As a leader in commerce and technology, and the largest global commerce platform for independent small retailers and entrepreneurs, eBay’s responsible use of artificial intelligence takes into consideration important values such as trust, ethics, fairness, explainability, safety and human fundamental rights.
At eBay we develop AI technologies and systems in respect of the principles of Accountable AI, and in coordination with government governance structures, laws and regulations. Below we illustrate some of the concrete applications we are investing in and how they support our vision for “transforming eBay into an adaptive AI-managed trusted marketplace & platform where the complex is made simple and customer experiences are fluid, natural and exceptional.”
We believe that policy-making whenever touching upon the topic of AI should draw from the key Accountable AI principles. At the end of each paragraph below, we illustrate how these principles can come to life in certain policy recommendations.
Explainable / transparent
Where AI will increasingly be used to supplement or replace human decision-making, this artificial process needs to remain rationally explainable to human beings. Not only will this make AI more acceptable to society, it will also allow developers to identify and fix potential biases that could otherwise be amplified through AI.
However, AI systems are complex and cannot always be fully explained without costly analysis. In other words, often significant resources must be dedicated in order for transparency to provide the intended effect, which is to make the explanation accessible to various types of audiences, and not just the most technical expert. As such, ideal explainability / transparency standards should ideally differentiate between various types of targeted audience.
Policy recommendation: We believe in different levels of explainability, which vary depending on the type of audience being addressed.
It seems obvious to say that a majority of AI users will be non-experts of the technology. What will be paramount for these users, above understanding the technical logic that led an AI system to make a certain decision, will be to determine whether to trust this decision or not. On the other hand, experts, including auditors and authorities aiming to control the technology, will seek a level of understanding that goes beyond trust, aiming to ensure that the AI works in respect of all the key principles. In turn, the result of this independent review will be key in producing trust among users. For this reason, we recommend that a difference is maintained between explainability (or understandability) for all users, and transparency for experts and auditors.
This difference may have broad implications in terms of what information AI developers are expected to deliver to the public. Should the explainability requirements go too far, they can risk infringing upon trade secrets or intellectual property rights, hereby removing the economic incentive to develop AI in the first place. As a result, we believe policy-makers should exercise caution when adopting explainability requirements on AI.
At eBay we believe that AI should not be reserved to a privileged elite, nor should its decisions lead to manipulation or discrimination of a given group of users.
An illustration of our goal for inclusion at eBay through AI can be found in our recent investments for image-based search. We aim to achieve the same level of query understanding whatever the method that the user chooses to look for an item on eBay. This is likely to help populations with difficult access to textual search (for example due to language barriers, or disabilities) feel more included into the eBay shopping experience.
Further, as we operate in more than 190 markets, machine translation is used as a tool to process multiple languages when buyers shop internationally. A study conducted by Brynjolfsson, Hui and Liu (2018) that investigates the effect of eBay's AI-based machine translation tool on international trade has shown that the system helped achieve a 17.5% increase in exports. As a result, our machine translation system removes language barriers that can significantly hinder inclusive trade.
Policy recommendation: To further inclusion in AI, we believe that global cooperation, standardization and governance are needed as AI technologies themselves are seldom developed in regional siloes. This means that any voluntary labelling system for AI applications should ideally receive not only European but global recognition.
Furthermore, we would caution policy-makers against obligations on the “representative quality of data sets” used in AI. We believe the privacy of users could be at risk, as AI developers are not meant to access their personal attributes such as gender or ethnicity. Besides, the notion of representativity itself should in this case be precisely defined in order to avoid all risk of subjectivity.
Protective of privacy and security
For almost 25 years, we have been committed to protecting the privacy and data of our hundreds of millions of global customers who use eBay’s trusted platforms. Millions of transactions happen every day on our site, and we value the trust users place in eBay safeguarding personal information. In 2009 eBay implemented as first Silicon Valley company Binding Corporate Rules, and since its adoption in 2016, we embraced the GDPR as an opportunity to demonstrate and deepen our commitment to protecting our users’ data, making enhancements to processes, products, contracts, and documentation to help support the company’s, and our partner’s, compliance with the GDPR.
AI at eBay does not in any way depart from this approach and is therefore developed in full consideration of GDPR rules. Privacy is inserted by design in all of our technology and that includes AI. For example, we make sure to anonymize data collected to train our AI, such as pictures of individuals used for image recognition in the fashion area (image-based search technology called “Shop The Look”). This is consistent with our founding privacy principles that our users’ data is collected, stored and used appropriately and in such a way that users understand what information eBay collects, and the value they receive in return. We provide clarity on the way eBay collects, uses, shares and manages personal data through our User Privacy Notice and that covers the data we use for our AI technology.
Policy recommendation: In order to ensure trust, security must be integrated by design in AI technologies. However, as previously mentioned, this might be contradicted by certain explainability and transparency requirements imposed by legislation. Maximal compatibility between both the explainable and the security principles must therefore be ensured, and this includes to factor in the additional risks on privacy, cybersecurity or on the manipulation of data prior to adopting such requirements. This challenge extends to the concept of human oversight which, if inadequately applied, could destroy the benefits of the AI application entirely (by undermining speed of execution and privacy).
At eBay, we want to foster “commerce with a purpose”, by enabling businesses to grow through insights and competitive analysis, hereby advancing retail economic prosperity, globally - across geographies and languages.
For example, we are offering eBay sellers around the world an option to “clean up” their images thanks to AI-reliant image recognition technology. Cleaned up images are more likely to attract customer attention and therefore sales. At the same time, the technology alleviates the need for sellers to invest in costly professional photography equipment.
We also use natural language processing to accelerate product listing time. This allows us to refine the search process and find the best matched item. It helps sellers to save time, as well as get better descriptions and details about products.
Similarly, we use AI to provide insights to our sellers harnessing the power of the collective eBay community. For instance, this allows us to recommend the best price and the best time for sellers to bring their inventory to the platform. Of course, we do so only by using aggregated and anonymized data, protecting each individual seller’s information and data to ensure fair competition.
Policy recommendation: In order to support the development of AI with a positive impact on society, its economic viability will have to be ensured. As such, and as mentioned above, explainability requirements will have to ensure the protection of key innovation protections such as IP rights and trade secrets.
Furthermore, we believe that, like for any product, the intended use and context of use should be factored in when looking at assigning liability for harm caused by AI. In fact, it is possible for ill intentioned users to create harm from the use of AI solely by diverting it from its intended use. For example, on eBay, fraudsters may attempt to “game” search algorithms to get their products to show more prominently, for instance by miscategorizing or otherwise inaccurately labelling their listing. In such a case, the algorithm itself is not at fault, but rather the use that was made of it. While AI developers will always, to ensure the prosperity and usability of their technology, attempt to reduce such opportunities for fraud (for example, by rolling out updates or by training the system to detect certain use patterns), assigning full liability may discourage the use of the technology entirely.