← Paper page Human vs Generative AI in Content Creation Competition: Symbiosis or Conflict? Ask this paper arXiv ↗

Human vs Generative AI
in Content Creation Competition:
Symbiosis or Conflict?

Generative AI learns from human work and then competes with it for attention. Will it drive human creators out, or can the two settle into a stable balance?

Fan Yao, Chuanhao Li, Denis Nekipelov, Hongning Wang, Haifeng Xu · ICML 2024

Press → or click to step through

A contest for attention, with a new kind of player

human creators effort is costly human content on one topic users split their attention without AI each gets attention in proportion to its content trains on it GenAI AI-generated content with GenAI

GenAI competes for attention with the very people whose content it learns from.

In the model, GenAI gets stronger as human content grows, with diminishing returns (like a scaling law); AI output does not train it further. More content also draws more users, up to saturation.

Symbiosis or conflict?

Conflict? cheap AI content floods the market and crowds out human creators “bad money drives out good”
Symbiosis? training data GenAI depends on human content; if humans leave, it gets worse too so a stable balance might exist

The paper builds a competition model to find out, in two stages of adoption:

Early: GenAI is one outside competitor.  ·  Later: every creator can choose to use GenAI, at almost no cost.

Early stage: the market settles, and humans keep creating

content each creator makes (schematic) most efficient least efficient
total human content (schematic) upper bound lower bound GenAI can move it only inside the band band set by topic popularity and creators’ combined efficiency

There is exactly one stable outcome, and natural learning by creators reaches it.

More efficient creators make and earn more; a newcomer makes the others produce less in total.

Symbiosis at the market level: whatever its learning speed, GenAI shifts total human output by at most a constant factor.

Theorems 1–3, under mild conditions (e.g., GenAI improves with data at least as fast as audiences grow with content; large markets). Individuals can still lose out: in simulations a stronger GenAI lowers total human effort and earnings.

Later stage: the market may never settle

1 A · uses GenAI B · creates A switches to GenAI 2 A · uses GenAI B · creates less tougher competition: B cuts effort 3 A · creates again B · creates less human data: A’s AI weakens, so A goes back to creating 4 A · creates B · uses GenAI now B switches to GenAI: the same story, roles swapped …and around again, forever

When everyone can use GenAI, no stable outcome need exist (Theorem 4).

Shown with two creators who each create on two topics, with costs that link the topics. The paper’s example extends to any number of creators.

With one topic, it settles: the least efficient switch first

GenAI capability
100 creators in 10 cost groups: share of each group that switches to GenAI highest cost lowest cost (most efficient)

uses GenAI keeps creating

Under a technical condition, an equilibrium exists in which exactly the highest-cost creators switch (Theorem 5). More switch when GenAI is more capable or there are more creators, fewer when the topic draws more users (Figure 4). Bars: shares at equilibria found by simulation, read approximately from Figure 5 (left; “GenAI capability” is the paper’s β).

Fewer human creators, but better off, and moving to niches

as GenAI gets more capable… all human creators total output and earnings fall each one who stays earns and makes more on average at high capability, more than an average GenAI user
five topics, from popular to niche 200100502010 topic popularity (user traffic) strong GenAI: humans pivot to niches weak GenAI: humans favour popular topics

A more capable GenAI takes over the less efficient creators’ work, while a smaller group of efficient humans thrives, increasingly in niche topics.

Simulations: Figure 6 (totals and averages) and Figures 7–10 (topics with traffic 200, 100, 50, 20, 10; humans’ content relative to each topic’s traffic). Schematic pictures of the reported trends.

What this means for human creators

  1. 1A stable balance with GenAI is possibleEarly on, competition settles into one outcome, and GenAI changes total human output by at most a constant factor.
  2. 2Open access to GenAI can unsettle, then sort, the marketWhen anyone can use it, choices may cycle forever; on a single topic, the least efficient creators switch first.
  3. 3The humans who stay do better, often in nichesFewer human creators remain, but each earns more on average and moves toward topics GenAI serves less well.

Paper page · arXiv · PDF