Internet Society Youth Standing Group
Veronica Piccolo, ISOC Youth Standing Group, Technical Community, WEOG
Emilia Zalewska, NASK, government, East Europe
Daniele Turra, ISOC Youth Ambassador, technical community, WEOG
Veronica Piccolo, Ca’ Foscari University, Academia, WEOG
Emilia Zalewska, NASK, Governmental organisation, East Europe
Umut Pajaro Velasquez, ISOC Gender Standing Group, Civil Society, GRULAC
Daniele Turra, ISOC IGF Youth Ambassador, technical community, WEOG
Targets: This proposal links to SDGs 8.2 and 8.3 by discussing the modalities to achieve higher levels of economic productivity through technological upgrading and innovation and the promotion of development-oriented policies that support productive activities, entrepreneurship and innovation; SDGs 9.1 by exploring how the use of technical infrastructure and equipment to set up cloud computing services and AI technologies can support economic development and whether there are setbacks; SDGs 17.10 and 17.16 by emphasising the importance of multistakeholderism as an AI governance model and multilateral cooperation to promote an open, rules-based and non-discriminatory trading system.
The talk will consist of a presentation followed by a discussion with the audience. The topic is based on a three-year Doctoral research meant to be divulged through several papers and presentations. This talk will serve to open a discussion in a non-academic environment in order to further the research in this area thanks to the input of all stakeholders.
The high availability of data, coupled with increasing computing power (from cloud to quantum), is paving the way for the rapid improvement and expansion of AI. As decision-makers struggle to understand the full dimensions and impact of this technology on human rights, it is still uncharted its impact on consumers. In fact, with the ramping and uncontrolled use of AI in data-driven markets, competition might be disrupted at the expense of end users, consumers and more fragile stakeholders. The presentation will show the results of a Doctoral research based on substantive data and legal analysis on competition policy and enforcement in the era of big data and artificial intelligence, which will highlight the role of multistakeholderism in AI governance even in this policy area.
The session will ensure the hybrid format following the next proposed agenda: - 10’: introduction and presentation - 25’: open floor for questions, comments, inputs and feedback from the audience both online and onsite. - 5’ for conclusions and key takeaways. The full participation of everyone in the session is guaranteed regardless of the location, both moderators will be granting the floor on a round-robin base.
The greater data availability and the mainstreaming of AI systems in digital markets might lead to new phenomena of algorithmic collusions.
The legal framework is uncertain and competition law enforcement unfit for the digital age.
The solution can be found in the ex ante approach by shaping AI governance to include competition law compliance programs.
# KEY POINTS
* PhD research on "Competition Law and Enforcement in the era Big Data and AI", with three main points (technical, legal, enforcement)
* Technical findings include the potential for AI-based systems to facilitate cartels under specific conditions, such as using machine learning/deep learning, unsupervised learning, and transparent oligopoly/duopolies in the market structure.
* Legal findings indicate that AI-enabled collusions may not always be considered cartels and sanctioned, except for some cases like the Messenger case where humans used AI to implement a cartel.
* Enforcement efforts by competition authorities involve various approaches, including AI-based systems for parameter analysis, monitoring and analytics, web scraping, and econometric instruments. Proving collusion can be challenging in some cases.
# FULL REPORT
Presents Veronica Piccolo and the format of the session: 15 mins for presentation, 15 mins for Q&A
Veronica Piccolo is a lawyer, originally from Italy, part of Youth Standing Group (YSG) of Internet Society. She currently works for the European Commission. All the opinions expressed here are not on their behalf. She just completed a PhD research in Law and Economics at Ca' Foscari University in Venice with the title of
"Competition Law and Enforcement in the era Big Data and AI". Using a comparative, multidisciplinary approach, she focuses on the legal and technological aspects.
Her PhD research focused on three main points:
3. Institutional and Enforcement
## Technical findings
The question here was: "Are AI based system able to facilitate cartels?"
Those findings are based on an experiment developed by Calvano et al. at the University of Bologna. The experiment involved getting simulated markets to interact with each other using Q-learning. The key objective here was to understand if two AI algorithms could possibly collide and set the same prices?
The findings suggest that when some specific conditions that are met, based on the same design, they could learn to interact to each other. The elements are the following:
* Design: use of Machine learning/Deep learning; most likely, with the same technology being deployed on the market, there are good opportunities that those interact with each other.
* Data: unsupervised learning is more likely to facilitate collusion
* Market Structure: transparent oligopoly/duopolies, with prices known and interchangeable products.
Some criticism was moved to the approach, such as the experiment not working under normal market dynamics. This was contrasted by Brown and Makay and found that over two years, the sellers that could sell using the algorithms set the price above marginal cost.
Also, Ezrachi and Stucke devised models of collusion:
Messenger: cartel set up by humans and algorithms implement it
Predictable Agent: AI used to monitor the market and swiftly react to competitors price change.
Hub & Spoke: presence of coordinators and coordinated nodes, such as a platform marketplace and the spokes. So, the prices of the spokes are coordinated by the hub.
Digital Eye: the "cartel of the future, where market players would be able to predict the price change by competitors using AI and adjust accordingly
## Legal FIndings
The question here was: "Can AI-enabled collusions be considered as cartels and sanctioned?"
The findings suggest no. Besides the hardly exposable Digital Eye, there are some real cases that can be explored. For the Messenger case, the UK competition authority (CMA, case Trod/GB Eye) determined that a cartel was determined by humans that used AI to implement it. In that case, CMA sanctioned Trod, but not GB, because the former applied for the, specific procedure for leniency. For the Predictable Agent case, we have examples of AI system used to monitor and swiftly change prices. Calvano et al. exposed dynamics of Q-learning, where the algorithms learn to collude and keep the price above marginal costs, more in particular between Bertrand-Nash price and monopoly price. Brown and Mackay bring the case of over-the-counter drugs, where algorithms are used to generate supracompetitive prices through non-collusive mechanism.
Sometimes, it can be very hard to expose this type of cartels. The legal interpretation of what a cartel can be does not correspond to what an economist would identify. Cartel investigations from an economical perspective look for very clear evidences, but for legal theory this can not always be done.
The question here was: "How are competition authorities tackling the issue?"
The European authorities have equipped themselves to tackle the complexity. The Italian AGCM is using AI-based systems for reverse-engineer parameters generated by the AI of the undertakings. In Poland, they are trying to find if there are terms not compliant with consumer right law. In Greece and Spain, authorities are performing augmented market monitoring and analytics to expose when the market price fluctuates too much, therefore signaling the need for an investigation. In the Czech Republic, the authority is doing web scraping and econometric instrument to detect manipulation.
In some cases, competition authorities might not be certainly able to prove collusion, also given that companies can provide information to prove the contrary.
# OPEN FLOOR
How can AI governance be shaped to include competition law compliance programs?
Q: As an umbrella association, we are working with other organization to understand how anti-trust law can be applied and communicated to the rest of civil society. Is it possible to adopt a multistakeholder to frame the governance of AI?
A: Compliance programs must be based on multistakeholder model, but still there is nobody representing smaller businesses, just marking another cost for them. Lowest cost for being compliant.
### 2 Francisco Livardia - Diplomat from Panama
Q: How would be the relationship of AI Governance and compensation of damages? How can the fair market price applied for compensating damages?
A: In Europe we have a poor framework for private enforcing of anti-trust law. In the USA and other countries, if you are a damaged competitor, you can go before a judge and ask for reimbursement. In the EU is up to public competition authorities to carry out investigations and fine accordingly. However, if a cartel is suspected and one of the cartel members applies for a lineancy, they can disclose the cartel themselves and do not get fines by exposing other cartel members.
Q: Can we talk about the new EU policies about data sharing? They seem to enforcing data sharing, actively damaging competition.
A: The Data Act is still just a proposal. However, interestingly, we have the Digital Market Act (DMA). Article 6 says that the gatekeeper has to make the data available to other businesses, including possible competitors. The open question here is, does this make easier to facilitate market collusions?
Q: Competition Law is just for big companies. Can we model an AI program as a information theory system and train it to report on how other models are treating data?
A: Compliance programs could benefit that, however this would require the datasets to be f
rozen in order to expand the state and run more and more iterations.