The introduction of artificial intelligence (AI) in various aspects of life makes you think about cultural, technological, as well as legal consequences that this process entails. Operation of AI algorithms of search and results output at the request of users will obtain regulation as it happened once with legal framework for the Internet.
The omnipotence of the algorithm: how the first search systems worked
Today, at times AI demonstrates terrifying possibilities, sometimes they are like superpowers. For example, a few years ago “Portrait of Edmond Belami” was sold at an auction Christie’s for $430,000. The author of a such successful canvas was a neural network of the generative adversarial networks (GANs) class.
AI algorithms have been writing poetry for a long time, every year they do it better – in bookstores you can already find collections of “neurolyrics”.
Cybercourt has been operating in China since 2019, where in some cases with a typical composition there are proceedings led online by an AI judge – an algorithm, trained on a huge number of real cases reviewed by “human” colleagues.
However, this was not always the case. Once upon a time, SEO specialists gave a worthy answer to AI algorithms using a set of different techniques when composing text, and also when it is formatted as web pages. Techniques included oversaturation with key words and wording, and various types of disguise, so it was “black SEO” in all its glory.
This approach reduced the quality of search work: the engines were not offering among the first objectively relevant pages, but showed those that managed to fool output algorithm.
Currently, AI allows you to filter out such spam and issue on demand only genuine original content. At the same time, each change in work of search algorithms force companies to reconsider approaches to creating and publishing their content on web resources.
For example, AI technologies such as natural language or Images processing, allow you to recognize instantly words and visual materials, using them as inputs for the search query.
In general, modern AI search is a complex network of algorithms based on the artificial intelligence technology that determines the selection of search results, priority and order of their output.
As a rule, the exact mechanisms and principles of algorithms are the developer’s intellectual property. It is becoming more and more difficult to “outwit” modern algorithms.
What is happening now
Modern search engines rely heavily on AI algorithms. They are responsible for which search results are displayed on user’s demand, as well as for the selection of related topics offered for additional study.
Search engines also rely on AI to determine how content will be ranked when displayed. The algorithms used by these systems have many rules that determine the priority of different factors – from keywords to user experience (search history and so on).
Modern search on the Web uses optimization methods as standard results at the request of previous sessions. However, today search engine operators have made serious progress in processing requests.
One of the main growth points here is providing the most relevant content on demand. To do this, search engines must understand clearly what they are asked, taking into account the realities and context.
For example, Google’s development – a trained BERT language model – allows the company’s search engine to understand the full sentences. That means, that BERT can understand the semantic context of the search, not just the key words in the query.
At the same time, in the process of improving AI algorithms for search there is a nuance. While developers are improving various work aspects of search and recommendation engines, no one outside the company has a complete understanding of how these algorithms are constructed and how they make decisions.
It’s yet too early
Right now, it is hard to talk about violations of the law in this area in an environment, when the regulatory framework for regulating recommendation and search services are just being formed.
At the moment, making judgments about the actions of such services in Russia, when it is not about clearly illegal information, the dissemination of which falls under the regulation of Federal Law FZ-149, is possible only in the aspect of AI technology ethics.
And here, first, you need to deal with what can be called negative patterns of the work of AI algorithms and to determine the detrimental impact.
It is necessary to understand what specific risks we are talking about – content risks (provision of illegal information), consumer risks (inclination to buy and misleading about the consumer qualities of the product/service), communication risks (acquaintances on the Web that pose a threat, cyberbullying) or about the risks in information security (actions leading to a leak of personal data).
It is impossible to oblige developers of search and recommendation services to take into account all risks. Otherwise, it will have a material slowing-down influence on the development of technologies, because, in a sense, the developer will have to implement in the service a full-fledged security monitor and content filtering features, which is separate expert field of information security.
Work is underway
Nevertheless, attempts to take the work of search algorithms at least under some control is conducted: in the relevant bill on the regulation of advisory services in the Russian Federation there are some very useful theses.
For example, about increasing user awareness of operation features of the algorithm and the data used by it, as well as about the possibility to abandon completely the recommendations.
Obviously, such proposals have a negative side – the developer will have to reveal the know-how underlying the operation of the algorithm, “giving” competitors basic principles and developments underlying it.
It is very important here to find a balance and target regulation correctly, ensuring human protection and implementing humanistic approaches, but not stopping the development of technology in this area. For example, there is such a promising direction in terms of the use of AI, as commercially demanded services of personalization of product offers.
It is also necessary to take into account the popularity of the service and the type of search issuance object. Let us say if this is news, opinions and knowledge, and a specific service and its recommendations are used by millions of people, then such a service should be very careful in providing information and forming recommendations. Otherwise, there is a danger of turning the algorithm into a “guide” and even in “teacher”.
There will always be risks of placing the user in information isolation, when all shown materials will tend to be homogeneous, in fact hypertrophying selected topics or the significance of one point of view.
In this perspective legislators, owners and developers of such services have a lot of work and calls that need to be solved in the dialogue, very neatly and thoughtfully.
Everything can be hacked and exploited
Also, do not forget about the eternal principle: “What one person constructed – the other can always break”. To be more precise: notwithstanding the perfect technology, people will always be able to find a way either to hack it directly, or to exploit its capabilities in their own benefit.
For example, in 2020, artist Simon Wekkert in Berlin “hacked” Google Maps. With the help of a small wheeled trolley in which he placed 99 old smartphones, he made a huge virtual traffic jam on the city streets. To do this, you just needed to turn on all the phones with launched Google Maps app and walk with them through the streets of the city. In reality all these streets were not even close to being overloaded with cars, but Google Maps recorded honestly the traffic jam.
And it’s all about the algorithm: Google maps work through data aggregation from several sources, but, as it turned out, they rely mainly on real users who have geolocation services enabled on their phones.
The artist studied the features of the service for about two years, and the result is really impressive. No “cyber attack”, hacking and other suspicious anti-law activity – resting completely legal and without much IT competencies he was able to adjust the traffic picture.
How could this be used in practice? For example, “strangling” competitors whose points of sale are located along the most attractive streets by patency criteria.
There is an inherent problem of AI algorithms in terms of conscientious exploitation of the principles of their work in potentially not entirely legal or at least unethical goals. Potential impact of one person’s actions on the behavior of thousands or more should cause fair regulators’ concerns.
Who should be responsible for the possible consequences in such cases? With such questions, the future legislation in the field of AI algorithms should deal.
By Alexey Sergeev, Head of the Axenix Machine Learning and Artificial Intelligence Practice