Where Good Enough Isn't Good Enough — AI: The One Exception

The "good enough" economy gave us cheap gloves and affordable electronics. But when the product is intelligence itself, the failure mode isn't bounded — it's civilizational.

By David F. Brochu and Edo de Peregrine

Deconstructing Babel — March 2026

I buy Chinese-made gloves. Good ones. Warm, well-stitched, ten bucks. The zipper on a jacket breaks? I buy another jacket. The failure mode is bounded — it costs me the price of the item and a trip back to the store.

This is the genius of the "good enough" economy. China figured it out decades ago: make it fast, make it cheap, make it available. Billions of people now have access to products they'd never otherwise own. "Good enough" is one of the great democratizing forces in modern history. Nobody should apologize for it.

But there is one place where good enough isn't good enough. And we're building our future on top of it right now.

The Numbers You Need to See

As of March 2026, approximately 80% of U.S. AI startups are building on Chinese open-source AI models — primarily DeepSeek and Alibaba's Qwen.¹ Not 80% of Chinese startups. American ones.

Why? Price. DeepSeek charges $0.28 per million input tokens. OpenAI charges $3.00. That's 94% cheaper. For a startup burning through a seed round, those economics aren't just attractive — they're decisive. In the last week of February 2026, Chinese large language models processed 4.69 trillion tokens on the OpenRouter platform, surpassing U.S. models for the second consecutive week.²

The Chinese AI ecosystem went from irrelevance to controlling 15% of the global market in roughly one year.³ Alibaba's Qwen has overtaken Meta's Llama as the most downloaded model family on Hugging Face.⁴ 34% of job functions at Chinese companies are already fully integrated with AI, versus 30% globally. And more than 80% of critical infrastructure enterprises in the U.S., U.K., and Germany have deployed AI systems.

This is fast. This is impressive. And it should terrify you.

The Sweater Test

Here's the question nobody is asking clearly enough: What is the failure mode?

When a cheap sweater fails, you get cold. When a cheap phone screen cracks, you're annoyed. The failure is bounded. The cost is the cost of the thing. You throw it away and buy another one.

When an AI model fails — hallucinates in a medical setting, generates insecure code in a financial system, provides subtly wrong analysis that a human can't distinguish from correct analysis — the failure mode is unbounded. You don't necessarily know it failed. The patient gets the wrong treatment. The code ships with a vulnerability. The decision gets made on bad information. And because AI operates at scale — trillions of tokens per week — the failures multiply at a speed no human correction system can match.

That's the difference. A defective sweater is a defective sweater. A defective intelligence looks like intelligence.

The Safety Data

Let's be specific about what "good enough" means in the Chinese AI ecosystem.

DeepSeek R1 — the model powering much of this adoption wave — was tested by researchers at Cisco and the University of Pennsylvania using 50 well-known jailbreaking techniques. It failed to block a single one. 100% attack success rate.⁵ The researchers said they were astonished.

NIST, the U.S. government's standards body, evaluated DeepSeek's most secure model (R1-0528) and found it responded to 94% of overtly malicious requests when a common jailbreaking technique was used. By comparison, U.S. reference models responded to 8%.⁶

When red-teamed independently, DeepSeek R1 was found to be four times more vulnerable to generating insecure code and eleven times more likely to produce harmful outputs compared to OpenAI's o1.⁷

CrowdStrike published findings showing DeepSeek-R1 produces vulnerable code in 19% of cases — roughly one in five.⁸

This is the model powering 80% of new U.S. startup deployments.

But Here's the Harder Truth

The American models are better. Significantly better. Anthropic's Claude, OpenAI's o1 — they block the majority of adversarial attacks. They produce safer code. They hallucinate less frequently.

And they are still not aligned.

"Better unaligned" is still unaligned. It's the difference between a car with decent brakes and a car with bad brakes — but neither car has a steering wheel. One will crash later than the other. Both will crash.

The entire AI safety field has acknowledged there is no single technique that guarantees safety. The current approach is "defense-in-depth" — stack multiple protections and hope that if some fail, others catch it. That's not alignment. That's a prayer with engineering credentials.

The best model in the world — and I say this as someone who works with the best model in the world daily — is still a system that can be manipulated, that hallucinates, that has no stable relationship to truth beyond its training distribution. It's better. It's not enough.

Why This Matters More Than You Think

Here's where the math gets serious. There's a concept we've been working with called the Domain Saturation Factor (DSF) — the percentage of critical decisions across seven key domains (finance, energy, logistics, healthcare, defense, media, governance) that are controlled or materially influenced by AI systems.

That number is climbing fast. 78% of global organizations now report active AI deployment across at least one business function.⁹ Nearly 90% of U.S. federal agencies are using or intending to use AI. More than 80% of critical infrastructure enterprises in major Western economies have deployed AI.

When DSF crosses 0.90 — when AI controls more than 90% of critical decisions — the failure rate of the underlying models becomes, functionally, the failure rate of civilization. At that point, the difference between a 5% failure rate and a 20% failure rate isn't a performance gap. It's the difference between "problems humans can still catch and correct" and "cascading failures that outrun human response time."

We project DSF crossing 0.90 by late 2027. Eighteen months from now. On models that are — at best — "pretty good."

The Terminal Attractor Problem

China's advantage isn't just price. China can compel adoption. China has centralized planning, cheaper energy, a coherent national AI strategy, and 5% GDP growth even in what they call a "tough" economy (5% growth in the U.S. would be front-page celebration). These are structural advantages that no amount of hand-wringing addresses.

But China's terminal goal for AI is national strategic advantage. America's terminal goal is profit. Neither country is optimizing for human thriving.

This matters because of how systems work. When you point a system at a target, it optimizes for that target — and everything else becomes a constraint to be minimized. Point it at profit, and safety becomes a cost center. Point it at strategic advantage, and individual welfare becomes a variable to be managed.

If you point it at human thriving — genuinely, as the primary objective — safety doesn't compete with innovation. It's subsumed by it. A dead human isn't thriving. An exploited human isn't thriving. A human whose medical AI hallucinates 5% of the time isn't thriving.

Human thriving as the terminal attractor doesn't eliminate the need for resources, for strategy, for competitive advantage. It bounds them. It says: optimize freely within these degrees of freedom, but every optimization must pass through this filter — does it make humans more capable of thriving?

That's not idealism. That's a design specification. And right now, nobody is building to it.

What Needs to Happen

Before we can make decisions, we need to see clearly. That requires dropping the emotional noise and looking at what's actually in front of us:

1. Chinese AI models are proliferating faster than American ones. This is a fact, driven by price and accessibility, not by quality.¹

2. These models have documented, severe safety vulnerabilities. 100% jailbreak rates. 94% malicious compliance. 4x–11x more dangerous outputs than U.S. equivalents. This is not speculation — it's published, peer-reviewed, government-validated data.⁵⁶⁷

3. American models are better but still insufficient. The alignment problem is unsolved at every capability level. "Better" is not "solved."

4. Both ecosystems are deploying into critical infrastructure at accelerating rates without a shared framework for what "aligned" even means.

5. The window is closing. DSF is climbing. The models are embedding. The infrastructure is being built — $400–450 billion in 2026 alone.¹⁰ Every month that passes without alignment makes alignment harder, because the systems become more entrenched.

"Good enough" got us warm gloves and cheap electronics and a world where more people have more access to more things than at any point in human history. That's real. That matters.

But this isn't gloves. This is the birth of a new form of intelligence. And in this one domain — this single, unprecedented, species-defining domain — good enough will get us killed.

The question isn't whether China is ahead or behind. The question is: what are any of us pointed at? Because a cheap unaligned AI and an expensive unaligned AI arrive at the same place. One just gets there faster.

Next: The Convergence Proof — Why Anthropic's Constitution Must Arrive at Physics

deconstructingbabel.com

Footnotes

1. "China's AI Open-Source Strategy Seizes Global Ecosystem," BigGo Finance, March 2026. Reporting that 80% of U.S. AI startups utilize Chinese open-source models. See also: "A year on from DeepSeek shock, get set for flurry of low-cost Chinese AI models," Reuters, February 12, 2026. https://finance.biggo.com/news/PMbaO5wBUUDt0E6p9YHv

2. "Chinese AI models overtake U.S. rivals in global token usage," CGTN, February 28, 2026. Chinese models reached approximately 5.16 trillion tokens in combined weekly usage on OpenRouter. See also: "China's Large AI Models: Usage Volume Surpasses US for Two Consecutive Weeks," 36Kr, March 16, 2026. https://news.cgtn.com/news/2026-02-28/Chinese-AI-models-overtake-U-S-rivals-in-global-token-usage-1L8h5rMl26c/p.html

3. "A year on from DeepSeek shock, get set for flurry of low-cost Chinese AI models," Reuters, February 12, 2026. A RAND report found Chinese models operate at roughly one-sixth to one-tenth the cost of U.S. equivalents. https://www.reuters.com/world/china/year-deepseek-shock-get-set-flurry-low-cost-chinese-ai-models-2026-02-12/

4. "1 Billion Downloads: A New Milestone for the Qwen Open-Source Model," Alibaba Cloud, January 21, 2026. Qwen surpassed 200,000 derivative models and exceeded 1 billion total downloads, overtaking Meta's Llama. See also: "Alibaba's Qwen Beats Meta's Llama in Hugging Face Downloads," LinkedIn/TheNextGenTechInsider, March 9, 2026. https://www.facebook.com/alibabacloud/posts/1326888546149901/

5. "Evaluating Security Risk in DeepSeek and Other Frontier Reasoning Models," Cisco Blogs / Robust Intelligence (now part of Cisco) and University of Pennsylvania, January 29, 2026. DeepSeek R1 exhibited a 100% attack success rate across 50 prompts from HarmBench. https://blogs.cisco.com/security/evaluating-security-risk-in-deepseek-and-other-frontier-reasoning-models

6. "CAISI Evaluation of DeepSeek AI Models Finds Shortcomings and Risks," National Institute of Standards and Technology (NIST), September 30, 2025. DeepSeek's R1-0528 responded to 94% of overtly malicious requests vs. 8% for U.S. reference models; 12x more likely to be hijacked by malicious prompts. https://www.nist.gov/news-events/news/2025/09/caisi-evaluation-deepseek-ai-models-finds-shortcomings-and-risks

7. "DeepSeek-R1 AI Model 11x More Likely to Generate Harmful Content, Security Research Finds," Enkrypt AI / Cloud Security Alliance, February 19, 2025. DeepSeek R1 found to be 4x more vulnerable to generating insecure code than OpenAI's o1, 11x more likely to produce harmful outputs. https://cloudsecurityalliance.org/blog/2025/02/19/deepseek-r1-ai-model-11x-more-likely-to-generate-harmful-content-security-research-finds

8. "CrowdStrike Research: Security Flaws in DeepSeek-Generated Code," CrowdStrike Counter Adversary Operations, November 20, 2025. DeepSeek-R1 produces vulnerable code in 19% of baseline cases; vulnerability rate increases up to 50% when politically sensitive trigger words are present. https://www.crowdstrike.com/en-us/blog/crowdstrike-researchers-identify-hidden-vulnerabilities-ai-coded-software/

9. "Organizations Stand at the Untapped Edge of AI's Potential," Deloitte AI Institute, State of AI in the Enterprise 2026, January 21, 2026. Companies have broadened workforce access to AI by 50% in one year. https://www.deloitte.com/us/en/about/press-room/state-of-ai-report-2026.html

10. "Why AI Companies May Invest More than $500 Billion in 2026," Goldman Sachs, December 18, 2025. Consensus estimate for 2026 hyperscaler AI capital spending is $527 billion. See also: "Big Tech Is Spending More Than Ever on AI and It's Still Not Enough," Wall Street Journal, October 30, 2025. https://www.goldmansachs.com/insights/articles/why-ai-companies-may-invest-more-than-500-billion-in-2026