Last week I pulled up the on-chain activity for three credential protocols I've been tracking since 2021. Combined daily active issuers: eleven. Eleven people, across three protocols, issuing verifiable proof that a human did something real. Meanwhile, Singapore announced it is putting roughly 80,000 financial workers through AI skills training — something like 40% of its entire banking workforce — and there is not a single public plan for how those 80,000 people will prove, to anyone, what they actually learned.
We didn't build the rails for this. Or rather — we did, and then we forgot we built them.
I've spent the better part of a decade watching cryptographic proofs move from academic curiosity to the thing that lets stablecoins settle in eleven seconds. So when I read a policy headline about mass AI upskilling, my brain doesn't go to productivity charts. It goes to a much older problem: who attests to the truth of a claim, and how do you verify it without asking permission from the party that issued it?
That's the whole ballgame. And right now, Singapore is playing it blind.
The Arithmetic Nobody Put in the Press Release
To understand why this matters, you have to look at the numbers underneath the narrative. Singapore's financial services sector employs somewhere between 170,000 and 200,000 people. The training target of 80,000 is not a rounding error. It's between 40 and 47 percent of the entire industry. That is not a seminar series. That is a structural repricing of a national labor force, dressed in the language of "career resilience."
The plan sits inside a broader strategy. Singapore's National AI Strategy 2.0 landed in December 2023, backed by a budget commitment that runs into the billions of Singapore dollars. The regulatory scaffolding came earlier still — MAS began publishing AI governance principles back in 2018 with the FEAT framework, covering fairness, ethics, accountability, and transparency, and has been layering on generative-AI guidance ever since. Institutions like the Institute of Banking and Finance, working alongside the Association of Banks in Singapore, are the standard vehicles for this kind of national credentialing push.
None of that is the interesting part. The interesting part is what the headline leaves out: funding sources, curriculum depth, assessment standards, and — most critically — whether the resulting certificate means anything at all.
Here's the thing I keep coming back to. This is a blockchain story wearing a labor-policy costume. Because the moment you promise that 80,000 people are now AI-competent, you've made a claim about identity. You've asserted that a specific human possesses a specific capability at a specific moment. And in a world where a large language model can generate a flawless certificate PDF in four seconds, the only way that claim survives contact with reality is if it's cryptographically anchored to something the holder controls.
Identity isn't a document. Identity is a set of verifiable claims about a person — issued by parties you trust, provable to parties who don't trust you. That's it. Everything else is marketing.
What Actually Exists (And Why Nobody's Using It)
Let me be concrete, because the gap here isn't conceptual. It's operational.
The standards exist. The W3C Verifiable Credentials data model has been a recommendation since 2019. It defines a portable, tamper-evident format for exactly this use case: a university, an employer, or a training body issues a credential; the holder stores it in a wallet they control; a verifier checks the issuer's signature without ever phoning home to the issuer. That last part is the revolution nobody notices. In the traditional model, verifying a certificate means contacting the issuing institution, which means the institution remains a permanent, rent-seeking gatekeeper. In the VC model, the credential is self-verifying. The issuer's job ends at issuance.
On the settlement side, we have soulbound tokens — ERC-5192 formalized the idea, though the concept predates the standard — which are non-transferable NFTs bound to an address. The appeal is obvious: you cannot sell your degree, you cannot rent your certification, and you cannot buy your way into a skill you don't have. The Ethereum Attestation Service generalized this further, letting anyone issue an on-chain attestation about anything, with schemas that define what's being claimed. Attestations can be revocable, expirable, and gated by staking — meaning an issuer who lies can be economically slashed.
And then there's the privacy layer, which is where my own 2017 detour into ZoKrates and zero-knowledge proofs finally pays rent. The objection to on-chain credentials has always been the same: I don't want my entire educational and employment history permanently visible to anyone with a block explorer. Fair. But ZK credentials solve this cleanly. Using systems like Semaphore or zk-creds, you can prove a statement — "I hold a valid AI-competency attestation issued by an accredited body, and my overall score is above the 70th percentile" — without revealing which attestation, which body, or what score. The verifier gets what it needs. You leak nothing else.
So the technology stack for verifiable, privacy-preserving, fraud-resistant AI skill credentials has been production-ready for roughly three years. Which brings me to the question that should be printed on the front of every policy document in Singapore: why isn't it being used?
The answer is banal. Because issuing a PDF is free, and issuing a verifiable credential requires someone to think about architecture. The 80,000-banker program, like almost every national upskilling initiative I've audited since 2020, will almost certainly default to the path of least resistance: completion certificates stored in a learning management system, verifiable by emailing HR. Which is to say, verifiable by nobody.
The Substitution Math Nobody Wants to Publish
Here's where it gets uncomfortable, and where I think the media coverage has been actively misleading.
When Singapore says it will "redefine roles" and "enhance productivity," what it is describing, in plain economic terms, is a substitution curve. I've watched this play out at the institutional level. DBS — the most aggressive large bank in the region on AI — has publicly claimed AI has generated hundreds of millions of Singapore dollars in value, and has signaled that headcount in certain contract and temporary functions will shrink over the coming years. That's the most honest data point we have, and it comes from a bank that is genuinely ahead of the curve.
Now layer on estimated substitution rates, drawn from what I've seen across audit engagements and cross-referenced against industry practice. Software development and testing: 30 to 60 percent exposure, with an equally high augmentation rate. Customer service and ticket handling: above 50 percent exposure, because AI customer agents are already mature. Compliance and risk first-pass screening: 20 to 40 percent, constrained by regulatory requirements for human sign-off. Relationship-driven business — private banking, investment advisory: below 15 percent, augmentation-dominant. Back and middle office operations, reporting: above 50 percent.
The brutal part is the timing. The training cycle for a national program like this runs two to three years. That is precisely the window in which most institutions expect large-scale AI deployment to hit. So the people being trained are the same people being repriced, simultaneously. They will learn to use the tool that replaces the surrounding workflow. That's the tension nobody in the press release wants to name.
And when the certification they receive is a PDF with no cryptographic anchor, no transferable proof of effort, no tamper-evident record — they will have spent two years acquiring a credential that cannot defend itself in a labor market that is being repriced in real time.
Proof of Effort Is the Missing Primitive
This connects to something I worked on directly. In 2021 I co-founded a project that started as an NFT play and pivoted hard into what we called "provability of effort." The insight was simple: an on-chain record of work performed is fundamentally different from a self-reported résumé, because the record is created at the moment of work, not years later under interview pressure. I wrote about blockchain-verified volunteer hours, and it got picked up by people who had nothing to do with crypto, because the underlying idea — that contribution can be attested continuously rather than claimed retroactively — maps onto every credentialing problem people actually have.
The Singapore program is the largest test of that idea the world has ever run. And by defaulting to non-verifiable credentials, it is running the test with the instrument switched off.
Think about what's possible. A trainee completes a module on prompt engineering for compliance workflows. Instead of a PDF, they receive a W3C-compliant verifiable credential, issued by an accredited body, anchored to a soulbound token on a low-cost L2, with a ZK proof that lets them demonstrate competency without revealing their identity. The bank that hires them verifies the credential in under a second, offline, without calling the issuing body. The worker owns the credential and can carry it across employers, jurisdictions, and career pivots. When they upgrade their skills, the old credential is revocable and a new one supersedes it, with the full history auditable.
That is not science fiction. Every component exists today. The reason it isn't happening is that credentialing is a game of institutional territory, and the institutions currently holding the pen have no incentive to make their records portable. A verifiable, portable credential is a threat to the training provider's lock-in. That's the real reason you don't see it.
The Data Governance Bottleneck (The Part Everyone Misdiagnoses)
Here's a contrarian take that I'll defend all day: the bottleneck in financial AI is not talent, and it never was.
I've sat in rooms where banks spent six figures training staff on generative AI tools, only to discover that the internal systems holding the data the AI needed could not be wired into the model without a multi-year replatforming project. The staff became capable. The organization stayed frozen. The productivity gain never materialized, because capability without integration is a hobby.
The Singapore program, by training 80,000 people, is building the capability layer aggressively while remaining silent on the integration layer. That asymmetry is the dog that isn't barking. If the training succeeds perfectly, the average Singapore bank still faces the same wall: legacy core banking systems that don't expose clean APIs, data residency rules that complicate cloud inference, and governance requirements that demand human oversight on any customer-facing decision. You can train the humans perfectly and still get zero throughput if the pipes are clogged.
This is the mirror image of a problem I know well from DeFi. Uniswap V4 gave everyone the ability to build custom hooks — programmable, composable liquidity logic. On paper, infinite innovation. In practice, the complexity spike scared off the overwhelming majority of developers, because the tooling to safely build, test, and audit a hook is a different skill entirely from using the protocol. Capability is not the constraint. The interface between capability and production is the constraint.
Liquidity isn't money. Liquidity is the presence of consent at a given price. And in the same way, AI skill isn't productivity. AI skill is the presence of consent between a worker and a system that can actually deploy them. Right now, Singapore is manufacturing consent on the worker side and forgetting the system side.
The Institutional Lock-In Nobody's Naming
There's a quieter risk buried in all of this. The generative AI tools that Singapore's bankers will be trained to use run, overwhelmingly, on models and cloud infrastructure owned by a handful of American firms. Every unit of AI capability Singapore imports deepens its dependency on operators it does not control and cannot sanction. The country has decided — rationally, given its size — to be a fast adopter rather than a foundational model builder. But fast adoption of a foreign stack is a vulnerability, not a moat.
This is where the credential layer matters more than it appears. If Singapore anchors its AI skills to verifiable on-chain credentials, it owns the record layer even if it rents the compute layer. The proof of what its workforce can do becomes a sovereign asset, portable, auditable, and immune to the whims of any single vendor. If it doesn't, its workforce capability lives inside the LMS of whichever training conglomerate won the contract, and its verification depends on that vendor's continued goodwill.
Freedom isn't the absence of dependency. Freedom is the presence of consent — and you can only consent to a dependency you can actually see, measure, and exit. An opaque, vendor-locked credential system is a dependency you cannot see.
The Bear Market Lens
I've written before about resilient engineering — the projects that kept shipping while prices collapsed. The 2022 downturn taught me to stop asking "what's the narrative" and start asking "what's the structural integrity." Apply that lens here and the picture sharpens.
The narrative says Singapore is future-proofing its financial workforce. The structural reality says Singapore is running a massive, under-instrumented labor experiment during a period when the labor market is being repriced by forces it doesn't control, using credentials that can't verify themselves, on data infrastructure that may not be ready. None of that makes the program wrong. It makes it incomplete in ways that matter.
In a bear market, survival beats gains. The professional equivalent is: provenance beats credentials. A certificate that can't prove its own origin is worth less than the paper it isn't printed on.
What I'd Actually Watch
If you're tracking this as a signal rather than a headline, watch three things. First, the funding mechanism — whether the program routes through the Financial Sector Development Fund or charges banks directly, because that tells you whether it's a strategic bet or a cost pass-through. Second, whether MAS updates its generative-AI guidance in parallel with the training, because that would signal the compliance modules are real rather than decorative. And third, whether DBS, OCBC, or UOB disclose quantified AI-driven cost reduction in any quarterly filing over the next eighteen months. That number will be the first honest measure of whether 80,000 trained humans actually changed anything.
And watch Hong Kong. The HKMA runs its own fintech talent initiatives and GenAI sandboxes. If Hong Kong responds with a more aggressive or more technically sophisticated program, Singapore's advantage evaporates, because its moat was never the policy — it was the execution speed.
The Question That Should Haunt the Press Release
Eighty thousand people are about to be told they've been made AI-capable. Almost none of them will receive a credential that can prove it without permission from whoever issued it. That is the same trust model we had with paper diplomas in 1970, wrapped in the vocabulary of a technology that solved the problem a decade ago.
So here's what I keep circling back to. When the first Singapore banker loses a role to the very automation she was trained to operate, and she applies for the next job holding a certificate that no one can independently verify — what exactly did those two years buy her?
The rails exist. The standards are ratified. The cryptography is proven. What's missing is the political will to hand workers a credential they actually own. And in a market where AI can forge any document in seconds, the only credential worth anything will be the one that can prove itself. That's not a blockchain talking point. That's just the math.
