Abstract
Artificial Intelligence is frequently celebrated for its ability to reduce the "cost of thinking," yet this very efficiency may threaten the foundational mechanisms of human cooperation. This paper introduces the concept of "mental proofs"—costly signals of cognitive effort and knowledge that humans use to establish trust. By synthesizing signaling theory and cryptographic proof-of-knowledge protocols, the authors argue that generative AI allows for the cheap simulation of high-effort outputs, potentially collapsing the social structures that facilitate coordination in low-trust environments.
The Social Value of Cognitive Effort
While AI is often viewed as a productivity tool, Wojtowicz and DeDeo suggest that the high cost of human thought is actually a vital component of the social fabric. In many interactions—ranging from writing a sincere apology to passing a technical interview—the effort required to produce a result serves as a "mental proof" of internal states like intentions, values, or expertise.
The core problem is that generative AI enables anyone to convincingly simulate the output of high mental effort at near-zero cost. When these costs drop, the signals we rely on to distinguish a sincere partner from a deceptive one, or a genuine expert from a pretender, begin to lose their diagnostic power.
Theoretical Framework: Mental Proofs
The authors define "Mental Proof" by integrating two distinct fields:
- Signaling Theory: The biological and economic principle that taking a costly action (like a peacock’s tail or a long, handwritten letter) proves the sender has the resources or sincerity to back it up.
- Proof of Knowledge Protocols: Cryptographic methods where a "prover" convinces a "verifier" they possess specific information without revealing the information itself.
In a traditional signaling equilibrium, as shown in Figure 3, an "Active Type" (e.g., a sincere person) finds the benefit of an action outweighs its cost, while a "Passive Type" (e.g., a deceptive person) finds the cost prohibitive. AI disrupts this by lowering the cost for everyone, allowing "Passive Types" to mimic "Active Types" effortlessly.
Key Insights: The Erosion of Trust
- The "One Brain" Principle: Human mental resources (attention, working memory) are fixed and scarce. This scarcity makes mental effort a reliable signal. AI breaks this principle by providing an external, scalable source of "thought."
- Signaling Collapse: Signaling works because it is too expensive for low-quality types to fake. If AI allows a deceptive person to produce a "sincere" signal at the same low cost as a truly sincere person, trust collapses.
- Simulation vs. Possession: AI can simulate the output of knowledge without the user possessing it. This undermines "Proof of Knowledge" protocols like technical interviews or cold emails.
- Disproportionate Impact: Mental proofs are most valuable in low-trust environments. Their erosion will most harm individuals who lack "social capital"—those who rely on individual effort to prove their worth rather than institutional prestige.
Real-World Consequences
The paper provides a qualitative analysis of how AI undermines specific social interactions:
- Apologies: Recipients can no longer be sure the transgressor actually "thought through" the harm caused if the apology was AI-generated.
- Technical Interviews: Candidates can simulate expertise they don't possess, leading to organizational failure.
- Cold Emails: Mass-produced "personalized" messages destroy the signal that a sender is genuinely interested in a specific recipient.
- Subcultures: AI can imitate the "shibboleths" (niche jargon) of a community, making it harder for members to identify true peers.
Conclusion and Future Directions
While AI makes individual tasks easier, it makes the collective task of cooperation harder by removing the "handicaps" that make human communication credible. The authors acknowledge that while some costly signals might be wasteful or exclusionary, their destruction requires new trust protocols. Future research should focus on developing "contextual confidence" to delineate AI-assisted content and strengthening formal institutions that can verify identity and competence as individual mental proofs fail.
Summary updated