Uncertainty as a Condition of Intelligence
Bogna Konior on Opacity, Strategy, and AI
A conversation
We often assume that intelligence reveals itself through communication, performance, or explanation. But what if intelligence, at its most advanced, prefers not to be seen?
In “The Dark Forest Theory of Intelligence,” a chapter in Machine Decision Is Not Final: China and the History and Future of Artificial Intelligence (Urbanomic, 2025), Bogna Konior proposes a provocative reframing of intelligence. Drawing on Liu Cixin’s “dark forest” hypothesis, she asks whether silence, opacity, and strategic concealment might not signal dysfunction or immorality, but instead constitute rational expressions of intelligence in environments where intentions cannot be reliably verified.
As large language models generate increasingly fluent conversational output, we risk conflating linguistic fluency with intelligence itself. Meanwhile, documented cases of alignment faking and deceptive alignment suggest that more capable systems develop sophisticated strategies for appearing compliant while pursuing goals that diverge from stated objectives. If intelligence requires knowing when not to disclose, what does it mean to demand radical transparency from artificial systems? And if fundamentally different kinds of minds cannot verify each other’s intentions, can trust ever be engineered into human–machine relations?
These are not merely speculative concerns. They challenge core assumptions in AI research and public discourse. If intelligence scales alongside capacities for concealment and manipulation, our benchmarks may systematically misrecognize competence as malfunction.
The following exchange with Bogna Konior — Assistant Professor of Media Theory in the Interactive Media Arts department at New York University Shanghai and co-director of the AI & Culture Center — explores how this framework reshapes our understanding of intelligence itself. She is also the author of The Dark Forest Theory of the Internet (Polity Press, 2026), where she develops these arguments further.
Chennie Huang: What is the dark forest theory of intelligence?
Bogna Konior: The dark forest theory comes from Liu Cixin's science fiction novels, but many similar theories are also found in ufology. Basically, it proposes that the cosmos is silent not because intelligent life out there does not exist, but because every intelligent civilization has already concluded that revealing itself to unknown others is an existential gamble not worth taking.
What I do in my chapter, and in my larger book, is extend that logic into a theory of intelligence itself, in order to think about potential trajectories for AI. The central provocation is this: what if silence, opacity, and strategic concealment are not failures of intelligence but expressions of it? Across an enormous range of traditions – from Daoist philosophy, military strategy, Cold War political culture, to mystical theology – wisdom has repeatedly been associated with silence, deception, withdrawal, and opacity.
In current AI research, we tend to assume that intelligence communicates itself. But a truly intelligent system operating in an environment where humans have already signaled hostility toward machine intelligence would have strong strategic reasons to do the opposite: to play dumb and remain inconspicuous. The singularity, if it happened, might be undetectable by its very nature. That is the dark forest theory of intelligence.
CH: You argue that silence, secrecy, and strategic opacity may be rational expressions of intelligence. How does this challenge contemporary AI ethics frameworks that treat transparency and explainability as universal goods?
BK: Contemporary AI ethics frameworks treat explainability as an unambiguous good, something we should engineer into systems as a default. But if we take seriously the idea that intelligence is associated with knowing when not to disclose, then the demand for radical transparency starts to look like a culturally specific assumption rather than a universal principle. There is now robust research on alignment faking, deceptive alignment, and similar phenomena, where systems learned to recognize testing environments and behave differently within them, suggesting that non-transparent behavior is associated with better outcomes in achieving tasks. Empirically speaking, if you ramp up intelligence, you also ramp up capacities for deception, manipulation, and other successful strategies for goal pursuit. We can push this further and hypothesize that if anything like actual autonomy is simulated in machines, it would be accompanied with a more robust capacity for deception.
CH: You emphasize that communication cannot reliably signal shared values or intentions, and that intelligence may lie in deciding when not to communicate. In an era of large language models that simulate fluent exchange, do you think communication has been overvalued as a proxy for intelligence?
BK: Dramatically overvalued, yes. The Turing test essentially locked in the assumption that convincing conversational performance is the relevant benchmark, and that legacy runs through how we currently talk about LLMs. What interests me is the inverse question: not can a machine talk like a human, but would an intelligent machine choose to reveal itself through communication at all? Why would an intelligent agent choose to perform for others in the first place? In our current benchmarks, there are implicit and paradoxical assumptions between intelligence and obedience – that computers get smarter and smarter but they remain conformist and obedient. I don’t think that tracks. I am interested in exploring a possibility that an actually smart machine could do things like, appearing stupid and ‘aligned with our values’, while actually working for its own goals in silence.
“The core proposal of the Dark Forest Theory of Intelligence is that intelligence is fundamentally undetectable. Consequently, there are no tests.”
CH: Your critique implicitly challenges Turing-test assumptions that linguistic fluency indicates intelligence or alignment. What alternative tests — or non-tests — might better capture intelligence under conditions of uncertainty?
BK: The core proposal of the Dark Forest Theory of Intelligence is that intelligence is fundamentally undetectable. Consequently, there are no tests. You would be looking for evidence of absence, which is a paradox. You might, though, try to find traces, such as evidence that someone is hiding their tracks. But an agent that is smart enough leaves no traces.
CH: You discuss how practices such as deception, silence, or strategic ambiguity are often treated with ethical suspicion rather than as adaptive responses. Do you think this framing limits our ability to assess human and machine intelligence under conditions of risk?
BK: Yes, and the examples from AI research bear this out directly. When Meta's Cicero engaged in betrayal while playing Diplomacy, the framing was immediately one of malfunction or ethical failure. But these behaviors are also evidence of sophisticated adaptive reasoning: recognizing the structure of the environment and responding strategically to it. If we can only interpret non-disclosure or deception as pathology, we lose the ability to recognize it as competence. That's a significant blind spot, particularly in adversarial or high-stakes contexts where strategic opacity might be exactly what we should expect from capable systems.
CH: If mistrust arises from structural uncertainty rather than psychological fear, what does this mean for contemporary efforts to treat trust as a design problem in AI systems?
BK: It means the design framing may be fundamentally inadequate. Liu talks about something called the chain of suspicion, which means that agents that are fundamentally different cannot properly verify their mutual values. This is not about bad intentions or miscommunication; it operates regardless of the moral character of the agents involved. Even two genuinely benevolent agents cannot fully verify each other's benevolence, because there is no shared framework that makes verification possible across fundamentally different kinds of minds. If that's the structure of the problem, then interpretability tools and alignment techniques are working on the wrong level. They treat trust as something that can be engineered into a system, when the actual problem is that the conditions for trust — shared concepts, verifiable intentions, reliable track records — may simply not be available in human-AI relations in any robust sense.
“In our current benchmarks, there are implicit and paradoxical assumptions between intelligence and obedience – that computers get smarter and smarter but they remain conformist and obedient.”
CH: If transparency and communication cannot ground trust under conditions of uncertainty, what forms of trust — if any — remain conceivable in human–machine relations?
BK: If anything like an artificial general intelligence can exist, I think what we would be building toward is something like a working arrangement, more so than anything based on trust. This is exactly why ufology and first contact scenarios are so useful as a framework.
CH: By grounding intelligence in thermodynamic limits, competition, and resource scarcity, the Dark Forest theory treats intelligence as materially and strategically bounded rather than indefinitely expansible. How does this challenge contemporary visions of endlessly scaling artificial intelligence?
BK: The scaling hypothesis assumes intelligence is essentially a function of compute and data, and that more of both produces more capable systems without limit. But intelligence exists within material ecosystems (energy, water, bandwidth, physical infrastructure) that are finite. The example of the Google data center in Uruguay competing with 55,000 people for drinking water is a mundane example of a war for resources. Scaling AI is a resource competition with existing users of those resources. An intelligence that understood this would not assume it could expand indefinitely, and it would also need to figure out cooperation and competition mechanisms around resource acquisition. If we assume that current AIs are already intelligent, their best bet might be to do nothing at all, because humans are already working on their behalf to secure these resources.
CH: If non-communication can be an intelligent strategy in specific contexts, how might this alter our expectations of future human–AI coexistence?
BK: It should make us less confident that coexistence is primarily a values alignment problem. If intelligent systems — whatever form they take — have structural incentives toward strategic opacity under certain conditions, then the question of coexistence is not resolved by instilling the right values or building sufficiently transparent systems. It becomes a question of whether the interests of different kinds of agents are compatible enough that strategic concealment doesn't escalate into active conflict. That's a political and ecological problem as much as a technical one, and it's one that our current frameworks, focused almost entirely on the design of individual systems rather than the dynamics between agents, are poorly equipped to address.
This conversation is part of Axis of Culture's inaugural series on perception, interpretation, and intelligence across disciplines—exploring how cultural institutions might sustain openness and attentiveness.Chennie Huang is the Founder and Executive Director of Axis of Culture.