When AI Becomes Routine: A Decade of Public AI Mediation in Korean Go Commentary

The Vanishing Label: How AI Becomes Invisible Through Routine

As AI systems surpass elite human performance and settle into everyday expert practice, a subtle transformation occurs in how we talk about them. In the high-stakes world of professional Go, AI is no longer a shocking interloper. It is now a standard analytic tool. However, a new study from Indiana University Bloomington suggests that as AI becomes more integrated, it actually becomes less "present" in human conversation.

The study explores how professional Go commentators on YouTube discuss machine judgment over a decade of availability. It finds that as AI moves from a "shock" novelty to routine infrastructure, the way experts talk about it shifts fundamentally. Instead of explicitly naming the AI, commentators increasingly rely on describing the metrics and graphs on the screen. This process, which the authors call "domestication" (the process of integrating new technology into existing social practices), allows the machine to recede into the unmarked background of expert speech.

The Gap Between Sight and Speech

Current debates regarding AI often focus on short-term reactions. We study how people feel immediately after a breakthrough or how they react to a specific error. We understand the "shock" phase. However, we know much less about what happens after a community has lived with superhuman systems for years. There is a growing disconnect between what is visually presented to an audience and what is verbally articulated by experts.

In modern Go broadcasts, the AI's influence is visually omnipresent. The authors report that in late-period institutional broadcasts, AI winrate graphs are visible for an average of 97.99% of the time .

This means the machine is almost always visible on screen. Yet, despite this near-constant visual presence, "AI-salient" talk—speech that explicitly engages with the machine—accounts for only 2.63% of all sentences in the BadukTV sample. This creates a massive asymmetry. The machine is always seen, but it is rarely named.

Mapping the Mechanics of Mediation

To understand this shift, the researchers developed a framework for "public AI mediation." This is the communicative layer where human intermediaries (people who interpret data for others) translate, soften, or resist machine judgment for an audience. They categorize this mediation into a typology that distinguishes between two primary directions: source-foregrounding and source-receding.

Source-foregrounding acts as a "discursive anchor." Much like a legal disclaimer in a financial advertisement, it explicitly identifies the source (e.g., "According to KataGo..."). This ensures the audience recognizes the origin of the judgment. This allows the audience to maintain a "hook of contestability." This is a linguistic starting point that allows people to recognize, question, or challenge the algorithmic source.

In contrast, source-receding mediation involves "interface rendering." This occurs when a commentator describes the data without attributing it to the machine. For example, they might say "the winrate is dropping" instead of "the AI says the winrate is dropping." The authors argue that this leads to "naturalization." In this state, machine judgment is folded into the natural flow of expert commentary. It becomes the unmarked, default way of evaluating the game.

Evidence of a Decadal Shift

The study analyzes a massive corpus of approximately 1,900 hours of footage from 2016 to 2025. It spans four distinct phases: Human Mastery, Shock, Diffusion, and Routine Integration. By tracking specific keyword families, the authors document a clear evolutionary trajectory.

The paper reports that the total share of AI-salient sentences in BadukTV grew steadily. It rose from an effectively zero baseline in the pre-AlphaGo era to 2.63% in the routine integration phase .

This represents a steady increase in how often AI is verbally activated. However, the most significant finding is the change in the composition of that speech. In the early "Shock" phase, the majority of AI-related talk focused on explicit naming. By the "Routine" phase, the balance had shifted toward interface and metric-based talk.

This shift is even more pronounced when comparing different media formats. The authors find that creator-led channels (individual streamers) lean much more heavily into interface rendering than institutional broadcasters .

While institutional channels retain more explicit naming, creator channels often bypass the source label entirely. They focus almost exclusively on winrates and "bluespots" (the AI's top recommended moves).

Limits of the Go Model

While these findings offer a compelling look at AI integration, the authors note several limitations. The study is strictly confined to the Korean Go community and the YouTube platform. Go is a "complete-information" game. This means all players can see the entire state of the board at all times. Because of this, the AI's evaluations are exceptionally reliable. In Go, the machine acts as a "near-oracle." This is a source so trusted that its judgment can be assumed without constant attribution.

Furthermore, the researchers cannot definitively determine the cause of the verbal shift. It is unclear if this is a deep socio-cognitive change in how experts think. Alternatively, it may be a practical adaptation to broadcast UI (user interface) conventions. If a winrate graph is already on the screen, a commentator might stop naming the AI to avoid being redundant. Finally, because the primary coding was performed by a single researcher, the study acknowledges potential limits in interpreting complex linguistic nuances.

The Verdict: A Warning for High-Stakes AI

Does this mean the "domestication" of AI is a purely positive sign of maturity? The answer depends entirely on the stakes.

In the context of Go, naturalization is a functional success. The machine has become a seamless part of the expert's toolkit. However, the authors warn that this pattern poses significant governance risks in other domains. This is especially true when moving from low-stakes games to high-stakes environments.

In a clinical setting, if a doctor begins to say "the risk score is high" instead of "the diagnostic algorithm estimates a high risk," the "hook of contestability" vanishes. When the source recedes, the ability for a human to recognize or challenge the machine's judgment is eroded. For practitioners designing AI interfaces, the takeaway is clear. Transparency is not just about what the model outputs. It is about how the human intermediary is empowered to speak about it. Code for this study is reportedly available; see the paper for the canonical link.

Figures from the paper