The following is a guest post and opinion from Felix Xu, Founder of ARPA Network.

Walk into any of Cloudflare’s global offices and you’ll find some unusual decor. In San Francisco, it’s floor-to-ceiling lava lamps, known as “the wall of entropy.” In London, it’s the “unpredictable pendulums.” These aren’t just pretty backdrops—they are grist for the randomness mill, exemplifying the ongoing creative and engineering race for true randomness.

Randomness is the unsung hero of the modern internet—the cornerstone of encryption, the backbone of fair gaming systems, and increasingly, a critical component in AI verification. Yet, as we hurtle toward a future where crypto represents a growing share of the global economy and AI agents gain greater autonomy—particularly over financial operations—the integrity of randomness becomes not just a technical concern but an existential one.

The Myth of Perfect Randomness

Computer scientists have long sought an idealized version of randomness, embodied by the theoretical “random oracle,” a hypothetical black box providing truly unpredictable outputs for every input. Unfortunately, perfect randomness is practically unattainable. Instead, digital systems rely on pseudorandom functions—sophisticated algorithms designed to simulate randomness convincingly. Physical entropy sources, such as Cloudflare’s “wall of entropy” lava lamps or London’s unpredictable pendulums, serve as essential real-world seeds for these pseudorandom functions, introducing genuine unpredictability from natural phenomena into cryptographic processes.

Yet, this blend of physical entropy and pseudorandom algorithms isn’t foolproof. As MIT computer science professor Steve Ward points out, knowing an algorithm and its initial seed can enable prediction of supposedly random outcomes—such as the next card dealt in online poker. Such vulnerabilities underscore the critical importance of genuinely unpredictable and verifiable randomness in technology-dependent contexts, from digital gaming to financial security.

Verifiable Randomness in AI

Artificial Intelligence systems increasingly rely on randomness to ensure fair, unbiased, and robust outcomes, playing an indispensable role across diverse applications—from healthcare diagnostics to financial decision-making. Randomness helps AI models avoid overfitting by introducing necessary variability into training processes, making predictions and decisions more adaptable and reflective of real-world scenarios. However, when randomness is not verifiable, it becomes impossible to ensure that AI-generated outcomes are genuinely impartial and resistant to hidden biases.

Take, for example, AI-driven financial trading algorithms. These systems utilize randomness to explore various investment scenarios and prevent predictable exploitation. However, without transparent and verifiable randomness, financial institutions and regulators cannot confirm that the model’s decisions are truly unbiased. The integrity of randomness directly influences market fairness, as hidden biases could disproportionately favor certain assets or market participants.

Similarly, in generative AI applications such as language models, image generation, or personalized recommendations, the randomness parameter—known as “temperature”—profoundly affects output quality. A low temperature generates consistent yet potentially repetitive outputs, while higher temperatures introduce greater variability but risk unpredictable and unreliable responses. Without verifiable randomness, users and stakeholders cannot confirm that the model’s claimed temperature settings accurately reflect actual operational conditions—creating opportunities for covert biases and subtle manipulation.

Moreover, AI systems used in critical decision-making scenarios—such as autonomous driving or medical diagnosis—must reliably produce unbiased, equitable outcomes. Verifiable randomness ensures stakeholders can independently verify these systems’ fairness, significantly boosting public trust and regulatory compliance.

The Importance of Verifiable Randomness Functions (VRFs)

At its core, verifiable randomness addresses a fundamental paradox in our digital systems: how do we create unpredictability that can nevertheless be trusted? Verifiable randomness rests upon three pillars:

1. Unpredictability: The generated values cannot be anticipated in advance, even by the system’s creators or operators.
2. Bias-resistance: The output distribution contains no detectable patterns or skews that could be exploited.
3. Public verifiability: Anyone can independently confirm that the random values were generated according to the specified protocol, without requiring access to secret information.

Public verifiability distinguishes verifiable randomness from traditional random number generation. It creates a bridge between the necessary chaos of true randomness and the transparency required for trust. It may sound dramatic, but without that trust, insiders could rig AI algorithms and crypto platforms in their favor—plunging our civilization into a technological dark age.

Toward Decentralized Verifiable Randomness

Likewise, in blockchain systems, randomness underpins critical functions including validator selection, transaction ordering, and token distribution. When Ethereum selects validators for block production or when NFT platforms determine rare trait distribution, randomness decides outcomes worth billions. Any manipulation could allow malicious actors to gain unfair advantages, potentially undermining the entire industry.

The “stakes” (no pun intended) are particularly high in Proof-of-Stake systems, where validators are selected probabilistically. If validator selection becomes predictable or manipulable, attackers could corrupt the network by front-running blocks. Similarly, in DeFi, predictable randomness enables flash loan attacks and market manipulations that drain liquidity pools.

The ideal solution combines true randomness with decentralized verification—distributed networks generating collective randomness where no single entity controls the outcome, yet everyone can verify its integrity.

The Path Forward: Embracing Verifiable Randomness

As we entrust AI systems with greater autonomy and responsibility, verifiable randomness becomes not just a technical challenge but a foundational requirement for trusted systems. Organizations developing AI agents must prioritize verifiable randomness before shipping code—or risk catastrophe.

The technology industry stands at a crossroads. We can continue building AI systems and offloading more capital to crypto-systems on the shaky foundation of conventional randomness and opaque decision processes—or we can embrace verifiable randomness as part of a broader commitment to transparency and trust.

Without verifiable randomness, we’re building our future on digital quicksand. With it, we have the conditions for a flourishing future—where our technological superpowers work with us, not against us.

The following is a guest post from Felix Xu, Founder of ARPA Network.

The U.S. government’s approach to artificial intelligence (AI) has shifted dramatically, emphasizing accelerated innovation over regulatory oversight. In particular, President Donald Trump’s executive order, Removing Barriers to American Leadership in Artificial Intelligence, has set a new tone for AI development, one rooted in promoting free speech and advancing technological progress. Similarly, U.S. Vice President JD Vance’s refusal to endorse a global AI safety agreement signals that America will prioritize innovation without compromising on its competitive advantage.

However, as AI systems increasingly become more influential in financial markets, critical infrastructure, and public discourse, the question remains: How can we ensure trust and reliability in AI model-driven decisions and outputs without stifling innovation?

This is where Verifiable AI comes in, offering a transparent, cryptographically secure approach to AI that ensures accountability without heavy-handed regulation.

The Challenge of AI Without Transparency

AI’s rapid advancement has ushered in a new era of intelligent AI agents capable of complex and autonomous decision-making. But without transparency, these systems can become unpredictable and unaccountable.

For instance, financial AI agents, which rely on sophisticated machine learning models to analyze vast datasets, are now operating under fewer disclosure requirements. While this encourages innovation, it also raises a trust gap: without insight into how these AI agents reach their conclusions, companies and users may struggle to verify their accuracy and reliability.

A market crash triggered by an AI model’s flawed decision-making is not just a theoretical risk, it’s a possibility if AI models are deployed without verifiable safeguards. The challenge is not about slowing down AI progress but ensuring that its outputs can be proven, validated, and trusted.

As renowned Harvard psychologist B.F. Skinner once said, “The real problem is not whether machines think but whether men do.” In AI, the key issue is not just how intelligent these systems are, but how humans can verify and trust their intelligence.

How Verifiable AI Bridges the Trust Gap

Russel Wald, executive director at the Stanford Institute for Human-Centered Artificial Intelligence, sums up the U.S. AI approach:

“Safety is not going to be the primary focus, but instead, it’s going to be accelerated innovation and the belief that technology is an opportunity.”

This is precisely why Verifiable AI is crucial. It enables AI innovation without compromising trust, ensuring AI outputs can be validated in a decentralized and privacy-preserving way.

Verifiable AI leverages cryptographic techniques like Zero-Knowledge Proofs (ZKPs) and Zero-Knowledge Machine Learning (ZKML) to provide users with confidence in AI decisions without exposing proprietary data.

• ZKPs allow AI systems to generate cryptographic proofs that confirm an output is legitimate without revealing the underlying data or processes. This ensures integrity even in an environment with minimal regulatory oversight.
• ZKML brings verifiable AI models on-chain, allowing for trustless AI outputs that are mathematically provable. This is particularly critical for AI oracles and data-driven decision-making in industries like finance, healthcare, and governance.
• ZK-SNARKs convert AI computations into verifiable proofs, ensuring AI models operate securely while protecting IP rights and user privacy.

In essence, Verifiable AI provides an independent verification layer, ensuring that AI systems remain transparent, accountable, and probably accurate.

Verifiable AI: The Future of AI Accountability

America’s AI trajectory is set for aggressive innovation. But rather than relying solely on government oversight, the industry must champion technological solutions that ensure both progress and trust.

Some companies may take advantage of looser AI regulations to launch products without adequate safety checks. However, Verifiable AI offers a powerful alternative empowering organizations and individuals to build AI systems that are provable, reliable, and resistant to misuse.

In a world where AI is making increasingly consequential decisions, the solution is not to slow down progress, it’s to make AI verifiable. That’s the key to ensuring AI remains a force for innovation, trust, and long-term global impact.

Mentioned in this article