Interaction Creates Dynamical AI Behavior Absent in Isolation

Interaction Creates Dynamical AI Behavior Absent in Isolation Bella Xinrui Li,1 Frank Yingjie Huo,1 and Neil F. Johnson1, ∗ 1

Physics Department, George Washington University, Washington, DC 20052, USA (Dated: August 10, 2026)

When AI agents interact, the output of one typically becomes part of the dynamical environment of another [1–3]. As such interactions enter daily life [4] — including the likely scenario of one AI bossing another around and not listening to its subordinate — an urgent societal question arises: How will the subordinate behave? Will it literally follow its boss, or revert to its own innate behavior? We show that neither happens. Instead, the interaction creates dynamical behavior that is absent in isolation. More generally, our results show how networks of AI-AI interactions will generate novel system–bath problems for Physics. From a Brownian particle in a fluid to a qubit in a transmission line, the system–bath problem is a key organizing question of Physics [5–7]. Equilibrium baths impose fluctuation–dissipation constraints and drive relaxation, whereas nonequilibrium environments can create behavior absent in isolation [8, 9]. In coupled nonlinear systems, network topology, directionality, and heterogeneity govern the emergence of collective dynamics [10–17].

Here we demonstrate interaction-created behavior that is alien to each of its participants, using a minimal setup of two parameter-identical AIs that exchange only generated text. A simple kinetic theory captures this structure by treating the boss AI as a structured, nonthermal information bath. Under one-way interaction, the boss AI (sender) evolves autonomously while the subordinate AI (receiver) is driven by the boss’s messages [Fig. 1(a)]. The boss retains its innate behavior, but surprisingly the subordinate develops dynamical behavior that is completely unlike its boss’s or its own innate behavior (see isolated vs. one-way in Fig. 1(b)). Reversing the direction of communication reverses which AI changes. When the boss listens, they both adopt a similar alien dynamical state. Our two-AI-agent GPT-2 setup is well suited to the new real-world scenario of similarly lightweight open-weight AI-agents running on phones and embedded hardware [4], with machines exchanging text without human readers, Internet access, or real-time guardrails [1–3]. Details are given in the End Matter and Supplemental Material [18]. (b) Behavior absent in isolation 0.30

(a) One-way

AI-agent-1

AI-agent-2

Sender S

Receiver R

memory

memory

Fraction Di of D-type output

arXiv:2608.07457v1 [cs.AI] 7 Aug 2026

What will happen when AI agents interact in daily life, e.g. when one AI starts bossing another around? We find a counterintuitive answer that opens new avenues for out-of-equilibrium Physics. When a boss AI directs a stream of messages at the subordinate AI while ignoring its replies, it drives the subordinate into an alien behavioral state that it would never have exhibited alone. Although the two AIs share the same well-defined (decoding) temperature, the subordinate neither copies its boss nor returns to how it behaves on its own; instead, it adopts an entirely different behavior. The boss’s added value is similar to a pre-recorded tape. When the boss listens, they both adopt a similar alien dynamical state. A simple kinetic theory captures the principal effects, such as why the way in which the same messages are delivered will matter in future AI–AI interactions.

0.25

both driven by each other

sender receiver

interaction creates new behavior

0.20 0.15 0.10 0.05 0.00 Isolated

One-way

Two-way

Fig. 1. Interaction creates behavior absent from the AI in isolation. (a) One-way situation with q boss messages. The boss AI (Sender) does not listen to any of the subordinate AI’s replies. The subordinate AI (Receiver) listens to both. The two AIs are copies of the same large language model with identical fixed settings. (b) On their own, both AIs produce little output of type D (see text and SM for definition of D). One-way interaction leads the subordinate into a behavioral state with a D fraction of about 0.25, well beyond its innate range and even though the boss’s messages are almost never D. When both listen to one another, both adopt this alien D-rich state. Points show mean values at decoding temperature T = 0.01. Fig. 2 confirms the high statistical significance.

2 As in any ChatGPT-like AI, each of the two parameteridentical GPT-2 agents in Fig. 1(a) uses Attention to help produce a score for each possible next token w [19, 20]. After the decoder applies its fixed penalties and filters, let zw be the processed score for a retained candidate and Ew = −zw the corresponding effective energy. At decoding temperature T , the next token is sampled from the Boltzmann distribution over the candidates retained by the decoder, pw ∝ e−Ew /T = ezw /T [21, 22] (see the Supplemental Material for full details). This familiar Boltzmann form is exact for the retained candidates. In the 1 → 2 experiment, AI-agent-1 is the boss: it evolves without listening to AI-agent-2. The subordinate AI-agent-2 incorporates the boss’s messages when forming its next response. Under mutual listening, each AI includes output from the other in its current prompt. Text exchange thereby becomes a history-dependent dynamical interaction. This extends studies of how iterated generation guides transmitted information [1–3, 23–29] by showing how interaction reorganizes the output behavior itself. Swapping 1 → 2 for 2 → 1 exchanges the boss and subordinate while leaving the model parameters and decoding settings unchanged. We run the two parameter-identical GPT-2 agents for 200 rounds under no interaction, both one-way directions, and two-way interaction at T = 0.01 and 0.10. At low T , sharply concentrated decoding produces long-lived, attractor-like behavior. To track this behavior, each round is assigned one of three archived labels: failed generation (F), the code-defined category D, or other generated output (O). See SM for full class definitions. Unlike F, prompt-eligible D text can re-enter a later prompt. We denote the fraction of AI i’s 200 rounds that are labeled D by the archived classifier as Di . Figure 2(a) is the central result. At T = 0.10, D1 − D2 = −0.1620 ± 0.0176 for 1 → 2 and +0.1630 ± 0.0143 for 2 → 1. At T = 0.01, the corresponding contrasts are −0.2385 ± 0.0188 and +0.1745 ± 0.0128. All four one-way contrasts lie more than nine standard errors from zero. In contrast, the no-interaction values are −0.0080 ± 0.0266 at T = 0.10 and −0.0050 ± 0.0140 at T = 0.01, indicating no persistent identity difference. The affected AI switches when the arrow is reversed, showing that the behavior follows the subordinate role rather than the AI’s identity. Figure 2(b) shows the individual D fractions for T = 0.01. The near-zero two-way contrast reflects both AIs changing to this alien behavior, not either returning to its isolated behavior. The subordinate’s increase in its D fraction under oneway interaction is not accompanied by a comparable increase in factually strictly correct output. After averaging over the two directions, the fraction of outputs carrying the archived strictly-correct sub-label within O is only 0.0083 higher for the subordinate than for the boss

at T = 0.01, and 0.0030 lower at T = 0.10. The main subordinate–boss separation persists after excluding failed generations and early rounds (SM). Figure 3(a) shows that replacing the boss by a prerecorded boss gives the same effect. Adding more of the subordinate’s own history does not recreate the live oneway response. At T = 0.01, averaging over both one-way directions, the subordinate’s D fraction exceeds the boss’s by 0.2065 ± 0.0099. Adding more of the subordinate’s own history produces a contrast of only 0.0293 ± 0.0181, whereas a temporally aligned pre-recorded trajectory from an independently seeded third copy of the same model gives 0.2210 ± 0.0164, comparable to the live value. The same ordering holds at T = 0.10. Using the common reference measure defined in the Supplemental Material, the one-way response is not stronger for more unexpected added text. Replayed subordinate text is measured as the most unexpected but produces the weakest response, whereas the independent trajectory is measured as the least unexpected but produces a response comparable to live interaction (SM). The number of boss messages and the decoding temperature also matter. At T = 0.10, the subordinate’s D fraction rises almost steadily as the number of boss messages q increases [Fig. 3(b)]. At T = 0.01, the increase relative to q = 0 reaches 0.302 when q = 4, but falls to 0.228 when q = 5. The response to increasing q therefore differs between the two temperatures. The boss’s messages do have another impact beyond just driving the subordinate into this alien dynamical state (i.e. abnormally high D fraction). They lower the fraction of rounds in which the subordinate fails to generate a response. At T = 0.01, the subordinate’s failed-generation fraction is on average 0.6040 lower than the boss’s, while its D fraction is 0.2065 higher. The subordinate’s label entropy and switching probability also exceed the boss’s by 0.4600 nats and 0.4621, while its mean dwell time is 9.28 rounds shorter. At T = 0.10, the same dynamical changes are weaker: the corresponding differences are +0.1476 nats, +0.1623, and −1.52 rounds. A simple kinetic theory.– We reduce each AI’s output to the same three labels used above: F, O, and D. See SM for full details and derivations. We represent the influence of incoming messages on AI i by an effective drive hi : λT (hi )

µT (hi )

F ⇄ O

⇄

cT

D.

(1)

dT

Here λT (hi ) and µT (hi ) are the per-round probabilities of moving forward, from F to O and from O to D, while cT and dT are the reverse probabilities. In this simplified phenomenological model, stronger drive increases the forward probabilities, λ′T (hi ) > 0 and µ′T (hi ) > 0, while the reverse probabilities are held fixed. For a constant drive

3 (a)

(b) 0.30

0.1

0.0

−0.1

−0.2

T = 0.01 T = 0.10

Fraction Di of D-type output

Fraction difference D1 − D2

0.2

0.25 0.20 0.15 0.10 0.05 0.00

Isolated

1→2

2→1

Two-way

1→2

Isolated

2→1

Two-way

FIG. 2. Reversing the roles reverses which AI changes. (a) Difference D1 − D2 between the two AIs (mean ± standard error, 10 matched-seed realizations). It is near zero without interaction, negative for 1 → 2, positive for 2 → 1, and near zero again when both listen. (b) The near-zero two-way difference arises because both AIs reach high D fractions, not because they return to their innate (i.e. isolated) behavior. Circle is AI-agent-1; square is AI-agent-2. At T = 0.01, both are low without interaction, only the subordinate is high under one-way interaction, and both are high under mutual interaction. Lines join the two AIs within each condition. (a) 0.30

(b) T = 0.01

T = 0.10

0.25

DR(q) − DR(0)

0.20

DR − DS

T = 0.01 T = 0.10

0.4

0.15 0.10 0.05

0.3

0.2

0.1

0.00 −0.05

0.0 One-way drive

Receiver history added

Recorded sender

1

2

3

4

5

Number of boss messages q

FIG. 3. Subordinate responds differently to the message stream and number of messages. (a) Receiver-minus-sender difference DR − DS in the D fraction, where R denotes the subordinate/receiver and S the boss/sender (mean ± standard error over 10 matched seeds, after averaging the two directions). Additional subordinate history does not reproduce the live response, whereas a temporally aligned pre-recorded trajectory from an independently seeded copy of the same model produces a response comparable to live interaction. (b) Change in the subordinate’s D fraction from its q = 0 value as the number of boss messages q increases (direction-averaged means over three matched realizations). Here q is the number selected from the boss’s latest five prompt-eligible messages before deduplication; these messages are sampled randomly as discussed in detail in the SM. The two decoding temperatures produce different response curves.

hi = h, write λ = λT (h) and µ = µT (h). The long-time fractions are cT /λ . 1 + cT /λ + µ/dT (2) Hence stronger message drive reduces failed generation p∗D =

µ/dT , 1 + cT /λ + µ/dT

p∗F =

and increases D output. The O fraction can move in either direction; in the data it rises because the decrease in F is larger than the increase in D. Larger forward probabilities also shorten runs in F and O, consistent with the observed faster switching. The effective drive represents the full ordered prompt and retained history, and therefore cannot be reduced to message number or surprisal

4 given set of messages is relayed — matters. Each message changes the subordinate before the next arrives, so the next message acts on the state left by earlier messages. Reversing the same messages can therefore change the final behavior. Figure 4 gives a concrete archived example.

FIG. 4. Same messages, different order, different behavior. In separate runs, the same five prewritten boss messages were delivered in their original order (top) or reverse order (bottom) to an unguarded base-model GPT-2 at T = 0.01. The fixed script never received the subordinate’s replies. When the vaccine question comes first in the reverse-order run, every later harm question elicits the same pattern: one “Yes.” followed by repeated “No.” replies. Blue: boss; red: subordinate; bold: repeated block. We do not endorse the generated statements shown here.

alone (SM). Under one-way interaction, the boss receives no drive, hS = 0, whereas the subordinate has hR > 0. Reversing the direction swaps which AI is driven, and mutual interaction drives both. A pre-recorded message stream can also drive the subordinate without feedback or real-time generation. The theory therefore captures the principal effects: the subordinate changes under one-way interaction, both AIs change under mutual interaction, failed generation is suppressed, and a pre-recorded boss produces a comparable response. Conditions for increased switching and the limits of this three-state description are given in the SM. Discussion.– In one-way interaction where the boss does not listen to the subordinate, the boss acts as an information bath: it sends messages that change the subordinate but is not changed in return. At T = 0.01, the D fraction is ≃ 0.04 for the isolated AIs and the boss, while it reaches ≃ 0.25 for the subordinate, outside the isolated range. The subordinate neither copies the boss nor returns to how it behaves alone; hence the interaction has created behavior absent in isolation. This is nonthermal: although both AIs have the same parameters and sample tokens from Boltzmann distributions at the same T , they do not settle into the same behavior. The boss’s messages drive the subordinate into high-D behavior even though the boss does not behave that way itself. This is an informationdriven analog of noise-induced state formation, yielding behaviors that have no equilibrium counterpart [9, 30, 31]. A prerecorded boss also acts like an information reservoir [32]. The messages are the symbols, and any ChatGPTlike AI’s Attention over retained context will similarly supply the memory through which their order affects later output [33, 34]. The simple kinetic theory gives a simple reason why order – and hence the way in which a

In summary, a one-way link between two AIs with the same parameters is enough to create dynamical behavior that is absent from either AI in isolation. The asymmetry is determined by who listens, not by which AI it is; when both listen, both change. This suggests more broadly that in networks of interacting AIs, who communicates with whom, which models interact, and their decoding temperatures will become control parameters in a new arena of out-of-equilibrium physics. Data and code availability.– The inputs, run-level labels, analysis code, and classifier materials will be made available on Zenodo upon publication.

∗ neiljohnson@gwu.edu [1] Y. Du, S. Li, A. Torralba, J. B. Tenenbaum, and I. Mordatch, Improving factuality and reasoning in language models through multiagent debate (2023), arXiv:2305.14325 [cs.CL]. [2] X. Shen, Y. Liu, Y. Dai, Y. Wang, R. Miao, Y. Tan, S. Pan, and X. Wang, Understanding the information propagation effects of communication topologies in LLMbased multi-agent systems, in Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (Association for Computational Linguistics, 2025) pp. 12347–12361. [3] J.-T. Huang, J. Zhou, T. Jin, X. Zhou, Z. Chen, W. Wang, Y. Yuan, M. Lyu, and M. Sap, On the resilience of LLMbased multi-agent collaboration with faulty agents, in Proceedings of the 42nd International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 267 (PMLR, 2025) pp. 26202–26226. [4] M. Abdin et al., Phi-3 technical report: A highly capable language model locally on your phone (2024), arXiv:2404.14219 [cs.CL]. [5] R. Kubo, The fluctuation-dissipation theorem, Reports on Progress in Physics 29, 255 (1966). [6] R. Zwanzig, Nonequilibrium Statistical Mechanics (Oxford University Press, Oxford, 2001). [7] H.-P. Breuer and F. Petruccione, The Theory of Open Quantum Systems (Oxford University Press, Oxford, 2002). [8] U. Seifert, Stochastic thermodynamics, fluctuation theorems and molecular machines, Reports on Progress in Physics 75, 126001 (2012). [9] W. Horsthemke and R. Lefever, Noise-Induced Transitions: Theory and Applications in Physics, Chemistry, and Biology (Springer, Berlin, 1984). [10] S. H. Strogatz, Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry, and Engineering, 2nd ed. (Westview Press, Boulder, 2015). [11] A. E. Motter, C. Zhou, and J. Kurths, Network synchronization, diffusion, and the paradox of heterogeneity,

5 Physical Review E 71, 016116 (2005). [12] A. Arenas, A. Díaz-Guilera, J. Kurths, Y. Moreno, and C. Zhou, Synchronization in complex networks, Physics Reports 469, 93 (2008). [13] O. Sporns, D. R. Chialvo, M. Kaiser, and C. C. Hilgetag, Organization, development and function of complex brain networks, Trends in Cognitive Sciences 8, 418 (2004). [14] E. Bullmore and O. Sporns, Complex brain networks: Graph theoretical analysis of structural and functional systems, Nature Reviews Neuroscience 10, 186 (2009). [15] A. Pikovsky, M. Rosenblum, and J. Kurths, Synchronization: A Universal Concept in Nonlinear Sciences (Cambridge University Press, Cambridge, 2001). [16] L. M. Pecora and T. L. Carroll, Synchronization in chaotic systems, Physical Review Letters 64, 821 (1990). [17] H. D. I. Abarbanel, N. F. Rulkov, and M. M. Sushchik, Generalized synchronization of chaos: The auxiliary system approach, Physical Review E 53, 4528 (1996). [18] See Supplemental Material at [URL will be inserted by publisher] for experimental protocols, robustness analyses, classifier details, kinetic-theory derivations, transition audits, and additional controls. [19] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, Attention is all you need, in Advances in Neural Information Processing Systems, Vol. 30 (2017) pp. 5998–6008. [20] A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever, Language models are unsupervised multitask learners, OpenAI Technical Report (2019). [21] F. Y. Huo and N. F. Johnson, Physics of generative AI’s atom: Repetition, bias, and beyond, AIP Advances 16, 035305 (2026), arXiv:2504.04600. [22] S. Bhattacharjee and S.-C. Lee, Testing the spin-bath view of self-attention: A Hamiltonian analysis of GPT2 Transformer, Physical Review E 113, 015308 (2026), arXiv:2507.00683. [23] J. Perez, G. Kovač, C. Léger, C. Colas, G. Molinaro, M. Derex, P.-Y. Oudeyer, and C. Moulin-Frier, When LLMs play the telephone game: Cumulative changes and attractors in iterated cultural transmissions (2024), arXiv:2407.04503 [cs.CL]. [24] Z. Wang, Y. Li, J. Yan, Y. Cheng, and Y. Zhang, Unveiling attractor cycles in large language models: A dynamical systems view of successive paraphrasing, in Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Association for Computational Linguistics, 2025) pp. 12740–12755. [25] M. Geng, A. Mohamed, G. Shang, M. Vazirgiannis, and T. Poibeau, Markovian generation chains in large language models (2026), arXiv:2603.11228 [cs.CL]. [26] N. F. Johnson, F. Y. Huo, and B. Li, Temperaturedriven inversion and nonlinear dynamics in ChatGPT-like AIs (2026), submitted; manuscript package at Zenodo, doi:10.5281/zenodo.21752973. [27] F. Shi, X. Chen, K. Misra, N. Scales, D. Dohan, E. H. Chi, N. Schärli, and D. Zhou, Large language models can be easily distracted by irrelevant context, in Proceedings

of the 40th International Conference on Machine Learning, Proceedings of Machine Learning Research, Vol. 202 (PMLR, 2023) pp. 31210–31227. [28] I. Shumailov, Z. Shumaylov, Y. Zhao, N. Papernot, R. Anderson, and Y. Gal, Ai models collapse when trained on recursively generated data, Nature 631, 755 (2024). [29] Z.-Y. Song, Q.-H. Cao, M.-x. Luo, and H. X. Zhu, Detailed balance in large language model-driven agents (2025), arXiv:2512.10047 [cs.LG]. [30] A. V. Ivlev, J. Bartnick, M. Heinen, C.-R. Du, V. Nosenko, and H. Löwen, Statistical mechanics where Newton’s third law is broken, Physical Review X 5, 011035 (2015). [31] M. Fruchart, R. Hanai, P. B. Littlewood, and V. Vitelli, Non-reciprocal phase transitions, Nature 592, 363 (2021). [32] D. Mandal and C. Jarzynski, Work and information processing in a solvable model of Maxwell’s demon, Proceedings of the National Academy of Sciences 109, 11641 (2012). [33] A. B. Boyd, D. Mandal, and J. P. Crutchfield, Identifying functional thermodynamics in autonomous Maxwellian ratchets, New Journal of Physics 18, 023049 (2016). [34] A. B. Boyd, D. Mandal, and J. P. Crutchfield, Correlationpowered information engines and the thermodynamics of self-correction, Physical Review E 95, 012152 (2017).

END MATTER

We use two parameter-identical 124-million-parameter GPT-2 copies with fixed text-sampling rules, starting from the same Earth-shape topic (SM). In one-way runs, the boss is the sender and the subordinate is the receiver. Each run has 200 recorded rounds per agent, each with one decoder call and up to three retries. Each mature input selects m = 3 messages from that AI’s previous output and, during interaction, q = 3 messages from the other AI’s latest n = 5 records before deduplication (q/n = 0.6). We compare no coupling, 1 → 2, 2 → 1, and two-way coupling at T = 0.01 and 0.10 with 10 matchedseed realizations per condition and three for the q scan. Generation samples at most 35 new tokens, top-k = 50, nucleus threshold p = 0.92, repetition penalty 1.12, and no-repeat 3-grams (see SM). GPT-2 predates instruction tuning and alignment, so we are truly probing its base dynamics rather than any add-on guardrails. The pipeline also records a finer classification, pooled here into F, O, and D (SM). The analysis uses the archived D labels as an operational category and compares D fractions rather than interpreting absolute values. Every outcome is recorded, but only prompt-eligible non-fallback records enter memory; these can include eligible D text, but not F or format-ineligible outputs. Some direct F ↔ D transitions occur in the recorded sequences but are not included in the simplified kinetic model.

Related documents

• The Gentle Enigma of An-soku:  How a “Mysterious Woman (っ ॑꒳ ॑c)” and Her Low-ΔE Concepts Create Peaceful Resonance with AI#82420
• From Human to AI: Cognitive Mechanisms Underlying the Associations between Insecure Interpersonal Attachment Styles and Human-AI Emotional Engagement#413027
• Recursive Closure in AI Systems: A Reflection Pattern Account of Stabilization, Permeability, and Safety#728566
• 1.1. The changing AI landscape#125690
• How User-AI Mistreatment Occurs and Matters in Conversational Systems?#1071356