Atoma CEO Hisham Khan: ‘Web3 Innovation Could Be Accelerated by AI Privacy’

• Hisham Khan, CEO of Web3-focused AI privacy platform Atoma, believes that AI can achieve full privacy and verifiability through decentralization.
• AI traditionally operates in centralized environments, needing vast amounts of data and content to be trained.
• If success is found in Web3 AI privacy, it could encourage adoption across traditional industries that need private AI processing.

In the world of Web3, privacy stands as a crucial cornerstone, reflecting its foundational principles of decentralization and user control.

But as developers look to bring more AI systems into the ecosystem, the foundations of privacy begin to get more complicated.

AI traditionally operates in centralized environments, needing vast amounts of data and content to be trained. This directly conflicts with Web3’s ethos, where users demand a complex balancing act of both privacy and transparency.

Hisham Khan, CEO of Web3-focused AI privacy platform Atoma, believes focusing on AI privacy will accelerate Web3 innovation by enabling capabilities that traditional AI cannot achieve.

The Need for AI Privacy in Web3

AI is increasingly becoming a crucial part of Web3 applications. AI agents are used for everything from analyzing market data in decentralized finance systems to chatbots guiding new users through an ecosystem.

“Without privacy assurances, these applications risk exposing user data or compromising decision-making integrity,” Khan told CCN in an interview.

Khan explains that data ownership is an integral part of Web3 as it challenges traditional practices from companies that harvest user data from digital footprints, buying habits, and AI interactions:

“Companies like Reddit, Stack Overflow, and social media giants often sell or auto-opt user data into AI model training without explicit consent, creating products based on our digital profiles and exploiting users in the process.”

In Web3, users can take control of their digital identities and turn them into private AI assistants that can generate personal insights and tailored recommendations.

Ensuring AI datasets maintain integrity while preserving privacy during inference, the part where an AI model makes predictions, remains challenging.

“Traditional systems protect data at rest and in transit, but inference often remains vulnerable,” Khan said. “This gap undermines trust in Web3 and highlights why privacy is a pressing concern,” he added.

AI systems usually process inference in centralized data centers, exposing sensitive data to its hardware owners. However, most decentralized AI networks are permissionless, meaning anyone can own a node that processes inference requests.

This means that the owners of these nodes can see what data is being processed, introducing risks of exposing sensitive data to malicious actors. “If centralized entities accessing your data is concerning, it’s even riskier to have random, anonymous nodes process it without safeguards,” Khan said.

Implementing Web3 AI Privacy

Khan believes that through decentralization, AI can achieve complete privacy and verifiability without the need to trust centralized entities.

Atoma addresses AI privacy threats through its Trusted Execution Environment (TEE)-based model. This isolates data during processing, ensuring that even node operators cannot access sensitive information. Khan shared that TEEs also maintain integrity and confidentiality, ensuring compliance with privacy standards while delivering verifiable results.

He says that using crypto-economic principles can enhance and incentivize AI privacy in the Web3 space. For example, using cryptographic methods like on-chain attestations and confidential computing “ensures transparency while maintaining security.”

“These techniques allow users to verify that data is processed securely without exposing sensitive information,” Khan said.

Nvidia Rules

Nvidia has been the dominant player in hardware in the AI space for the past few years, and the future looks no different.

Khan notes that Nvidia’s landmark Hopper architecture in the company’s H100 and H200 GPUs gave more companies access to private AI inference and training for large-scale computing.

For AI privacy on this leading hardware, businesses usually rely on TEEs to secure sensitive data.

“However, privacy measures still significantly impact performance, particularly for training,” Khan notes.

Looking to the future, Khan predicts that Nvidia’s Blackwell chips could mark a “major step forward in the adoption of private AI at scale.”

Nvidia’s next generation of chips, unveiled in March, has been marketed by the company as significantly enhancing AI performance and efficiency.

The upcoming B100 and B200 models promise to deliver up to  30 times greater inference performance while consuming 25 times less energy than their predecessors for large-scale AI models.

AI Safety Gold Standard

Khan said that Fully Homomorphic Encryption (FHE) is “often touted as the future gold standard for privacy.”

FHE is a cryptographic technique that allows users to process data without decrypting it. This means that data can be encrypted and sent to a cloud environment for processing without any sensitive data being exposed.

With FHE, companies can offer online services without ever needing to see their users’ data.

However, Khan remains a realist when it comes to the groundbreaking technique:

“Despite the hype, it remains far from production-ready, with current implementations requiring weeks or months for a single AI inference, making it impractical in the near term.”

Web3 Testing Ground

As Web3 matures, it could offer a sandbox to test privacy-preserving AI technologies that could eventually be adopted across various industries. Khan continued saying, “The Web3 community is innovation-driven and willing to adopt emerging solutions when clear benefits are demonstrated.”

He added that the power of privacy is critical and non-negotiable across industries like healthcare and law, where sensitive data is used daily:

“Without strong privacy guarantees, decentralized AI is unlikely to achieve mass adoption by companies. Businesses cannot risk exposing sensitive or proprietary data, making privacy a non-negotiable requirement for Web3 AI to gain widespread trust and use.”

Ultimately, if the experiments and trials on Web3 are a success, it could lead to a set of new standards on how AI should handle privacy, making it a critical area for innovation.