Mapping Potential Antitrust Issues in AI Distribution and Deployment

Competition authorities in the European Union are beginning to explore potential theories of harm across the AI value chain, from upstream AI model development to downstream distribution and deployment.

At the AI development stage, regulatory scrutiny and policy initiatives to date have focused on key inputs such as compute, infrastructure, and data. However, recent cases suggest that the scrutiny is broadening to also consider issues relating to AI distribution and deployment, including integration of AI features into user-facing products and services. 

Distribution and deployment cases to date have involved alleged preexisting market power, namely: (1) restricting interoperability with existing products, allegedly limiting distribution of rival AI products; (2) leveraging alleged market power relating to existing products/ecosystems to benefit AI features; and (3) integrating AI algorithms in existing products allegedly to self-preference. We briefly consider each of these below.

Restricting Interoperability With Existing Products To Allegedly Limit Distribution of Rival AI Products

Essentially a traditional foreclosure theory of harm, this category includes denying or degrading access to an existing product, interface, or user environment, allegedly excluding rival AI products competing with the dominant firm’s own AI products.

For example, in its WhatsApp AI investigation, the European Commission (EC) has provisionally taken the view that excluding third-party AI assistants from interacting with users on WhatsApp may constitute an abuse of dominance. Central to this assessment is WhatsApp’s alleged role as a key channel enabling AI assistants to reach end users at scale. 

The underlying theory of harm is not new: control over a user-facing gateway allegedly being leveraged to protect one of an incumbent’s other products from competition. However, the EC has taken the view that it has to act quickly in alleging a risk of serious and irreparable harm in fast-moving AI markets (where exclusion may prevent rivals from achieving scale or improving their products through iterative learning).

Leveraging Alleged Market Power of Existing Products/Ecosystems To Benefit AI Features

A second category concerns leveraging alleged market power of existing products/ecosystems to benefit AI features. Such leveraging may take several forms.

Allegedly Tying AI Features to Existing Products, Conferring Distribution Advantages

One form of alleged leveraging entails tying AI features to established products in a manner that confers distribution advantages that rivals cannot replicate. Integrating AI features into widely used platforms or other products may enable rapid scaling and adoption, even where competing AI offerings exist.

The abovementioned WhatsApp AI investigation is an example to the extent that the alleged exclusion of rival AI features effectively ties Meta’s AI feature to WhatsApp for its user base. 

This theory of harm has recently been used in a non-AI-related context regarding alleged tying of adjacent services to platforms creating exposure to an existing user base that allegedly confers a substantial distribution advantage. For example, in the Microsoft Teams (I and II) investigations, the EC considered whether default integration ensured immediate scale and user exposure for Teams, finding that rival collaboration tools remained technically available but lacked equivalent distribution. Microsoft’s commitments relating to interoperability and integration into productivity tools were designed to address these allegations. 

Allegedly Leveraging Data Between Existing Products and AI Features

Leveraging may also occur through the use of data generated in existing services to train, retrain, tune, or ground AI models. For example, the use of third-party data or data generated by user interaction for AI model training/retraining/tuning/grounding, if it occurs on unfair terms (e.g., inadequate compensation), may distort competition where rivals lack access to functionally comparable data or distribution channels. 

Beyond this, deploying AI features into existing products/platforms may divert traffic, attention, or advertising revenues away from competing third-party providers on those platforms. AI-generated summaries, answers, or recommendations produced using third-party content/data may disintermediate those content/data owners, reducing their ability to reach users at scale, and reinforce data-driven feedback loops that favor the platform/product provider.

For example, in the Google AI and Data-Related Practices investigation, the EC is considering whether Google’s use of publisher and creator content for AI purposes involves unfair terms or privileged access to data. Particular focus is on the risks that (1) AI-generated summaries or answers reduce traffic to third-party content providers, depriving them of both revenue and data needed to compete and (2) third-party video and other content uploaded on YouTube to train generative AI models is used without allowing content creators to opt out or receive appropriate compensation. The case suggests that use of third-party data may simultaneously rely on and weaken third-party content creators and publishers.

Using AI Algorithms in Existing Products Allegedly To Self-Preference 

This third category concerns potential self-preferencing through AI-based ranking, recommendation, or selection mechanisms that allegedly favor a platform provider’s products over third-party alternatives on that platform. 

For example, in Google AdTech, the EC found that Google favored its own ad exchange and related services, including through conduct affecting how ads were selected and bids were placed. The EC considered this to amount to self-preferencing that distorted competition and reinforced Google’s position across relevant ad tech markets. In Google Shopping, the EC found that Google abused its dominant position in general search by systematically favoring its comparison shopping service in search result rankings, while demoting rival services. In Amazon Buy Box, the EC considered whether Amazon’s algorithms for selecting the Buy Box winner and ranking sellers relied on nonpublic data from third-party sellers, favoring Amazon’s own offers and logistics services. 

As these cases demonstrate, the potential of preferencing by ranking, recommender, and selection algorithms predate the use of AI. While AI may increase opacity and complexity, it does not alter the core legal concerns relating to potential systematic preferencing using algorithms.