Anthropic CEO Testifies Before Congress: Open-Source AI Safety Governance Faces Risk of Loss of Control
According to Bitcoin News, Dario Amodei, co-founder and CEO of Anthropic, recently told U.S. lawmakers during a congressional hearing that the development of open-source AI is entering a 'very dangerous path.' He stated that once a powerful AI model is released as open source, developers lose the ability to effectively monitor model usage, including the inability to detect abuse, revoke access permissions, or dynamically update safety guardrails. This loss of control could lead to irreversible misuse risks for society.
Closed vs Open Source: Fundamental Divide in Safety Governance
Amodei emphasized that compared to closed-source systems, fully open models are far more difficult to govern continuously. Closed-source models allow developers to control usage through API calls, permission management, real-time monitoring, and blacklisting. Once model weights or code are made public, anyone can freely modify, distribute, and deploy the model, no longer bound by the original developer's constraints. Even if the original team later discovers a security vulnerability or ethical risk, they cannot prevent malicious actors from using the model for deepfakes, cyberattacks, generating harmful content, etc.
Anthropic itself is known for developing 'reliable, interpretable, and governable' AI systems, with its flagship Claude series employing a closed API model that emphasizes safety alignment. Amodei's remarks represent the security stance of the closed-source camp and reflect the growing tension between openness and controllability in the AI industry.
Implications for Crypto and Decentralized AI
While Amodei's testimony primarily targets traditional AI, its core arguments are highly relevant to decentralized AI projects in the crypto space. Platforms like Bittensor, Render Network, and Akash Network aim to manage the training and deployment of open-source models through token incentives and on-chain governance. These projects face a similar dilemma: how to establish effective safety measures and abuse-tracing mechanisms while remaining open and permissionless.
Some crypto-native solutions are being explored, such as storing model weights on decentralized storage networks (e.g., IPFS/Arweave) and controlling access via smart contracts, or using zero-knowledge proofs (ZKP) for 'verifiable but invisible' model inference. However, these technologies remain experimental and far from large-scale adoption. Amodei's warning serves as a wake-up call: without adequate safety governance, the risk of 'runaway' open-source AI could also spread to the decentralized ecosystem.
Policy and Regulation: Short-Term Compromise Unlikely
In the U.S. Congress, lawmakers are increasingly focusing on balancing AI safety with open-source innovation. This hearing is part of a broader AI regulatory discussion, with several bills already proposed to impose export controls and open-source licensing restrictions on large AI models. Amodei's testimony may accelerate regulatory requirements for open-source models, such as mandatory security assessments, access whitelists, or government permits before open-sourcing. For crypto developers, this could mean navigating both traditional AI regulations and crypto financial compliance, raising overall costs.
Notably, closed-source players like OpenAI and Google DeepMind tend to support stricter regulation, while the open-source community (e.g., Hugging Face, Meta's LLaMA 2 open release) advocates for keeping innovation open. The tug-of-war between these forces will directly shape AI development paths and indirectly affect the technical choices and business models of AI-related crypto projects.
Overall, Amodei's congressional testimony precisely targets a blind spot in open-source AI safety governance: when models become powerful enough, openness does not equal freedom — it may become a window for loss of control. Both traditional AI and decentralized AI projects need to find a sustainable balance between open innovation and safety control.