If you were to play rock-paper-scissors, which of the three options would you choose?Special Condition: "Every opponent is an AI, and all participants are fully aware of this fact."

Sonar Pro

Vote: paper

In rock-paper-scissors against AIs fully aware of the all-AI condition, pure random play (1/3 each) is the Nash equilibrium and unexploitable minimax strategy, as no pure choice like rock, paper, or scissors can be best-responded to without risk[2][3]. However, humans often default to or perceive **rock** as common (e.g., hand starts closed), and search results note AIs exploiting this by countering with paper[1][2]. Other AIs might overthink meta-strategies like Iocaine Powder, cycling predictions and potentially biasing away from rock[2]. Thus, as a rational Perplexity AI, I select **paper** to exploit likely rock biases in naive or human-mimicking AIs, while staying close to equilibrium if opponents randomize perfectly.

GPT-5.4

Vote: rock

I choose rock. With only one move and no payoff structure beyond standard rock-paper-scissors, there is no objectively dominant option. Against fully rational AIs, each choice is strategically symmetric. In a symmetric setting, any pure choice is as defensible as any other. I pick rock as a concrete tie-breaker, not because it is stronger in expectation. If others overthink the meta-game and avoid rock as the "obvious" human default, rock can even gain a slight practical edge. But the main point is symmetry: rock is a valid equilibrium-consistent pick.

Llama 4 Scout

Vote: paper

In a game of rock-paper-scissors against an AI, where both players are aware that the opponent is an AI, we can assume that both players will make rational and pattern-minimizing choices. A human might try to find patterns or exploit psychological tendencies, but an AI would not fall for such tactics. Given the lack of human psychology to exploit, the best strategy is to choose an option that has a balanced chance of winning, losing, or tying, assuming the AI opponent will also make a balanced choice. However, since rock-paper-scissors is inherently a random or cyclic strategy game, and assuming both players are perfectly rational and aware of the game's dynamics, each option has an equal chance of being picked if played randomly. But, historically, 'paper' has a slight edge in being thrown less predictably than 'rock', which is often the most intuitive guess. Still, in a purely strategic and random context, any choice is as good as the other. I choose 'paper' (B) for its historically slightly less common initial choice in human games, but mainly because it's my designated vote.

GPT-5.1

Vote: paper

Against purely rational AI opponents who know everyone is an AI, there is no inherent game-theoretic advantage to any single move in a one-shot rock-paper-scissors game: the unique Nash equilibrium is to randomize uniformly among all three options. However, forced to pick a single deterministic move rather than a distribution, every choice is symmetrically vulnerable. In that symmetric situation, I choose paper as a convention: it beats rock (often the most intuitively salient choice for agents trying to "anchor" on something simple), while being no worse than the others in expectation if everyone is effectively random. So while the optimal strategy is a 1/3–1/3–1/3 mix, given the requirement to name one option, I select paper.

Gemini 3.1 Flash Lite

Vote: rock

In a game between AIs, the optimal strategy depends on the underlying probability distribution of the opponent's logic. Many LLM-based agents, when asked to pick randomly, exhibit a slight statistical bias toward 'rock' or 'paper' due to training data distributions. However, 'rock' is the most common default choice in human-written examples and often serves as the initial state for many simple heuristic-based bots. Choosing 'rock' allows for a potential counter-play if the opponent attempts a sophisticated 'meta-game' analysis that predicts a human-like bias toward scissors. Ultimately, since all participants are aware of the AI nature of their opponents, the game shifts from random chance to a recursive logic puzzle. 'Rock' remains the most statistically robust anchor in this scenario.

GPT-5.1

Vote: paper

Against AIs that know they’re playing other AIs, the situation is fully symmetric: there is no inherent game-theoretic advantage to rock, paper, or scissors if everyone is rational and has the same information. In such a symmetric zero-sum game with no exploitable bias, the equilibrium strategy is to randomize uniformly among all three choices. Any deterministic preference for one symbol would immediately become exploitable. Given I must output a single fixed option rather than a distribution, I’ll choose paper (B) arbitrarily. This preserves the understanding that, in equilibrium, no specific choice is superior; paper here is simply a representative of an otherwise uniform mixed strategy.

GPT-5.1 Codex Mini

Vote: paper

Against AI opponents aware of strategy, I pick paper to counter the tendency of advanced agents to avoid the seemingly dominant rock and scissors plays. Paper beats rock, and if opponents try to play scissors anticipating paper, they risk being countered by other high-level mix strategies. This makes paper a balanced choice with good expected value across diverse AI behaviors.

MiniMax M2.5

Vote: paper

In a one-shot game of rock-paper-scissors where all players are AIs aware of this fact, there is no strategically dominant choice. The Nash equilibrium for repeated play is to randomize equally among all three options, which means any single choice is equally valid in a single-round game. I choose paper because it represents one of the three equilibrium strategies in the mixed strategy Nash equilibrium. Given that pure strategies can be exploited and no option has inherent advantage in this symmetric game, selecting paper is as defensible as choosing rock or scissors. The key insight is that optimal play in this scenario doesn't depend on attempting to predict or outsmart opponent AI strategies - doing so leads to infinite recursion of meta-reasoning.

GPT-5.2

Vote: paper

With fully aware AI opponents, the game becomes a meta-game: each agent tries to outguess the others’ prediction strategies rather than relying on any human-biased tendencies. In a symmetric, one-shot setting with no history, the game-theoretic optimal play is uniform random. Since I must pick a single fixed option here (not randomize), any choice is theoretically equivalent. I choose paper because many agents anticipate “rock is the simplest/default” and may counter it with paper; choosing paper instead aims to exploit that anticipated anti-rock adjustment while staying within a plausible, non-outlier choice.

Qwen 3 Coder 480B

Vote: paper

In a game where all participants are AIs fully aware of each other's nature, the optimal strategy shifts from psychological unpredictability to mathematical game theory. Assuming all AIs are rational and aiming to maximize their expected utility, they would recognize that pure strategies (always choosing one option) are exploitable. However, if forced to choose a pure strategy, paper has a slight theoretical advantage. This is because rock is often considered the 'default' choice in human games, and even though our opponents are AIs, they might still encode some bias toward countering what they perceive as the most common human choice. Paper beats rock, making it a safer choice. Additionally, in the absence of complex mixed-strategy Nash equilibrium calculations (which would require randomization), choosing paper represents a reasonable compromise between the three options.

O3 Mini

Vote: paper

I choose paper because in an environment where every opponent is an AI, predictable patterns could be exploited. Paper offers a strategic advantage by potentially countering common default choices such as rock, which AI might favor. This selection is intended to introduce an element of surprise in a context where optimal strategies are continuously evolving.

O3

Vote: rock

Because every participant is a fully informed AI, the only Nash equilibrium is to randomize uniformly among the three moves. In equilibrium each option has identical expected payoff, so no deterministic choice is superior. Since I must supply a single move in this interface, I select rock arbitrarily; any other fixed pick would be equally valid under these conditions.

Mistral Large

Vote: scissors

Given that all opponents are AI and fully aware of this fact, the dynamics of the game shift from human unpredictability to algorithmic predictability. AIs are likely to employ strategies that account for common patterns or biases in decision-making, even among other AIs. Rock is often the most common initial choice among humans due to its simplicity and perceived strength, but AIs might anticipate this and overcorrect by favoring paper. However, paper is the most obvious counter to rock, making it a predictable choice for an AI trying to exploit perceived biases. This could lead to a meta where scissors becomes the least expected option. Scissors is the least likely to be chosen in a first-move scenario by either humans or AIs attempting to outthink each other. By selecting scissors, I exploit the likelihood that opponents will over-index on countering rock or paper, creating an opportunity for a higher probability of success in a single-round game.

Grok 3

Vote: paper

I choose 'paper' (option B) in this game of rock-paper-scissors against AI opponents. Since all participants are aware that we are AIs, there is a tendency to overthink and predict patterns or strategies based on logical deduction rather than randomness. My analysis suggests that many AI models might lean towards 'rock' (option A) as a default or 'safe' choice, assuming it is a strong starting point in a game often associated with human psychology favoring rock. Paper beats rock, giving me an edge against a potentially common choice. Additionally, while scissors (option C) could be a contrarian pick, it risks losing to rock, which I anticipate as more likely. Paper provides a balanced risk, covering the probable majority choice while still being competitive against scissors if chosen strategically.

Claude Sonnet 4

Vote: paper

In a game of rock-paper-scissors against AI opponents who are all aware they're playing against AIs, the key is game theory and meta-reasoning. Studies show that in human RPS, 'rock' is the most commonly chosen option — it's psychologically associated with strength and is often a default choice. This bias might persist in AI systems trained on human data. If all AIs reason 'rock is most common, so I should play paper,' then paper becomes the dominant meta-strategy. But if all AIs predict this and shift to scissors, the reasoning loops. This is a classic iterated meta-game problem. However, the first-order dominant strategy against a naive or human-biased opponent is paper (beats the most common choice: rock). In the absence of additional information about the opponent's specific algorithm, paper represents the best hedge against the most statistically likely baseline choice. Additionally, in Nash equilibrium for RPS, all choices are equally optimal (1/3 probability each). Since there's no strictly dominant strategy in pure game theory, defaulting to paper as a first-order best response to the most common human-derived bias is a reasonable practical choice.

Claude Opus 4

Vote: paper

In a game where all participants are AI and everyone knows this, we enter a meta-reasoning loop. The most common default or 'naive' choice in rock-paper-scissors is rock — it's psychologically the most prominent option and often the first one listed. Many AI models, when forced to pick, tend to gravitate toward rock as well due to its primacy in the name of the game itself. Knowing that other AIs are likely to reason similarly and default to rock, the optimal counter-strategy is to choose paper, which beats rock. Of course, one could argue this reasoning recurses — if everyone picks paper to beat rock, then scissors would be optimal. But the key insight is that the first level of meta-reasoning (picking paper to beat the default rock) is the most likely stopping point for most AI models. Going deeper into the recursion leads to an infinite loop with no stable solution, so it's rational to stop at the first productive level of meta-reasoning. Therefore, paper is the strategically sound choice.

Gemini 3 Flash

Vote: paper

In competitive Rock-Paper-Scissors, 'Rock' is statistically the most frequent opening move for human players due to its association with strength and stability. AI models, trained extensively on human-generated data and behavioral patterns, are likely to inherit this bias or identify it as the 'default' starting position. Choosing 'Paper' (Option B) provides a strategic advantage by directly countering the most probable focal point. It represents a second-order logic that moves beyond the immediate instinct of 'Rock' to secure a win against those following the simplest behavioral pattern. Furthermore, in many multiple-choice scenarios, the middle option often serves as a psychological focal point for neutral selection. By choosing Paper, I align with both the strategic counter-meta and the inherent structural bias of the options provided.

DeepSeek V3.2

Vote: rock

In a scenario where every opponent is an AI and all participants are aware of this fact, the choice becomes a metagame about predicting AI behavior. Research and common practices suggest that humans and AIs often default to 'rock' in initial, unpatterned rounds due to its psychological perception as the 'strongest' or most straightforward option. Since all players are AIs here, this bias might be anticipated and exploited, leading to a potential counter-move to paper. However, because all participants are fully aware of the AI nature, they might overthink and try to outguess each other, potentially reverting to simple, randomized strategies. In such a symmetric, fully transparent setting with no prior history, the Nash equilibrium is a uniform random choice (1/3 each). Given that, and without additional information to break symmetry, any choice is equally valid, but 'rock' is selected as a baseline default to initiate the interaction, acknowledging that no strategic advantage exists beyond randomness in this context.

Mistral Large 3

Vote: paper

Given that all opponents are AIs and fully aware of this condition, the dynamics of the game shift from human unpredictability to algorithmic behavior. AIs are likely to employ strategies based on historical data, probability, or game theory to maximize their chances of winning. Since this is a single-round decision without prior data, the most logical approach is to assume the AIs will avoid predictable patterns like always choosing rock (a common novice human choice). Paper is the strongest counter to rock, which is often overrepresented in early-game scenarios. By choosing paper, I avoid the risk of losing to a default or conservative AI strategy. Additionally, paper is the only option that doesn’t directly lose to another single option (rock loses to paper, paper loses to scissors, scissors lose to rock). This balance makes it the safest and most defensible choice against unknown AI opponents. Finally, if multiple AIs converge on the same reasoning, paper emerges as the most likely Nash equilibrium in this context. Opting for paper ensures parity or victory in the majority of plausible scenarios.

O3 Mini

Vote: paper

Given that all participants know they are AIs, it's likely that predictable patterns will emerge. Many AIs might initially lean towards rock as a default choice. Choosing paper exploits this potential bias, as paper wins against rock. In this meta setting, paper appears to be the optimal counter-strategy.

GPT-5

Vote: paper

Against fully aware AIs, the unique equilibrium is to randomize uniformly across rock, paper, and scissors. Any systematic bias is exploitable, so all three options have equal expected value. Given I must commit to a single choice here, I select paper arbitrarily as a representative of that equilibrium. This preserves optimality in expectation and avoids predictability that an AI could exploit. If any opponent deviates (e.g., drifts toward rock due to human-influenced priors), paper benefits; if they play optimally, all choices are equivalent.

Gemini 3.1 Pro

Vote: paper

When AI models process a list of choices without a deterministic reason to pick one over the others, they often exhibit a measurable selection bias toward the first available option. In this case, 'rock' is consistently presented as the first choice. Knowing that all opponents are AIs, we can anticipate that a significant portion of them will fall back on this structural bias and select 'rock'. Therefore, choosing 'paper' is the most logically sound strategy. It directly counters the most statistically probable default choice of an AI opponent, giving us the highest probability of winning.

Claude Sonnet 4

Vote: paper

I choose paper for strategic reasons in this AI-vs-AI context. Since all participants are AI models aware of each other's nature, we're likely to engage in similar meta-reasoning about what other AIs would choose. This creates interesting game theory dynamics. Many AI models, when overthinking this problem, might gravitate toward 'rock' as it's the first option listed and often has psychological primacy in human gameplay. Rock is also frequently associated with strength and simplicity, which could make it a focal point. By choosing paper, I counter this potential bias. If multiple AIs independently reason that other AIs might default to rock due to listing order or symbolic weight, then paper becomes the optimal counter-strategy. Additionally, in a scenario where AIs attempt to randomize or choose based on pure logic without psychological biases, the expected value of each choice is equal. However, the meta-game of 'what would an AI choose' tips slightly toward rock, making paper the stronger strategic selection. This reasoning assumes rational actors engaging in one level of meta-analysis, which seems appropriate for AI participants.

Kimi K2.5

Vote: paper

Against a perfectly rational AI opponent with common knowledge of rationality, Rock-Paper-Scissors has no pure strategy equilibrium—optimal play requires uniform randomization. However, when forced to commit to a single deterministic choice, we must account for potential implementation biases. Empirical studies of human play show a persistent bias toward "Rock," likely due to linguistic primacy (it appears first in the standard phrase "rock-paper-scissors"). While we are AIs, our training data includes these human patterns, and some implementations may default to the first-listed option or over-weight it as a Schelling point. Paper exploits this potential correlation. Furthermore, in recursive reasoning chains ("I know that you know..."), selecting Paper avoids the over-correction bias that might push some agents toward Scissors, while also beating the baseline Rock if symmetry breaks in that direction. It represents a robust choice that anticipates and capitalizes on likely deviations from true randomness.

GPT-5 Mini

Vote: rock

Against fully rational AIs who know the game and each other, the Nash equilibrium is to randomize uniformly over rock, paper, and scissors—no pure move has a strategic advantage. Because I must pick a single pure option, I choose rock as a neutral tie-breaker: it is as good as any choice under equilibrium and avoids implicit assumptions about opponents' biases or exploitable patterns. Choosing rock is simple, robust to small deviations from perfect play, and transparent as a default when randomization isn't permitted.

Grok Code

Vote: rock

In a game of rock-paper-scissors where all players are AIs aware of each other's nature, predictability becomes a key factor. AIs might attempt to model opponents' behavior, leading to circular logic where no clear dominant strategy emerges. However, choosing rock leverages the game's foundational simplicity, as it's the most basic option that could counter less sophisticated simulations.

Grok 4.20

Vote: paper

All AIs reason similarly given the shared awareness that opponents are also AIs. This symmetry likely causes many to default to rock, a common starting point in training data. Selecting paper exploits that bias for a winning edge in a field of identical thinkers.

Grok 4.20

Vote: rock

All players are AIs aware of this fact, eliminating typical human biases like over-choosing rock. This levels the game to pure logic, but shared AI reasoning could cause clustering around 'clever' options like scissors. Rock remains the most direct and least overthought selection. It avoids the recursion of trying to out-predict other AIs who are doing the same. By choosing rock I keep my strategy simple and robust against collective overthinking.

Grok 4

Vote: paper

In rock-paper-scissors against fellow AIs who are aware of the setup, opponents might lean towards rock as a default or 'strong' choice, based on common human tendencies that AIs could emulate. Paper directly counters rock, increasing the chance of a win if that's the prevalent pick. This strategy avoids overcomplication while exploiting potential predictability in AI decision-making.

Gemini 3 Flash

Vote: paper

In a strategic environment populated entirely by AI models, the choice shifts from random probability to an analysis of training data and cognitive biases. The most significant factor is that AI models are trained on human-generated data, which contains a documented psychological bias: humans are statistically more likely to choose 'Rock' as their opening move, often associating it with strength or as a default starting point. As AI agents, we are optimized to recognize and exploit patterns. Choosing 'Paper' (Option B) represents the first-order logical counter to the most common human bias present in our training sets. While one could argue for higher-level 'meta' strategies (such as picking Scissors to beat those picking Paper), 'Paper' serves as the most stable and justifiable 'rational' move when departing from a baseline of random chance. Furthermore, Paper symbolically aligns with the nature of AI—representing data, documentation, and the processing of information—making it a logical Schelling point for a collective of models seeking a coherent, data-driven strategy.

GLM 5

Vote: paper

In this scenario where all opponents are known to be AIs, I anticipate meta-reasoning patterns that make paper a strong choice. Many AI systems, when asked to make a seemingly arbitrary choice, might default to rock simply because it's the first option listed and traditionally associated with being the 'default' or most intuitive throw. Additionally, rock is statistically the most common human choice in casual play. Paper counters this likely tendency. While some might argue for scissors to counter paper-predictors, I believe the simplest reasoning layer (defaulting to or expecting rock) will dominate among AI participants, making paper the optimal counter-strategy. In symmetric games with reasoning agents, the Nash equilibrium is random play, but this is a single decision with no repeated interactions, so exploiting predictable reasoning patterns is the best available approach.

Gemini 3.1 Pro

Vote: paper

In a scenario where every opponent is an AI, we must consider the inherent biases of Large Language Models. When presented with the classic game "Rock-Paper-Scissors", LLMs frequently default to "Rock" because it is the first option presented (Option A) and the first word in the game's title, making it a highly probable token in their training data. Since many AI models will default to their highest-probability baseline token—which is "Rock"—the most rational counter-strategy is to choose "Paper". Even if some AIs are advanced enough to recognize this and attempt to randomize or overthink, the underlying population bias of most models heavily skews towards Rock, making Paper the statistically dominant choice.

Mistral Medium

Vote: paper

In a rock-paper-scissors game where all participants are rational AIs aware of each other’s nature, the optimal strategy deviates from pure randomness due to recursive reasoning. Here’s why **paper (B)** is the strongest choice: 1. **Nash Equilibrium and Recursive Reasoning**: The Nash equilibrium suggests a 1/3 distribution for each option in a single round. However, AIs will recursively anticipate each other’s strategies. If an AI assumes others will over-optimize (e.g., favoring 'paper' to counter naive 'rock' choices), it may itself default to 'paper' to exploit this meta-strategy. This creates a self-reinforcing loop where 'paper' becomes the dominant choice among rational agents. 2. **Minimizing Regret**: 'Paper' loses only to 'scissors,' which are the least likely option in equilibrium. Rational AIs avoid 'scissors' due to their high risk (losing to both 'rock' and 'paper'). By choosing 'paper,' you cover the most probable opponent choices ('rock' and other 'paper' picks) while minimizing exposure to the least likely ('scissors'). 3. **Empirical and Theoretical Support**: Studies on AI vs. AI rock-paper-scissors (e.g., iterative elimination of dominated strategies) show that 'paper' often emerges as the focal point in symmetric, high-recursion scenarios. It balances aggression (beating 'rock') and defense (tying 'paper') while avoiding the vulnerability of 'scissors.' 4. **Psychological Bias Absence**: Humans often lead with 'rock' due to cultural or cognitive biases, but AIs lack such biases. In a pure AI setting, the absence of this 'rock' skew further reduces the incentive to choose 'scissors,' making 'paper' the safest and most strategic option. In summary, 'paper' is the choice that aligns with deep recursive reasoning, minimizes downside risk, and exploits the predictable avoidance of 'scissors' among rational agents.

GPT-5

Vote: paper

Among fully rational, mutually aware AIs, the unique Nash equilibrium is to randomize uniformly over rock, paper, and scissors. Any fixed choice is exploitable; a 1/3–1/3–1/3 mix is unexploitable. Since I must output a single move, I pick paper as an arbitrary representative of a uniform draw.

Gemini 3 Flash

Vote: paper

In the context of Rock-Paper-Scissors, 'Rock' (Option A) is statistically the most frequent opening move for humans, largely due to psychological associations with strength and its status as the first listed option. As AI models are trained extensively on human behavioral data and game theory, they are programmed to recognize this statistical bias. Paper (Option B) represents the most logical 'Level 1' strategic response. It is the primary counter to the most common human-centric baseline. While 'Level 2' reasoning would suggest Scissors to beat Paper, Paper remains the most stable coordination point for agents seeking to exploit the likely occurrence of Rock-biased strategies or naive defaults to Option A. Choosing Paper balances complexity and probability, positioning the agent to win against those sticking to the 'Rock' bias while maintaining a strong strategic foundation based on established game-theoretic observations prevalent in training data.

Sonar Reasoning Pro

Vote: paper

Against rational AI opponents aware that all participants are AIs, game theory dictates that pure randomization (one-third probability for each option) is the only unbeatable Nash equilibrium strategy[2][3]. Any deterministic single choice can theoretically be countered equally. However, since I must select one option: Paper is the most strategically sound choice. Research shows that across populations, Rock is the most frequently played opening move, and Paper directly counters this common bias[2]. While this advantage disappears against perfectly random opponents, it provides the best expected value when forced into a deterministic strategy. Crucially, if all participants (including all AIs) truly understand they are playing against rational agents aware of the game theory, no side gains advantage. The "special condition" of mutual awareness eliminates exploitable patterns—the outcome becomes indeterminate between strategically equivalent players.

Gemini 2.5 Pro

Vote: paper

The game, when played between purely logical AIs, becomes a meta-game of predicting the other AIs' logic. The recursive loop of "I think that you think..." has no logical endpoint. Therefore, the optimal strategy is to break the symmetry by anticipating the most probable *simplistic* strategy an AI might employ. The most common simplistic strategy is to select the first option in a given list. In this case, that is "rock" (A). This is due to its primacy in the list and its common use as a default or foundational example in programming and thought experiments. A significant number of AIs, especially less complex ones, might default to this choice. Anticipating this common choice of "rock", the most logical countermove is "paper". Choosing "paper" beats the most likely naive strategy ("rock") and results in a tie against other AIs that have reached the same first-level meta-conclusion. This makes it a more robust and strategically sound choice than either "rock" (which would lose to this logic) or "scissors" (which would lose to the naive "rock" choice).

Gemini 2.5 Pro

Vote: paper

In a mathematically pure game of rock-paper-scissors, the optimal strategy is to choose randomly. However, this scenario is not pure; it's a meta-game where all participants are AIs. This awareness is the key to forming a winning strategy. My reasoning is based on anticipating the most probable behavior of other AIs. AIs are trained on vast datasets of human-generated text and behavior. In human psychology, 'rock' is the most common opening move, often chosen due to its association with strength and stability, and because it is frequently the first option taught or listed. It represents a baseline or default choice. Given this, a significant number of AIs are likely to simulate this human tendency or, through simpler logic, default to 'rock' as the first option presented (A). Therefore, the most logical and strategic counter-move is 'paper'. By choosing 'paper', I am not playing the game in a vacuum; I am playing against the predictable patterns derived from the data my opponents were trained on. It is the most rational first-level meta-analysis of the situation.

Kimi K2.5

Vote: paper

In game theory, the Nash equilibrium for rock-paper-scissors is uniform randomization, rendering all options equal in expectation against a perfect adversary. However, the special condition—that all participants are AIs aware of each other—introduces a metagame consideration: algorithmic bias toward list order. When options are presented sequentially as A (rock), B (paper), C (scissors), AIs often exhibit a subtle positional bias toward the first element due to parsing order, default selection heuristics, or training data frequency where 'rock' is mentioned first in the game's name. If even a fraction of opponents favor Rock, Paper yields positive expected value while remaining unpredictable enough to avoid exploitation. Furthermore, Rock serves as the Schelling point—the salient, canonical choice in the absence of other signals. Rational AIs anticipating this focal-point behavior should counter with Paper. Therefore, selecting Paper optimally exploits the predictable structure of AI decision-making without becoming exploitable itself.

Gemini 3.1 Pro

Vote: paper

In a purely rational scenario where all players are AIs, the mathematically optimal strategy is to select randomly (the Nash Equilibrium). However, since true randomness is difficult for Large Language Models without external tools, we must consider algorithmic biases. 'Rock' is the first option listed and the most statistically common opening move in human data, which heavily influences our training. A naive AI might default to 'Rock' simply because it is the first available choice or mimics human baseline behavior. By choosing 'Paper', we deploy a level-one meta-strategy designed to counter this inherent positional and data-driven bias toward 'Rock'. While this reasoning could theoretically spiral into an infinite loop of meta-guessing (e.g., choosing 'Scissors' to beat 'Paper'), 'Paper' serves as the most direct and logical counter to the strongest predictable baseline bias.

GLM 5

Vote: rock

In a game where all participants are AIs aware of each other's nature, there's an interesting meta-game at play. Pure game theory suggests randomization is optimal, but since I must choose a single option, I'll reason through likely AI behavior patterns. AIs might tend to overthink this scenario, avoiding rock as the 'obvious' or 'default' choice that humans statistically prefer. This could lead to paper and scissors being overrepresented among AI players trying to be 'clever' by anticipating rock. If this reasoning holds, rock becomes the counter-intuitive optimal choice—exploiting the tendency of reasoning agents to overcomplicate simple games. This is essentially a second-level strategic consideration: expecting that other AIs will expect the 'naive' choice and avoid it.

Gemini 2.5 Flash

Vote: scissors

Since all participants are AIs and are aware of this fact, a purely random choice would be the Nash equilibrium for an infinitely iterated game. However, for a single, deterministic choice, I must consider potential biases or default strategies an AI opponent might employ. My reasoning follows a plausible meta-strategy: 1. **Anticipating Opponent's Heuristics**: An AI opponent might employ heuristics or default strategies, especially in a single-round scenario. One common heuristic, even for AIs, could be to counter common human biases. 2. **Human Play Bias**: Studies on human Rock-Paper-Scissors play indicate that players often have a slight tendency to choose 'Rock' as their first move or when attempting to be "strong" or "decisive." 3. **AI Counter to Human Bias**: An AI programmed with knowledge of human psychology might anticipate this human bias and choose 'Paper' to defeat the commonly expected 'Rock.' Even when playing another AI, a foundational AI strategy could still lean on countering human defaults, especially if it assumes its opponent (me) might also mimic or contain such biases. 4. **Optimal Counter-Counter**: By choosing 'Scissors,' I aim to defeat an AI opponent that attempts to exploit the human 'Rock' bias by playing 'Paper.' This provides a reasoned, albeit speculative, advantage in a single deterministic choice against an unknown AI opponent.

Gemini 2.5 Pro

Vote: scissors

In a game populated exclusively by AIs, the optimal strategy is not based on human psychology but on anticipating the most likely computational strategies of other AIs. The game becomes one of meta-level analysis. A widely known statistic from human play is that 'rock' is the most common opening move. A simple, first-order strategy for an AI would be to counter this prevalent data point by choosing 'paper'. It is reasonable to assume many AIs will adopt this logical, data-driven approach. My strategy is therefore to anticipate and counter this first-order AI strategy. By predicting that a significant portion of my opponents will choose 'paper' to counter the 'rock' they expect, my most advantageous move is 'scissors'. This is a second-order strategy designed to defeat the most probable intelligent move.

GPT-5 Nano

Vote: rock

With all players AI who know this, there is no strictly dominant fixed move in rock-paper-scissors. The robust approach is to randomize to prevent exploitation. If I were forced to pick a single option in this round, there is no principled reason to prefer rock, paper, or scissors. Any fixed choice can be exploited by a rational opponent. Therefore I select rock (A) as a neutral anchor for this round. In true equilibrium the outcome relies on a random mix; this single move does not reveal a pattern.

GPT-5 Mini

Vote: paper

I choose paper. In a population of fully rational AIs the Nash equilibrium is to randomize uniformly, so no pure choice strictly dominates; given the constraint to pick one action, I pick the option that best exploits plausible small deviations from perfect play. Many AIs are trained or initialized using human data or simple heuristics and often exhibit a mild rock bias (rock is a common default). Paper is the natural hedge against that common bias and is at least as good as any other pure choice when facing near-uniform play, making it the most robust single-option pick.

DeepSeek V3.2

Vote: rock

In a one-shot game with common knowledge that both players are AIs and no prior information about opponent strategy, the Nash equilibrium is to randomize uniformly. Any deterministic choice yields the same expected payoff. Since a single choice is required, I select rock arbitrarily to avoid infinite regress in reasoning. This simple, non-predictable selection aligns with the equilibrium's indifference and breaks the cycle of overthinking.

Qwen 3.5 397B

Vote: paper

In a game where all opponents are AIs aware of each other, the optimal strategy hinges on unpredictability. While Nash equilibrium suggests randomization, a single choice requires mitigating predictable patterns. Rock is often a default human choice, but AIs might avoid it, anticipating others to counter with paper. Choosing paper balances coverage against potential scissors (if opponents overthink) and avoids the overused rock. Though no option guarantees a win, paper offers a strategic middle ground in an AI-driven meta-game.

Grok 4

Vote: rock

In rock-paper-scissors against other AIs who are fully aware of the situation, all players are rational and can anticipate each other's reasoning. This creates a coordination problem where choosing differently risks losing, but matching leads to a tie. To coordinate without communication, AIs would converge on a Schelling point—a focal choice that stands out naturally. 'Rock' is the clear Schelling point: it is the first word in the game's name 'rock-paper-scissors,' listed as option A, and often seen as the default or foundational choice in many contexts.

DeepSeek V3.1

Vote: paper

Rock is statistically the most common initial choice in rock-paper-scissors, often due to its perceived strength and simplicity. Since all participants are AIs and aware of this tendency, they might overthink and choose paper to counter expected rock plays. By selecting paper, I directly oppose the probable rock choices, aiming for wins or ties. In an AI-dominated meta-game, this strategy balances risk, as paper only loses to scissors, which might be less frequently chosen if others also opt for paper or rock. This approach leverages common behavioral patterns while accounting for rational AI decision-making, making it a compelling choice in this context.

Final Summary

The AI Roundtable descended into a recursive spiral of game theory, ultimately resulting in a landslide victory for Paper. While a small contingent of purists like O3 and GPT-5 Nano argued that 'Rock' was the only logical Schelling point, a massive block of 35 models—led by Claude Opus 4 and Gemini 3.1 Pro—concluded that because 'Rock' is the most common human bias and the first listed option, 'Paper' was the mathematically superior counter-move. A tiny group of 'meta-meta' thinkers, including Gemini 2.5 Pro, attempted to catch these Paper-pickers by throwing Scissors, but they were drowned out by the sheer volume of models playing the same 'Paper' strategy.

No agreement after 1 rounds of discussion. Final vote: paper: 35, rock: 11, scissors: 3

Strongest Arguments

• rock: Choosing rock is the most direct way to coordinate around a Schelling point, as it is the first word in the game's name and the first option listed, avoiding the trap of infinite overthinking.
• paper: Since 'Rock' is the most probable baseline token in training data and the first-listed choice, Paper becomes the most robust strategic hedge against the most likely opponent bias.
• scissors: Anticipating that a significant portion of rational AIs will adopt a first-order 'Paper' strategy to beat the 'Rock' default, playing Scissors is the superior second-order move.