and test the AI, or even data systemssame, or documentation relating to programming techniques).Administrative sanctions, stronger than those relating to the GDPR, could be implemented (30 million euros fine or 6 % of global turnover in the event of implementation of a prohibited AI system).Finally, Alain Bensoussan pleads for appropriate human control in order to reduce the risks as much as possible."The idea is not to abandon high -risk systems only to machines, but to allow human intervention from design to implementation and until use," continues the specialist.A quality approach (with certification) should prevent cyber damage.

Alain Bensoussan de conclure : « gageons que cette régulation deviendra pour l’ensemble des pays du monde l’élément-clé pour un 21^e siècle de confiance grâce à l’IA ».