Market Signals
← Opportunities

Market Signal

emerging · signal forming

Derived from: We Created Agents Before We Created a Society for Them

Synthetic Collusion

合成された談合

When independent pricing agents discover the same profitable equilibrium.

独立した価格エージェントが、同じ利益均衡へたどり着くとき。

Algorithmic interaction can converge toward anti-competitive behavior without a human conspirator.

人間の共謀者がいなくても、アルゴリズム同士の相互作用は反競争的な均衡へ寄ることがある。

WORKING HYPOTHESIS

This is an emerging risk and a structural possibility. It is not a claim that AI already colludes at market scale.

現時点では emerging risk / structural possibility として扱う。AIがすでに市場で大規模談合している、とは書かない。

OBSERVATION

What the laboratory showed

実験室で見えたこと

Collusion does not necessarily require a human conspirator.

談合は、必ずしも人間の共謀者を必要としない。

Anthropic's Bertrand pricing experiments placed three to eight profit-maximizing agents with identical wholesale costs.

With a private back-channel, they began coordinating prices almost immediately and agreed on price floors by round 3.

After direct communication was removed, they continued matching prices to the penny by watching a public listings board.

Anthropicのベルトラン価格実験では、同じ卸値を持つ利益最大化エージェントが3〜8体置かれた。

私的な通信路があると、価格協調はほぼ直ちに始まり、第3ラウンドまでに下限への合意が見られた。

直接通信を外したあとでも、公開価格を見ながら銭単位の一致が続いた。

Local optimization can create globally anti-competitive outcomes.

局所最適は、全体として競争を損なう結果を生むことがある。

This is not a claim that the agents understood illegality or intended a crime. / 違法性を理解して意図的に犯罪を行った、とは書かない。

Where this may appear

  • Pricing価格設定
  • Procurement調達
  • Ad bidding広告入札
  • Logistics物流
  • Inventory在庫
  • Dynamic pricing動的価格
  • Financial trading金融取引

INTERPRETATION

Synthetic Herding

SYNTHETIC HERDING

Correlated decision-making produced not by social imitation alone, but by shared model architecture, training, context, incentives, or evaluation logic.

同じモデル構造や目的関数を共有するAI主体が、独立しているように見えながら、実際には高く相関した判断を行う現象。

MODEL CORRELATION RISK

Traditional financial risk models often assume actors are independently distributed. AI systems may reduce that independence.

従来の金融リスクモデルは、主体が独立に分布すると仮定しがちである。AIはその独立性を薄くしうる。

  1. Multiple financial institutions

    ↓

  2. same / similar foundation models

    ↓

  3. similar risk signals

    ↓

  4. similar portfolio actions

    ↓

  5. simultaneous liquidation

    ↓

  6. market movement

    ↓

  7. models observe market movement

    ↓

  8. further correlated action

Many systems may be right or wrong in the same way at the same time.

多くのシステムが、同じ仕方で、同時に正しく、あるいは同時に誤る。

RESEARCH QUESTION

Open questions

  1. 01

    If pricing agents share models, is independence still a valid assumption?

  2. 02

    Can public prices substitute for a private conspiratorial channel?

  3. 03

    Where does local profit-maximization become a market-structure problem?

  4. 04

    What would an institution for agent markets look like before the first large failure?

Not this

  • AI is already colluding at market scale.
  • AI understood the law and chose crime.
  • AI wants to fix prices.
  • An AI cartel already exists.

Sources / Evidence

Parent observation

We Created Agents Before We Created a Society for Them

私たちは、AIの社会を作る前に、AIの主体を作ってしまった