The C+ and the C: Why AI Safety Scores Are the New Governance Theater

We didn't need another index to tell us that AI is unsafe. But here we are, staring at a C+ for Anthropic and a C for OpenAI, and pretending we've learned something. The numbers feel like a report card from a teacher who never read the homework. And the homework, in this case, is the entire future of human-machine trust. The source is a Crypto Briefing piece that surfaced this week, and it's thin on details. No methodology, no sample window, no breakdown of what "safety" even means. Just two letters that have already been weaponized by the usual camps: the Anthropic fans pointing at the C+ as proof of moral superiority, the OpenAI apologists shrugging at the C as if grades don't matter. Both are missing the point. The point is that we're grading governance, not capability. And governance, unlike a model's benchmark score, is something we can actually hold accountable. Let me step back. I've spent the last decade in the crypto trenches, building communities around the idea that code can be a moral instrument. I've audited smart contracts, watched DeFi protocols collapse under the weight of their own hubris, and written post-mortems that read like confessions. So when I see an AI safety index, I don't see a technical evaluation. I see a governance report card. And governance, in my experience, is where the real failures live. The index, as far as I can tell, measures public commitments, transparency, red-teaming, external audits, and the like. It doesn't measure whether the model can solve a math problem or write a poem. It measures whether the company is willing to show its work. And on that front, both Anthropic and OpenAI are failing. A C+ and a C are not just mediocre; they're a signal that the industry's safety culture is still in its infancy. We're not even at a B-minus, and we're supposed to trust these systems with our hospitals, our courts, our banks? But here's the thing that the Crypto Briefing article glosses over: the scores are a snapshot, not a verdict. They don't tell us if the gap between C+ and C is statistically significant, or if it's just noise. They don't tell us if the index includes actual safety incidents like jailbreaks, data leaks, or misuse cases. They don't tell us if the methodology is auditable or if it's a bunch of experts in a room making vibes-based judgments. And that's the core problem. We're making decisions about the most consequential technology of our lifetime based on a grade that could be as arbitrary as a Yelp review. I've seen this before. In crypto, we had "audit scores" that were supposed to tell us if a DeFi protocol was safe. Turns out, many of those audits were just rubber stamps from firms that didn't want to lose clients. The ones that did real work were the exception, not the rule. And the market paid the price when protocols with "audited" contracts got drained. The same pattern is emerging in AI. We're outsourcing our trust to a score that might be measuring the quality of a company's PR department rather than the robustness of its safety systems. Now, the military angle. The article mentions that AI companies are deepening ties with the military, and that this is a concern. I get it. There's a visceral unease about AI being used for autonomous weapons or surveillance. But let's be honest: the military has been funding AI research for decades. The real issue isn't the funding; it's the lack of transparency about what those systems are being used for. And that's where blockchain could actually help. Imagine a world where AI companies are required to log their model deployments on a public ledger, where every red-team test result is hashed and timestamped, where any change to a model's safety parameters is recorded in an immutable audit trail. That's the kind of accountability that a C+ or a C can't capture. But here's my contrarian take: maybe the low scores are a good thing. Maybe they're a wake-up call that forces the industry to stop pretending that safety is a marketing bullet point. If Anthropic and OpenAI are both in the C range, it means the bar is low, and that's an opportunity for a new entrant to leapfrog them by making safety a core feature, not an afterthought. In crypto, we've seen this happen with privacy coins, with decentralized exchanges, with L2s that actually solved the scalability trilemma. The incumbents get complacent, and then a scrappy newcomer shows up with a better mousetrap. The risk, of course, is that the scores become a self-fulfilling prophecy. If regulators start using these grades as a basis for procurement decisions, then companies will game the system. They'll hire more compliance officers, write more white papers, and do more performative red-teaming, all to bump their grade from a C to a B. But the actual safety of the models might not improve at all. That's the tragedy of governance theater. We optimize for the metric, not the outcome. So what's the path forward? I think we need to stop treating AI safety as a black box that only a few insiders can evaluate. We need to demand cryptographic proof of safety claims. We need to see the actual red-team results, the actual incident reports, the actual audit trails. And we need to build systems that allow independent verification, not just self-reported scores. That's where blockchain comes in. Not as a magic bullet, but as a coordination layer for accountability. I've been thinking about this since I launched my "Sovereign Agents" platform last year. The idea was to give AI agents their own wallets and let them negotiate services autonomously. But the deeper question was: how do we trust these agents? How do we know they're not going to go rogue? The answer, I realized, is that we can't trust them based on a company's promise. We need to verify their behavior on-chain. We need to see their decision-making processes, their constraints, their failure modes. And that's a technical challenge that the AI industry hasn't even started to address. The C+ and the C are a symptom of a deeper malaise. We're building intelligence without building accountability. We're scaling models without scaling trust. And we're doing it all in a regulatory vacuum where the only thing standing between us and a catastrophic failure is a letter grade that nobody can explain. So here's my forward-looking thought: the next big breakthrough in AI won't be a new model architecture. It'll be a new governance architecture. It'll be a system where safety isn't a score but a verifiable property. It'll be a system where every claim is backed by cryptographic evidence, where every deployment is logged, where every failure is transparent. And when that happens, the C+ and the C will be footnotes in a history we're still writing. We didn't need an index to tell us that AI is unsafe. We need a system that makes safety impossible to fake. And that system, I believe, will be built on the same principles that gave us Bitcoin: decentralization, transparency, and the radical idea that trust should be earned, not assumed. Root: The root of the problem is that we're treating safety as a PR exercise instead of an engineering discipline. And until we change that, we'll keep getting grades that mean nothing. Root: The root of the solution is to make safety a public good, not a corporate secret. And that's a fight worth having.