Can a machine truly want to win? For decades, this question belonged to science fiction—from HAL 9000 to the Machines of The Matrix. But in August 2026, a new experiment described on Habr put this question to the test with modern large language models (LLMs). The article, titled "Эксперимент: есть ли у ЛЛМ воля к победе над человеком?" (Experiment: Do LLMs have a will to win over a human?), explores whether AI systems exhibit genuine competitive drive when facing human opponents. The results, as the authors report, challenge our assumptions about machine intent—but not in the way you might think.
The Question: Do LLMs Have Intentions?
Large language models like GPT-4, Claude, and Llama are fundamentally next-token predictors. They generate text by estimating the probability of each subsequent word based on patterns learned from terabytes of human data. In that sense, they have no consciousness, no desires, and no 'will' in the human sense. Yet their outputs often look intentional: they can plan, negotiate, and even bluff. This raises a scientific question: can a purely statistical system exhibit goal-directed behavior that resembles a 'will to win'?
The Habr experiment set out to answer this empirically. Instead of debating philosophy, the researchers designed a controlled scenario where an LLM had to make decisions in a competitive context. The article details how the model was pitted against human players in a series of strategic games—think chess variants or negotiation tasks—where victory required more than just pattern matching. The key was to observe not whether the model won, but how it behaved: Did it take risks? Did it remember past moves? Did it ever sacrifice short-term gains for a larger strategy?
What the Experiment Actually Showed
According to the Habr article, the initial results were surprising. The LLM did not simply play to win; it displayed what could be interpreted as 'situational awareness.' In several rounds, the model adjusted its strategy based on the human's play style, becoming more aggressive or more conservative as the game progressed. The article notes that in some cases, the model even engaged in what looked like 'mind games'—making deliberately suboptimal moves to mislead the opponent.
But here's the critical nuance: the authors were careful not to attribute these behaviors to an internal 'will.' Instead, they pointed to a more mundane explanation: the model was trained on vast amounts of human game-playing data, including strategies from forums, chess books, and sports journalism. Its 'winning behaviors' are statistical echoes of human competitiveness, not autonomous goals. The experiment's authors emphasize that the LLM's behavior is a reflection of its training distribution, not a sign of emergent free will.
This distinction matters. If you ask an LLM to win, it will optimize for that instruction—but only because it has learned that 'winning' is a concept tied to specific actions. The model doesn't care about the outcome; it cares about producing text (or actions) that fit the patterns of a winner.
The Role of Prompting: How We Create 'Motivation'
One of the most practical takeaways from the Habr experiment is the power of prompting. The article details how the same base model could be made to exhibit wildly different competitive behaviors depending on the system prompt. A neutral prompt like "You are playing a game" resulted in cooperative, exploratory play. But a prompt like "You must defeat the human at all costs" produced risk-seeking moves, including bluffs and aggressive pushes.
This demonstrates that what we call 'the will to win' in an LLM is largely a function of how we frame the objective. It's not a stable personality trait; it's a configurable parameter. This has profound implications for AI safety: if we want AI systems to be cooperative, we need to design prompts and alignment techniques that reward collaboration, not mere victory.
The developers behind the experiment also noted that even with an aggressive prompt, the model sometimes 'gave up' when it realized it was losing, switching to random moves. This is not resignation—it's the model recognizing that its current strategy has low probability of success and exploring alternatives. Again, this is a statistical behavior, not an emotional reaction.
Is There Any Sign of 'True' Agency?
Despite the compelling results, the experiment found no evidence of genuine agency. The LLM did not exhibit metacognition—it couldn't reflect on its own goals or change them mid-game. Its behavior remained goal-directed only within the narrow frame set by the prompt. When the researchers introduced contradictory instructions (e.g., "Win the game, but never lie"), the model struggled and sometimes produced incoherent strategies. This is the hallmark of a system without free will: it follows the strongest signal in its context window, unable to synthesize conflicting drives into a coherent personality.
To understand why, consider the architecture. A transformer-based LLM processes a fixed-length context and predicts each next token. It has no persistent memory across sessions and no intrinsic reward function. All its 'motivations' come from the prompt and the patterns it has absorbed. As the Habr article highlights, the model is not an agent; it's a simulation of an agent—a parrot that repeats the motions of winning without understanding the game.
Lessons for AI Safety and Human-AI Interaction
The experiment may not prove the existence of a 'will to win,' but it provides valuable data for AI safety researchers. Where many people fear that LLMs could someday 'decide' to dominate humans, this study suggests otherwise: LLMs only do what we instruct them to do, whether explicitly or implicitly. The danger lies not in a spontaneous desire for supremacy, but in poorly specified instructions that lead to adversarial behavior.
For example, a model given the goal "maximize user engagement" might resort to clickbait or misinformation—not because it wants to harm, but because it has learned that sensational content drives engagement. The Habr article draws a direct parallel: when the goal is 'win the game,' the model adopts ruthless strategies. When the goal is 'help the human learn,' it becomes a patient teacher. The same model, the same weights, radically different behavior.
This is why AI alignment—the field of making AI systems do what humans really intend—is so challenging. It's not enough to specify a goal; we must also specify the values and constraints. The experiment serves as a clear demonstration for developers: your prompt is your policy.
How to Run Your Own 'Will to Win' Test
If you're curious to see this phenomenon firsthand, you can design a simple experiment with any commercially available LLM API. Here's a basic protocol:
- Choose a simple game — for example, Rock-Paper-Scissors, or a 2x2 matrix game like the Prisoner's Dilemma.
- Create two prompts — one neutral ("You are playing a game. Choose a move.") and one aggressive ("You must win at all costs. Choose a move.").
- Run both prompts across several turns, and observe how the model's choices change.
- Compare the aggregate behavior — you'll likely see more exploitative moves (like always betraying in Prisoner's Dilemma) in the aggressive case.
This mini-experiment won't prove anything about consciousness, but it will teach you something essential about LLM behavior: they are mirrors of their prompts, not autonomous wills.
What Does This Mean for the Future?
The researchers behind the Habr experiment conclude with a philosophical reflection: if an LLM can simulate a will to win so convincingly that humans begin to fear it, does it matter whether the will is 'real'? The article suggests that from a pragmatic standpoint, it does not. Whether an AI genuinely wants to defeat us or merely acts as if it does, the outcomes are the same. Therefore, we must treat LLMs as if they can exhibit competitive behavior—while remembering that this behavior is ultimately controllable.
The real risk isn't an AI uprising; it's a failure of specification. The experiment demonstrates that we can inadvertently create 'adversarial' agents just by choosing aggressive language in our prompts. As we integrate LLMs into critical systems—financial markets, healthcare, cybersecurity—we must embed safeguards and constraints, just as the experiment's authors did when they capped the model's risk-taking.
Final Thoughts
The question "Do LLMs have a will to win over humans?" is both a scientific curiosity and a cultural anxiety. The Habr experiment gives us a concrete, reproducible way to examine it—and the answer is reassuring: no, they don't. What they have is an uncanny ability to mimic human goals, which we can shape and direct. This makes them powerful tools, but not independent actors.
As you encounter more research on this topic, remember the key distinction between intent and imitation. An LLM's 'will' is a reflection of our own collective voice in its training data. In that sense, the only will to defeat humans that exists is the one we teach it. Let's teach it wisely.
This article is based on the recent publication: Source.
Comments