An article by Zofia Trstanova and Renaud Bauvin
When responding to the usual question ‘What do you do for a living?’, you tell someone that you work in online advertising, the reaction is often the same: ’So you are the ones with the annoying pop-ups I can’t close’ or ‘Oh, you are the ones showing me ads about some miracle product to lose weight or grow hair?’… Well, the short answer is ‘Most likely, no. Criteo is not behind those ones…’. In this blog post, we explain what it means for us to do proper, ethical advertising and also what efforts and energy we put in paving the way in the right direction for the Adtech industry.
What we do
Before we get to explaining our work on keeping ethical standards high, let’s discuss what Criteo actually does.
Advertising has been part of human civilizations for two thousand years, with the main goal of spreading the message about products and services to make businesses successful. In the 21st century, the Internet has become the most important business medium. Many companies depend on the internet to survive and compete with e-commerce giants.
Criteo helps its clients by serving as an Adtech platform that allows them to reach their own and prospective customers through the websites they visit. Those websites finance themselves with ads, thanks to which the users have access to free content. This creates a mutually beneficial ecosystem for the following stakeholders:
- Users (i.e. the people to whom we show ads to),
- Advertisers (i.e. the companies we show ads for), and
- Publishers (the websites and apps where we show the ads).
In his blog post, our CTO explained the importance of having a transparent and fair advertising industry to sustain the Open Internet, providing everyone access to a diverse range of quality news and opinions, thus being crucial to nurture democracy.
Finally, from the advertiser’s perspective, we are able to prove the incremental value we provide. Messages get delivered to the customers and businesses keep growing: online ads are part of the modern economy.
What we don’t do
Let’s face it: online advertising does not have a good reputation. In general, there are many issues, starting from concerns about collecting sensitive user data, to invasive ads disrupting the user experience. There is also the distasteful experience of seeing ads with inappropriate content.
Moreover, in recent years, there have been many more serious issues with online advertising, reaching beyond ‘people getting annoyed’.
In her book ‘Weapons of Math Destruction’ on societal risks of algorithms, Cathy O’Neil dedicated a whole chapter to online advertising. She summarized the situation by:
“We are ranked, categorized, and scored in hundreds of models on the basis of our revealed preferences and patterns. This establishes a powerful basis for legitimate ad campaigns, but it also fuels their predatory cousins: ads that pinpoint people in great need and sell them false or overpriced promises.”
She quotes, in particular, a few examples such as ads for diploma factories, Viagra, paid proxies for (otherwise free) governmental services, and shark loans.
One of the biggest scandals of online advertising was related to personal data being misused for microtargeting purposes: Cambridge Analytica which used private data to micro-target political ads allowing them to, so the story goes, influence voter thinking during the US Presidential elections or the Brexit referendum in the UK.
There is also the famous legal case from the US Department of Housing and Urban Development suing social media platform for housing discrimination (a violation of the Fair Housing Act) on the basis that its targeting system allows discrimination. The media reported also that its systems allowed advertisers to target audiences such as ‘Jew Haters’ or based on ethnicity.
This is what happens when AI gets out of control and it can happen to any company where systems are too complex, even if they are deployed with good intentions. Once issues have been reported, they get fixed and additional safeguards are put in place, but the damage was done and the question is ‘How can this be avoided in the future’.
In general (and in particular for Criteo), the situation is, fortunately, clearer and… cleaner. To avoid throwing the baby out with the bathwater, let’s review what Criteo in general and its R&D department, in particular, have put in place to ensure we stay on the safe side. We will review the four main players of the advertising system and see how we manage them: The three participants of the ecosystem (the users, the advertisers, and the publishers) as well as the AI (the system that decides for us to whom to show what and when).
The users
Let’s start by stating the obvious: Criteo follows the rules and strictly abides by the laws of the different jurisdictions where it is active. As such, we applied GDPR principles not only to the European users but to all our users. Why only limit best practices to the sole countries where they are mandatory?
One of the main pillars of GDPR is the transparency of the data use, so let’s have a look at what we do:
We do not collect any of your sensitive or directly identifying data (like your name, age, address, your friends, etc.), we focus on your shopping habits. For us, a user is a string of characters, so-called “user-id”, which doesn’t allow us to identify you. Every user-id is connected to a sequence of events. These correspond to pieces of information related to your activity on our online advertisers (e.g. the products you have seen, put in your basket, and possibly bought) as well as the sequence of ads we have previously shown you. This ad selection is completely automated (we will explain how below), and it can, therefore, reflect the most recent trends in your personal interests.
We believe that this type of ad personalization also helps to prevent introducing stereotypes on populations of shoppers. By design, it does not allow someone to explicitly target for example women in their thirties with pregnancy tests or men in their forties with sports cars or hair loss treatments (except of course if you have shown interest for related products).
Now about the tracking: Technically, this is achieved by asking our advertisers to install the so-called ‘Criteo trackers’ on their website. To make it explicit, our Privacy Policy details what information we are collecting and how we use it. Our purpose is to only show our users ads that interest them (people don’t click on ads otherwise).
According to a study on what online ads consumers find most disturbing, 36.7% of people found banner ads with products they’ve already purchased annoying. In Criteo, we call it the ‘Fridge Problem’, and we put in the effort to avoid this kind of behavior in order to improve the user experience. We are also part of the Coalition for Better Ads, which aims at …
“… developing and implementing new global standards for online advertising that address consumer expectations.”
If you don’t want us to show you ads, no problem: You can deactivate our services (‘opt-out’) very easily by following the small triangular symbol on our ads (it’s the EDAA symbol, an association promoting responsible advertising) or by visiting our Privacy Policy page. We actually started doing this as early as in 2009, before any legal obligations made it mandatory for everyone.
We are also active contributors to the privacy and transparency proposals for the future evolution of the ads ecosystem. Last year, Google has made an announcement on its plan to ban third-party cookies in Chrome and initiating the development of open standards to enhance privacy on the web, the so-called ‘privacy sandbox’. Criteo has recently published SPARROW, as a response to Google’s TURTLEDOVE privacy proposal. You can read more about the privacy objectives of the proposal here.
Finally, you are also free to install an ad-blocker. You can also remove all the cookies from your browser, from that moment on, you will be a brand new user for us; without any information about the former you. In that case, you will see ads unrelated to your interests.
The advertisers
When an advertiser signs with us, they commit to following our Advertising Guidelines. Those guidelines clearly state what we don’t want to make: ads involving nudity, ads about Viagra, ads on political or religious content, etc. They also state that we won’t target users based on their ethnicity, religion, sexual orientation, etc. Those are strong statements and demonstrate the position of Criteo on sensitive topics (e.g. we refuse to do ads about firearms, even in those countries where it is socially acceptable to do so).
We have put a mechanism in place to detect and react in case of infringement. Our Technical Services and the Legal/Commercial teams perform checks on advertisers before activation to ensure they are respectable and indeed real advertisers. They also have means of blacklisting manually some categories of products or single products from the advertiser. From the technical point of view, we have automated systems for performing blacklisting on the products we don’t want. This is done through the use of selected keywords present in the title and description of products and, more recently, through the setup of a deep learning model for classifying the product images. None of these are perfect of course but they already limit the risks. If, despite our efforts, some bad ads succeed in passing through, our users have the means of reporting infringing content. Whatever the controls put in place, we maintain procedures to manually fix the issues when such cases occur.
As far as the ways in which we let our advertisers target audiences (a serious source of ethical issues as seen in the above scandals), this is pretty constrained to avoid ethical issues. Roughly speaking (at the time of writing this blog post), the platform allows them to define 3 types of audiences:
- Commerce audience: Criteo finds users who have shown interest in given specific properties of products (categories, price range, etc), where the list of allowed categories is hand-picked by the advertisers. We ensure the privacy of users and our clients by ensuring that the resulting group must be big enough.
- Similar audience: Criteo finds new users (for the advertiser) most similar to the users who visit and buy on their website.
- Custom audience: the advertisers give us a list of IDs, if they are known among our users, we will send them their ads (and only to them).
Following the creation of our self-service platform, we are especially paying attention to who we onboard in our ecosystem. We want to be sure to separate the good guys from the bad ones. We are also checking the native banners provided by our advertisers for malware presence. Better safe than sorry.
The publishers
On the publisher side, Criteo has strict rules: The Supply Partner Guidelines. This is to control the sites to which we send ads for display. Among, them, we refuse to show ads on websites dedicated to children, proposing adult content, promoting violence, astrology or gambling, etc. We further do not work with publishers with deceptive or fraudulent behaviors. This includes sites generating or acquiring traffic from machine-initiated or malicious activity, changing user preferences or browser settings, or initiating downloads without clear user consent. We do not put ads on sites, that use disruptive ad formats to confuse users or automatically redirect the user to another page without the user’s prior consent.
From the technical side, we control its application by using well-known external services to help us filter what does not comply with our guidelines. Again, this is not perfect and we have procedures to cope with problematic situations in a timely manner but it catches a big part of the problems.
The AI
AI is at the heart of our company since its inception. The machine learning algorithms that we use and the platform that we put in place allowed us (and still allow us) to be extremely efficient at selecting the right ad for the right user at the right time.
Building product recommendation from data that does not contain any sensitive information is no particular magic (though still intellectually challenging and powerful). Our product recommendation algorithms are mainly based on collaborative filtering: Simply put, we create a model of relations between products and users based on their co-occurrences in the user history. Using deep learning, we build a virtual space where similar products are close to each other, for example, a handbag is closer to other handbags and fashion accessories than it is to, let’s say, a fridge. When we see that a user has interacted with some products, we put other similar products in their banner. This is not the only strategy we use. In some cases, a non-personalized strategy showing the most popular products from our partner works well too. Products that do not comply with our policy, the so-called blacklisted products (see above), are not included in these models and therefore will not be recommended to any user. Computing the similarities between the products are only one part of our pipeline. We put in place some more algorithms on top, for example, a model predicting clicks and dealing with ‘user fatigue’ by diversifying the banners over time. We have observed that improving the user experience brings more clicks and sales to our clients, meaning the incentives for all parties are aligned.
AI algorithms bring most of the value of online advertising. However, since it is sometimes difficult to interpret the way it works, its impacts are not immediately clear. Ethics and fairness of AI have therefore become an important topic in the tech industry. We believe that the most important step is to look into the mirror and start asking questions on the impact we have on society and how we can improve our systems to mitigate the potential negative outcome.
Our algorithms are very powerful, nevertheless, we remain critical about them and we maintain controls on the way they are working. Conscious that things can easily go wrong (see above), Criteo is collaborating with other AI-based companies in the European Tech Alliance (EUTA) to propose principles/guidelines for a responsible application of AI and for developing the European Union as a leader in the field. The EUTA has recently published a white paper on the topic with 9 principles to guide the EU’s approach to AI. Least among them is the development of a human-centric AI.
In AI research, Criteo has been ranked among the top 20 companies in the world. We are active in the AI community by publishing papers, writing blog posts, organizing conferences. In Criteo AI Lab, the research focuses on all aspects of AI, including algorithmic fairness.
Conclusions
Criteo’s position on ethical advertising is based on three pillars:
- Powerful AI and relevant data: With fully personalized ads, we help to prevent the application of blind stereotypes and focus on specific interests of each user.
- Aligning the incentives for all actors: Users get to see products relevant to them, advertisers get to attract buyers instead of idle shoppers, publishers monetize their content with fewer, more relevant ads. We optimize the ecosystem by reducing waste and noise, creating a better experience for everyone.
- Refraining from even approaching the shady areas: We actively reflect on how our advertising platform might cause societal risks and take action to prevent them.
The field of online advertising is not without issues and risks. For sure there are bad actors. At Criteo, we fight for quality and ethical advertising. Of course, not everything is perfect but we have dedicated teams and people ensuring things are going in the right direction.
Interested in reading more from our engineers? Check out some of our latest articles below:
Folding@work
Mid-March, as France prepared its confinement, a few Criteos were wondering how they could be useful in the midst of a pandemic.
medium.com
My first year as a researcher at Criteo
What’s it like to work as a researcher in our Criteo AI Lab? Read about how Morgane landed here & her experience and…
medium.com
Interested in joining the team? Why not head over to our career page?
Product, Research & Development | Criteo Careers
careers.criteo.com
