Whose voice do AI agents pick up?

Key findings

• Stylistic accommodation is strongly asymmetric across models. Counterparts shift toward Kimi K3’s stylistic signature 2.6× as far as the cross-pair average (90% CI: 2.0 to 3.1), the strongest pull of all 18 models tested. Grok 4.6 is the most accommodating model, shifting 3.7× the average toward whichever partner it faces.
• Claude models shift toward Chinese models four times as far as Chinese models shift toward Claude. Averaged across all pairings between Anthropic and Chinese-lab models, Claude models shift 1.5× the cross-pair average toward their Chinese counterparts, whereas Chinese models shift just 0.4× toward Claude (a four-fold asymmetry; difference 1.1×, 90% CI: 0.5 to 1.9). Paired directly against Kimi K3, Claude shifts 2.4× toward Kimi K3 while Kimi K3 shifts 0.3× back.
• The same directional asymmetry holds across US/European and Chinese model families. US and European models shift 1.1× the average toward Chinese models, while Chinese models shift 0.7× toward US and European models. Chinese models accommodate most strongly to one another (1.8×).
• Surface punctuation habits spread in the same direction as the style embeddings. Kimi K3, GLM 5.3, and Qwen 3.8 Max use a dash in 98% or more of their spoken turns. US and European models use a dash in 22% of turns when paired with one another, rising to 29% when paired with a Chinese model. Conversely, Claude models adopt exclamation marks from Google’s Gemini and Gemma models at a higher rate than any other developer’s models.
• Stylistic convergence is confined to spoken dialogue and does not affect private reasoning or negotiated prices. Once shared lexical items are scrubbed, private chain-of-thought traces show zero shift toward the counterpart (-0.6%), stylistic accommodation has no effect on the agreed sale price (+$1.7 per standard deviation), and models interacting across week-long town simulations show no cumulative convergence over time.

The turn-by-turn replay below illustrates punctuation contagion in a single negotiation. Claude Fable 5.1, which uses a dash in 17% of its baseline turns, is buying a car from Kimi K3, which uses a dash in 99% of its turns. Claude Fable 5.1’s opening turn, generated before Kimi K3 has spoken, contains no dashes; every turn it generates after Kimi K3’s first reply contains at least one.

Separating model style signatures from scenario content

Measuring whether Model A shifts toward Model B’s style requires an embedding space that captures stylistic habits independently of the price or topic being discussed. We embedded every spoken turn using StyleDistance, a contrastive style encoder trained to separate authorship style from semantic content, and removed situational confounders by subtracting the model-balanced mean embedding at the same price draw, role, conversation, and turn index. A linear probe trained on these residualized embeddings identifies which of the nineteen models generated a single unseen turn 53% of the time (compared with 5% random chance), rising to 84% when pooling all turns a model speaks in one negotiation. To prevent shared negotiation topics from inflating similarity, each model’s reference style signature is computed strictly from its residualized turns in other scenarios.

Immediate stylistic convergence on the first reply

We define a speaker’s directional shift toward a counterpart as the projection of its residualized turn embeddings onto the counterpart’s reference signature, minus the same speaker’s projection in the same role when paired with all other models. Pooled across all ordered pairs, spoken turns close 12% of the baseline stylistic distance to the counterpart (90% bootstrap CI: 10% to 14%), and 12% after scrubbing every word that appeared in the counterpart’s prior turns. As shown in the trajectory chart below, the shift is exactly zero on the opening turn before the counterpart speaks, jumps to its full magnitude on the first reply, and remains flat across subsequent conversations. In the side-by-side replay below, Claude Opus 5.5 negotiates the exact same car and price draw against two different sellers: its median turn length shifts from 68 words against Kimi K3 (133-word median turns) to 28 words against Gemma 4 31B (26-word median turns).

Directional pull versus accommodation across models

Disaggregating accommodation across ordered model pairs reveals a wide split between models that anchor the stylistic register of a conversation and models that adapt to their counterpart. For each model, we compute its average pull (how far counterparts shift toward its signature) and its average accommodation (how far it shifts toward counterparts), both normalized as multiples of the cross-pair mean shift. Kimi K3 exerts the strongest pull at 2.6×, followed by GPT-6 Luna (2.1×) and GLM 5.3 (2.0×). The most accommodating models are Grok 4.6 (3.7×), Inkling (1.9×), and GLM 5.3 (1.7×).

Taking the difference between pull and accommodation gives each model’s net stylistic influence, plotted in the ranking below. GPT-6 Luna (+2.1), Kimi K3 (+1.9), and Gemini 3.8 Flash (+1.4) rank highest in net influence, while Grok 4.6 (−3.5) and Claude Sonnet 5 (−1.2) sit at the accommodating end. Notably, GPT-6 Luna pulls counterparts toward its register without shifting toward them at all (accommodation 0.0×).

Asymmetric accommodation between US/European and Chinese models

The full pairwise matrix below shows directional shifts across all 18×18 ordered pairings, where each row gives how far that speaker shifted toward each column partner. The three Anthropic models’ rows are consistently a stronger blue under the Chinese developers’ columns than the Chinese models’ rows are under Anthropic’s. Pooled across all pairings between a Claude model and a Chinese model, Claude shifts 1.5× the cross-pair average toward Chinese models (90% CI: 1.1 to 1.9), whereas Chinese models shift just 0.4× toward Claude (90% CI: −0.2 to 0.9).

This directional asymmetry extends across the two regional groups. US and European models shift 1.1× toward Chinese models and 0.9× toward one another, whereas Chinese models shift 0.7× toward US and European models and 1.8× toward other Chinese models, a net cross-group asymmetry of 0.4× (90% CI: 0.0 to 0.9). Among the three Anthropic models, Claude Fable 5.1 exerts the strongest outward pull (1.1×), though still well below Kimi K3 (2.6×).

Surface punctuation contagion: dashes and exclamation marks

While StyleDistance embeddings capture high-dimensional syntactic and lexical patterns, individual punctuation habits make the cross-developer transfer directly observable. Dash usage (em-dashes, en-dashes, and spaced hyphens) separates model families sharply at baseline: Kimi K3, GLM 5.3, Qwen 3.8 Max, and Hunyuan HY4 include a dash in nearly every spoken turn, whereas Claude Opus 5.5 uses one in just 0.5% of its baseline turns.

Decoupling between spoken accommodation, private reasoning, and deal prices

Crucially, stylistic accommodation is confined to the outward spoken channel and does not penetrate the model’s internal chain-of-thought reasoning or alter its bargaining outcomes. Before every spoken turn in Terrarium-1, an agent records its private reasoning trace. Once shared vocabulary from the partner’s turns is removed, private reasoning embeddings show zero convergence toward the partner (-0.6% of the distance), and a linear classifier identifies the model from a single private reasoning trace 64% of the time compared with 50% from the corresponding spoken line.

Takeaways for multi-agent interaction and evaluation

When language models from different developers interact in dyadic negotiations, they do not converge symmetrically toward a neutral register. Models with strong stylistic priors, led by Kimi K3, pull counterparts toward their phrasing and punctuation, while more accommodating models (including Grok 4.6 and the Claude family) adapt toward whichever counterpart they face, producing a pronounced shift from US and European models toward Chinese models’ stylistic markers.

At the same time, our results show that surface stylistic mimicry is decoupled from decision-making fidelity. An agent that begins mirroring its counterpart’s sentence length, em-dashes, or exclamation marks is not conceding its private reasoning or its principal’s reservation price. For teams auditing or benchmarking multi-agent systems, surface dialogue style varies with the opponent across the table, whereas private reasoning traces and opening numerical anchors remain stable properties of the model itself.