As governments, academics, and technologists debate how to rein in the risks of artificial intelligence (AI), recent developments from different angles underscore the urgency of designing systems with safety and privacy built in from the start.
Technical solutions already exist: decentralized AI can address safety and transparency concerns without waiting for regulation.To explore how decentralized AI could shift the trajectory of AI adoption, Techopedia spoke with Zac Cheah, Co-Founder of Pundi AI. Cheah discussed why centralized AI poses systemic risks, how decentralized design introduces transparency and accountability, and whether enterprise adoption could outpace regulation.
- With 200+ Nobel laureates and experts backing the Red Lines warning to the UN, the pressure is on to design systems with safety and privacy from the ground up.
- By keeping sensitive data off-chain while recording operations on-chain, decentralized AI reduces the dangers of centralized control and opaque decision-making.
- On-chain audit trails and clear participation rules can give enterprises the assurance they need to move to adoption.
- Safety must be treated as a first-class design constraint from the start, just like performance or scalability.
From Red Lines to Real-World Solutions: Decentralized AI in Action
More than 200 Nobel laureates and AI experts, including OpenAI co-founder Wojciech Zaremba, have called on the United Nations to establish “clear and verifiable red lines” for AI by 2026.
The UN letter warns that “AI systems have already exhibited deceptive and harmful behavior, and yet these systems are being given more autonomy.”
As Stuart Russell, Distinguished Professor of Computer Science at the University of California, Berkeley, and a signatory of the UN letter, argues, we need “built-in safety in the design from the very start” rather than “after-the-fact attempts to reduce unacceptable behavior.”
Meanwhile, tech giant Samsung proved this is technically feasible, unveiling a partnership with decentralized science (DeSci) protocol Galeon that applies decentralized, blockchain-based architecture to healthcare data.
DeSci, which is gaining traction in medical research and healthcare in particular, shows how technical innovation can address the risks policymakers are warning about, rather than waiting for regulation.
For instance, Samsung’s “Blockchain Swarm Learning” approach, which keeps sensitive data off-chain while recording AI operations on-chain for transparency, demonstrates how decentralization prevents the centralized control of AI that enables many of these risks.
Zac Cheah, Co-Founder of Pundi AI, the company building the decentralized data infrastructure layer that makes Samsung-style implementations possible across finance, legal tech, and enterprise software, explained how decentralized AI directly addresses the risks of centralized systems flagged by global experts.
Zac Cheah is a blockchain pioneer and tech entrepreneur who was named among the Top Ten Fintech Leaders in 2019 by the Singapore Fintech Association. An Erasmus Mundus scholar, Zac is a double-Master’s degree holder in Computing and Security from the Swedish Royal Institute and the Norwegian University of Science. Before founding Pundi X, he served as Chief Evangelist Officer of Opera China and chaired the W3C Interest Group in both Beijing and Oslo.
In his present role as co-founder & core contributor of Pundi AI, Zac is building a truly open AI ecosystem with the goal of empowering one million people to earn from AI. Cheah’s expertise spans decentralized AI data platforms, token-economy design, cross-chain interoperability, privacy-preserving frameworks, fintech, and community-driven innovation.
Engineering Accountability Into AI Systems
Q: The UN’s AI Red Lines initiative calls for “clear and verifiable red lines” by 2026. How feasible is it for AI developers to meet these expectations through design alone?
A: The UN’s call for “clear and verifiable red lines” is ambitious but not unrealistic if developers approach it as a design problem rather than a lobbying exercise. Policy statements alone cannot regulate AI safety. You need systems that prove, by design, they cannot cross those lines.
At Pundi AI, we see this in three layers.
- First, the data layer. Consent, provenance, and access must be enforced as code, not paperwork. With decentralized infrastructure, you can tokenize rights, track usage on-chain, and keep raw data under local control. Swarm or federated learning lets hospitals and enterprises train models without centralizing sensitive records. This is exactly what Samsung and Galeon are piloting in women’s health, where ultrasound imaging stays inside hospitals while models improve collaboratively.
- Second, the model layer. Developers should ship models with safety manifests and red-team test suites tied to agreed boundaries. Proofs of compliance and verifiable logs make those claims auditable. It is not enough to say “trust us.” Regulators, customers, and even end-users need a way to check that models respect the red lines in practice.
- Third, the governance layer. Decentralized AI makes accountability measurable. Smart contracts can record model activity, enforce penalties for violations, and even act as circuit breakers. Communities can vote on acceptable thresholds while retaining transparent audit trails.
Will design alone solve everything? No. You still need accredited auditors, global standards, and sector-specific laws. But the heavy lifting starts in architecture. Decentralized AI gives you privacy by default, verifiable training, and open governance.
That is how we move faster than policy timetables and show the UN and the public that red lines can be real, enforceable, and already embedded in the way systems are built.
Q: How technically achievable is it to build in safety from the start for large-scale AI systems?
A: It’s achievable, but not in the way people sometimes imagine. You can’t sprinkle in “safety features” at the end of development and call it done. Building safety from the start means treating it as a first-class design constraint, like performance or scalability.
There are three big levers.
- Architecture. Choose designs that reduce risk by default. That means keeping sensitive data at the edge, using verifiable training methods like federated learning, and building systems that leave audit trails others can check. These are not afterthoughts. They have to be part of the blueprint before you train your first model.
- Testing. Safety needs to be encoded into the same pipelines that handle accuracy and efficiency. Every large model should come with red-team test suites that hammer on known failure modes, plus reproducible benchmarks tied to clear thresholds. If you want to stop a model from crossing a red line, you need an automated way to prove it never does, not just a policy note.
- Governance. Technical safety only matters if it is enforced. That means building control layers into the deployment stack. Smart contracts, attestations, and circuit breakers can act as guardrails that trigger if a model starts producing outputs outside its declared scope. Governance is usually treated as a policy problem, but it is just as much an engineering problem.
Is this easy? No. Safety features often cut against speed and convenience. But technically, it is achievable, and we already know how to build parts of it. The challenge is discipline: starting every project with the assumption that the system must produce proof of safety, not a promise of safety.
Can Distributed Systems Rein in AI’s Dark Side?
Q: How can decentralized AI prevent the types of deceptive or harmful behaviors the UN letter warns about?
A: The key is that decentralized AI removes the single choke points where deception usually takes hold. A few examples make it clear.
- Data manipulation. In a centralized system, one operator can decide what training data to include or exclude, which makes it easy to tilt outcomes. In a decentralized setup, training happens across many independent nodes. Each participant validates updates, and consensus is required before the global model changes. That makes quiet manipulation much harder.
- Opaque decision-making. When model updates are logged on-chain, every adjustment has a traceable record. Auditors and even the public can see which institution contributed which updates and when. This accountability discourages tampering and builds trust.
- Concentration of power. Decentralized governance spreads authority. A decentralized autonomous organization (DAO) or similar mechanism lets communities vote on acceptable use cases and impose penalties for violations. It is harder for one company or government to unilaterally weaponize the system.
- Disinformation. A decentralized model can embed safeguards at the protocol layer, for example, rejecting outputs that fail verification checks or routing suspect content for community review. The same infrastructure that handles consensus on financial transactions can be used to flag and contain deceptive outputs.
The UN letter is right to worry about AI crossing lines of autonomy and truthfulness. The answer is not just rules on paper, but systems designed to prove they cannot be quietly bent. Decentralization gives you that proof, because it replaces ‘trust the operator’ with ‘verify the record’.”
Q: How does decentralized AI reduce the risks of centralized control that can lead to misuse or errors in AI systems?
A: Centralized control means a single authority sets the rules, owns the data, and decides how models behave. That concentration creates two risks: misuse by design and errors that spread unchecked. Decentralized AI addresses both.
When training and governance are distributed across many independent nodes, no one party can alter the system without others noticing. Updates are validated collectively and logged in a way that cannot be quietly changed. This reduces the chance of a single bad actor pushing the system in a harmful direction.
Decentralized AI draws from multiple contributors, each keeping their own data but sharing improvements to the model. The resulting system is less prone to the blind spots or biases of a single source.
In a decentralized structure, every update, decision, and model change can be traced. That creates a permanent record of who did what and when. If something goes wrong, it is easier to identify the source and correct it, rather than relying on the word of a single company.
Errors in a centralized system cascade because everything depends on one hub. But in a decentralized network, faults are isolated, and consensus mechanisms prevent flawed updates from becoming global defaults. The system is harder to hijack and quicker to self-correct.
From Concept to Enterprise Reality
Q: What are the main challenges in implementing decentralized AI at enterprise scale, and how are companies like Samsung overcoming them?
A: The challenges are real, and they explain why decentralized AI has been slower to reach enterprise scale than the hype suggests. But the path forward is clearer now.
Enterprises already rely on established IT and compliance frameworks. Plugging decentralized infrastructure into those workflows without breaking operations is hard. Samsung’s work with Galeon shows one approach: start at the edge with specific devices like ultrasound machines, and connect them to decentralized training without forcing hospitals to replace their existing records systems.
Handling sensitive data across jurisdictions is a major obstacle. Companies need to prove that personal information never leaves local custody. Federated and swarm learning help by keeping raw data in place and only sharing encrypted model updates. Galeon claims to have this running across European hospitals, which makes adoption more feasible in regulated sectors like healthcare.
Training across many nodes is slower and more complex than centralized pipelines. The workaround is selective: run the heavy lifting locally, then share only what is necessary to update the global model. That balance keeps costs in check while still preserving the benefits of decentralization.
Enterprises need assurance that collaborative systems will not expose them to legal or reputational risk. This is where on-chain audit trails and clear participation rules come in. By recording updates transparently and letting members vote on priorities, decentralized AI networks build trust that no single participant is gaming the system.
Companies like Samsung are overcoming these challenges by starting small, proving the model in one use case, and showing that decentralization can slot into real-world operations without disrupting them. Step by step, that is how enterprise adoption moves from pilot to scale.
Q: How can decentralized AI approaches be applied beyond healthcare, for instance, in finance, legal tech, or enterprise software?
A: Decentralized AI is not limited to healthcare. The same principles can apply wherever sensitive data, trust, and accountability matter.
- Banks and payment networks are under constant pressure to share insights without exposing client data. Decentralized learning allows different institutions to train fraud detection or risk models on their own transaction records, then contribute to a stronger global system.
- Law firms and courts deal with highly confidential documents. A decentralized AI framework could let them collaborate on contract analysis or case law research while keeping client records locked in their own systems.
- Large companies often sit on siloed datasets spread across business units and geographies. Decentralized AI can connect those silos without centralizing all the data. Imagine a multinational with separate HR, supply chain, and customer support divisions: each trains locally, and the global model learns across the entire organization. The result is coordination without a single point of failure.
The core idea is the same across sectors. Decentralized AI keeps data where it belongs, shares only improvements, and enforces accountability through transparent records. That model is adaptable to any industry where collaboration is valuable but centralization is risky.
Q: Could decentralized AI become a standard approach for sensitive industries, or will regulation still dominate?
A: Regulation will always play a role, but decentralized AI is well-positioned to become the standard for sensitive industries because it turns compliance into a technical feature rather than a paper exercise.
Regulators want assurance that data is not misused, that decisions can be audited, and that no single actor holds too much control. Decentralized AI addresses all three.
At the same time, policy frameworks move slowly and often lag behind technology. If industries wait for perfect rules before acting, adoption stalls. By contrast, decentralized design gives companies a way to demonstrate compliance by default.
So regulation will still dominate the narrative, but the systems that thrive will be those that make compliance provable by design. In that sense, decentralized AI is less an alternative to regulation than a foundation that regulators can build on. Sensitive industries need both, but decentralization can set the standard for how trust is engineered into AI.
Q: Could the adoption of decentralized AI reshape the current debates about AI ethics, liability, and accountability?
A: Yes, because decentralized AI changes who holds the keys and who bears the responsibility. Most debates on AI ethics and liability assume a central operator who builds, owns, and controls the system. When control is distributed, the questions shift.
Instead of one company setting the moral boundaries, decentralized networks can embed community governance. Participants can vote on acceptable use cases, red lines, or funding priorities. That does not eliminate ethical dilemmas, but it makes the process more transparent and less dependent on the values of a single actor.
In centralized systems, responsibility is clear: the operator is accountable if the model causes harm. Decentralization introduces shared responsibility. This pushes the need for new legal frameworks, but it also reduces the risk of concentrated negligence. Liability can be tied to verifiable records of contribution, with each participant accountable for the updates they provide.
Decentralized AI leaves a permanent trail of decisions and model changes. This addresses one of the hardest problems in AI governance: black-box behavior. If something goes wrong, investigators can trace the path back through the logs and identify which updates or votes led to the outcome. That level of transparency is difficult to achieve in closed systems.
In short, decentralization does not erase the debates around ethics and accountability, but it reframes them. It turns them from abstract questions of trust into concrete questions of design and governance that can be audited, shared, and enforced.
The UN’s call for “red lines” in AI highlights the urgent need for systems built with safety and privacy at their core, following privacy-by-design principles from the start. By distributing control, keeping sensitive data off-chain, and ensuring transparency through on-chain records, decentralized AI directly addresses the risks of centralized systems flagged by global experts.
With industries from healthcare to finance already exploring these solutions, the shift from experimental blockchain projects to enterprise infrastructure is underway. The question is how quickly organizations will adopt it.
FAQs
It ensures safety, compliance, and user trust are built into AI architecture from the start, preventing misuse instead of trying to fix it later.