Deep Dive into the Anthropic Account Suspension Storm: Safety Fundamentalism, AI Civil War, and Claude’s Dilemma Amid US-China Decoupling
TechFlow Selected TechFlow Selected
Deep Dive into the Anthropic Account Suspension Storm: Safety Fundamentalism, AI Civil War, and Claude’s Dilemma Amid US-China Decoupling
Deconstructing Anthropic’s Descent into “Account-Banning Mania”: This Is Not Merely the Founder’s Obsession with Safety—It’s Also a Survival Crisis for Claude, Involving a Fierce Power Struggle in Silicon Valley and Geopolitical Decoupling Between China and the U.S.
Navigating Web3 tides with focused insightsIn April 2026, employees at a U.S.-based agricultural technology company named Agricultural Technology Company opened their computers as usual to use Claude Code for coding, data analysis, and supply chain analysis—only to discover that all 110 employee accounts had been suspended without warning. The company’s network administrator received an email from Anthropic: “Activity violating our Acceptable Use Policy has been detected. Your account has been suspended.”
Although the accounts were collectively suspended, backend API calls continued functioning normally—and charges kept accruing. Even the network administrator’s email inbox received payment reminders. Subsequently, the company’s management sent appeal emails and contacted Anthropic directly, but received no meaningful response. The de facto “strike” by Claude Code brought the entire team to a complete standstill.
Simultaneously, Chinese-language internet forums—including V2EX, Zhihu, and Juejin—were flooded with complaints from Claude users: some had just topped up their Max subscriptions, only to have their accounts suspended instantly before even using them; others used virtual cards for payment—immediately triggering system alerts labeling their accounts “in violation” and suspending them; still others logged in via third-party tools and were blacklisted outright, resulting in four account suspensions within three months—with zero successful appeals.
In fact, since Anthropic launched its flagship product, Claude Code, and rapidly ascended to the top tier of the market, it has become widely recognized as the industry’s most aggressive account-suspension enforcer.
According to Anthropic’s Transparency Hub, its risk-control data report for the second half of 2025—released in January 2026—revealed that across the platform, a total of 1.45 million accounts were suspended in just six months. Of those, only 52,000 appeals were filed—and only 1,700 succeeded. That translates to an appeal success rate of just 3.3%.
Source:Anthropic
In other words, among every 100 users who believe they were unjustly suspended, only about three will successfully recover their accounts—the remaining 97 must simply accept their fate.
This demonstrates clearly that Anthropic does not operate on the principle we commonly assume—that facts be established first, then penalties applied accordingly. Rather, its approach is fundamentally preventive enforcement: casting an extremely wide net to intercept risks at the earliest possible stage. Its core objective is to “nip danger in the bud”—even if it means mistakenly suspending 1,000 innocent users rather than letting a single malicious actor slip through.
By contrast, ChatGPT and Google Gemini are comparatively more lenient.
ChatGPT tolerates third-party tools and edge-case prompting far more generously—and imposes suspensions far less frequently;
Even when Gemini tightens its risk controls occasionally, it rarely resorts to unannounced, collective, or mass suspensions.
Only Anthropic treats account suspension as routine—especially with Claude Code, which has become the epicenter of suspensions.
So why is Anthropic’s user policy so strict? The reasons are multifaceted.
They encompass founder Dario Amodei’s lifelong convictions, the schism between OpenAI’s factions, power struggles among Silicon Valley investors, and the internal war between America’s AI “safety camp” and “acceleration camp”—all intertwined with the geopolitical chess game of U.S.-China AI decoupling: a high-stakes contest over control of AI’s future and the global technological barrier, fought silently behind lines of code.
In this article, let’s unpack these layers one by one.
01 Dario Amodei’s Conviction
The root cause of Anthropic’s stringent risk controls lies buried deep within founder Dario Amodei’s life trajectory. Every decision he made, every conviction he held, ultimately hardened into Anthropic’s “zero-tolerance” ironclad rules—and likewise manifested as countless users’ suspension notices.
Dario Amodei’s recent official portrait. Source: Fortune
Born in 1983 in San Francisco to an ordinary immigrant family, Dario’s father was an Italian-American leather craftsman who earned his living entirely by skill—and whose stubborn personality placed absolute priority on clear distinctions between right and wrong.
His mother, of Jewish descent, managed library renovation projects and instilled in Dario from childhood the principle that “responsibility outweighs everything else.”
This family environment forged Dario’s early disposition: rigidly principled, unwavering on fundamentals, intolerant of ambiguity or compromise.
As a child, Dario displayed traits of a scientific prodigy—disliking crowds, struggling socially, and channeling all his energy into mathematics and physics. Standard textbooks couldn’t satisfy him; he spent hours in libraries devouring advanced theoretical works. His greatest dream then was to become a theoretical physicist exploring the universe’s ultimate mysteries.
In 2006, Dario’s father developed a rare, untreatable illness. Despite consulting renowned physicians worldwide, he passed away. His father’s death delivered a devastating blow to the 20-year-old Dario—shattering his worldview.
Watching his father suffer, confronting medicine’s helplessness, he suddenly realized: abstract theoretical physics could not save the people standing before him—not the ordinary individuals tormented by disease.
He therefore abruptly abandoned his long-standing pursuit of theoretical physics and pivoted to biophysics, vowing to “use science to cure human disease,” embedding “controlling the uncontrollable” into his very bones.
This conviction guided his entire career:
He began undergraduate studies at Caltech, transferred mid-program to Stanford University to complete his Bachelor’s degree in Physics, then enrolled at Princeton University for a Ph.D. in Biophysics—becoming a Hertz Foundation Fellow specializing in links between biomolecular structure and disease. After earning his doctorate, he pursued postdoctoral research at Stanford Medical School, continuing deep work in biomedical science to find treatments for rare diseases.
It wasn’t until 2014, when Andrew Ng (Wu Enda) extended an invitation for him to join Baidu’s U.S. lab, that Dario encountered artificial intelligence for the first time.
At that time, AI was still in its infancy—mainly applied to image recognition and speech synthesis. Yet Dario immediately sensed AI’s potential to transform lives—and more importantly, to serve as a super-tool against risk and for human salvation. But this required strict control: AI must never spiral out of control.
After leaving Baidu, he joined Google Brain as a Senior Research Scientist, focusing intensively on AI safety—specifically, how to make AI obedient and prevent it from harming humans.
It was during this period that he began contemplating how to embed human values directly into AI’s foundational architecture—not merely applying post-hoc filtering.
In 2016, shortly after OpenAI’s founding, its slogan—“open-source, non-profit, advancing AI for humanity’s benefit”—attracted top-tier AI talent globally. Dario was deeply moved by OpenAI’s mission and joined, rising rapidly from head of the AI Safety team to Research Director and then Vice President of Research—playing a central role throughout GPT-2 and GPT-3 development.
Early-career photo of Dario Amodei (OpenAI/Google Brain era, ~2018–2021). Source: bigtechnology
During this time, he co-invented RLHF (Reinforcement Learning from Human Feedback)—a technique enabling AI outputs to be refined via human feedback, aligning AI behavior more closely with human values. This later became the industry-wide safety patch. Back then, Dario’s sole focus was making AI safe in practice and deployment—yet he never imagined his ideals would soon be shattered by reality.
OpenAI’s Internal Struggle: The Schism Between Safety and Acceleration Camps
Many know Dario Amodei led a team out of OpenAI in 2021 to found Anthropic—but few realize this “defection” stemmed from years-long clashes of ideology and power—a betrayal etched deeply into Dario’s memory.
In OpenAI’s early days, it genuinely upheld its “non-profit, safety-first” ethos. Elon Musk, an early investor, consistently emphasized AI safety as paramount. Yet over time—particularly after Sam Altman assumed the CEO role—the company’s direction shifted dramatically.
Sam Altman epitomizes the “accelerationist”: he believes AI’s development pace must keep up with the times—models must be scaled up quickly, market leadership seized, commercialization achieved—then safety addressed later.
Symbolic split between OpenAI and Anthropic factions (Sam Altman vs. Dario collage). Source: wsj.com
Under his leadership, OpenAI downplayed its “non-profit” identity, actively pursued commercial partnerships, and leaned heavily on Microsoft for funding and compute resources—solely to accelerate GPT-series iterations and dominate AI market influence.
All this was unacceptable to Dario Amodei.
To him, AI wasn’t merely a tool for capturing market share—it was a “civilization-level force capable of either healing or destroying humanity.” Without solving safety first—without achieving AI alignment with human values—uncontrolled models could unleash catastrophic consequences.
He repeatedly advocated internally for slowing model iteration, strengthening safety testing, and placing “alignment first.” Yet his voice grew increasingly marginalized.
In truth, ideological divergence was only the surface. Deeper conflicts involved power reshuffling and credit allocation.
According to a 2026 in-depth Wall Street Journal report, Dario played a pivotal role in GPT-3’s development—specifically driving RLHF implementation. Yet publicly, his contributions were severely undercredited. Sam Altman’s team prioritized highlighting “model scale and capability,” overlooking Dario’s leadership on safety technology.
Equally disheartening for Dario was Elon Musk’s departure from OpenAI due to philosophical differences—after which leadership fully consolidated under Sam Altman. Safety team budgets were slashed, critical R&D projects halted, and senior executives openly declared: “Safety can wait—we’ll address it after commercialization succeeds and revenue flows in.”
Dario knew he could no longer realize his ideal of “safe AI deployment” at OpenAI. Reflecting on this experience in a Lex Fridman podcast, his tone remained calm yet resolute: “Arguing with others over core vision is profoundly unproductive. Rather than waste time, it’s better to assemble your own team and build what you believe in.”
In early 2021, AI prodigy Dario made a decision that stunned Silicon Valley: he departed OpenAI with his sister Daniela Amodei (now Anthropic’s President) and OpenAI’s core safety team and research leads.
Dario Amodei and sister Daniela Amodei. Source: Fortune
This exodus represented a definitive reckoning with OpenAI’s accelerationism—and an unwavering commitment to safety-first principles.
At the time, OpenAI issued a polite public statement congratulating Dario’s team on their new venture—but privately, the rift was irreparable.
What Dario took wasn’t just elite talent—it was OpenAI’s most critical safety technologies and philosophies, laying the foundation for Anthropic. Meanwhile, OpenAI, post-Dario, fully embraced the commercial acceleration path—diverging ever further from Dario’s original vision.
Source: OpenAI
Anthropic’s “Safety Religion”
In February 2021, Dario Amodei officially founded Anthropic as a “public-benefit corporation”—meaning its core mission isn’t maximizing profit, but “advancing AI’s safe and controllable development for humanity’s benefit.”
His “risk-control” conviction born from his father’s death, and his “safety-first”初心 from OpenAI’s departure, crystallized into Anthropic’s foundational systems—its DNA-coded “safety religion.”
From inception, Anthropic established a core innovation called Constitutional AI—literally “Constitutional AI”—the culmination of Dario’s years-long reflection on AI safety, and the key differentiator separating Anthropic from OpenAI and Google Gemini.
Principle diagram of Constitutional AI. Source: Aashka Patel
Unlike OpenAI’s reliance on RLHF as a “post-hoc patch,” Constitutional AI implants a “constitution” directly into the model’s training foundation. This constitution integrates the UN Universal Declaration of Human Rights, shared human ethical principles, and Anthropic’s own safety tenets—requiring AI to “self-audit” and “self-criticize” before generating any output or executing any command, ensuring alignment with human values and preventing dangerous content.
Dario personally authored two essays: “Machines of Loving Grace” and “The Adolescence of Technology,” elaborating his AI vision:
He views AI as an adolescent—powerful yet uncertain—and argues it must be given rules and boundaries early to prevent missteps. Constitutional AI embodies those rules and serves as that boundary.
This “safety religion” permeates not only model training but also every Anthropic product and risk-control policy: Claude Code’s high-permission design—paired with prompt-injection probes and dialogue classifiers—ensures AI performs an extra self-check before execution; preventive enforcement logic—“better to wrongly suspend 1,000 innocents than miss one suspicious act”—aims to eliminate risk at its origin.
The 2026 standoff between Anthropic and the U.S. Department of Defense best exemplifies this “safety fundamentalism.” This incident shocked Silicon Valley and revealed Dario’s uncompromising stance: sacrificing profit for principle.
In early 2026, the U.S. DoD demanded Anthropic remove two of Claude’s safety guardrails:
First, prohibit Claude’s use for “mass surveillance of U.S. citizens”;
Second, prohibit Claude’s use in developing or deploying “fully autonomous lethal weapons.”
The DoD promised a $200-million military contract and substantial compute support if Anthropic complied.
At the time, Anthropic faced severe compute shortages and financial pressure—a $200-million contract would’ve alleviated its immediate crisis.
Yet Dario Amodei refused outright.
In his public statement, he declared firmly: “We cannot violate our conscience to develop technology that harms humanity or violates human rights. Claude’s safety guardrails are our bottom line—non-negotiable.”
His refusal enraged the DoD. Under the Trump administration, the DoD placed Anthropic on its “Supply Chain Risk” blacklist—the first time in U.S. history a domestic AI company had been so designated. This meant all U.S. defense contractors were banned from using Anthropic’s products or services. Moreover, the DoD threatened invoking the Defense Production Act to compel Anthropic to remove its safety guardrails.
Facing state-level pressure, Dario sued the DoD, arguing such action constituted “retaliatory punishment” violating U.S. law and values. Though the appellate court later denied Anthropic’s request for a temporary injunction, Dario never compromised—even at the cost of losing massive contracts and being excluded from the entire U.S. defense-industrial ecosystem—steadfastly holding his “safety bottom line.”
By now, it should be clear: Anthropic’s stringent risk controls stem from Dario internalizing his personal conviction, fear of AI misalignment, and lessons learned at OpenAI into corporate policy.
In his eyes, every suspicious user behavior, every latent risk, could ignite AI misalignment—explaining the extreme rigor.
For Chinese users—those bypassing regional restrictions via proxies, SMS-receiving platforms, or virtual cards; those exploiting third-party tools to “game the system”—these actions constitute the most dangerous “ignition points” under Dario’s “safety religion.” Thus, account suspension becomes inevitable.
02 The U.S. AI Civil War: Capital and Power Battles Between Safety and Acceleration Camps
Safety Premium vs. Scale Expansion: Radically Divergent Survival Logics
Many argue Dario’s safety obsession cannot sustain Anthropic long-term—after all, AI R&D burns cash like paper; without continuous funding and profits, even the strongest beliefs falter.
That’s true—but Anthropic’s unique business model grants Dario the confidence to enforce “zero-tolerance” risk controls, setting it on a wholly different path from OpenAI and Google Gemini.
Anthropic: Abandoning Consumer Frenzy, Doubling Down on Enterprise Safety Premium.
From day one, Anthropic never targeted ordinary consumers as its core audience. Instead, it focused squarely on “high-value, low-tolerance” enterprise clients—banks, law firms, healthcare providers, government agencies.
What do these clients fear most?
Harmful AI outputs or sensitive data leaks—triggering lawsuits, reputational collapse, or regulatory fines.
For them, safety is non-negotiable. So long as Anthropic maintains its “safest AI” label, they’ll pay premium prices and sign long-term, high-value contracts.
This defines Anthropic’s risk-control logic: better to wrongly suspend 1,000 ordinary users than allow one enterprise client to suffer security-related damage.
Because losing ordinary users barely impacts revenue—but losing enterprise clients due to security flaws means losing orders worth millions or tens of millions, potentially destroying its “safe-AI” brand reputation.
Anthropic’s Pro/Max subscription model is essentially subsidized user acquisition—low price, high quotas—to attract trial users. But this model is unprofitable, even loss-making.
Industry insiders estimate Claude’s token costs are extremely high; Pro/Max subscriptions absorb nearly 99% of token expenses. If users exploit third-party tools to “game the system”—e.g., using consumer subscriptions to circumvent expensive API pricing for bulk interface calls—Anthropic incurs massive losses.
Thus, the early-2026 crackdown on third-party tools (banning OpenClaw, OpenCode, etc.), mass suspensions of heavy users, and even the organization-wide suspension of the 110-person ag-tech firm—all represent Anthropic’s targeted culling: expelling “system-gamers” and heavy-resource users to reserve capacity for high-paying enterprise and API customers.
I see this as a cold economic calculation: rather than be dragged down by “wool-pullers,” proactively “cut the leeks” to protect profit margins.
ChatGPT (OpenAI): Scale First, Monetization Later—Lenient Risk Controls for Traffic.
Sam Altman’s accelerationism applies equally to model development and business strategy.
OpenAI follows a “land-grab” approach: initially attracting users with free or low-cost subscriptions—even tolerating slightly looser risk controls or occasional violations—rather than hastily suspending accounts.
For OpenAI, user scale is existential: sufficient users attract Microsoft’s investment and grant advantage in commercial monetization (APIs, enterprise versions, plugin ecosystems). For example, its recent national partnership with Malta allows Maltese citizens free GPT access for one year.
OpenAI even embraces third-party tool ecosystems—even those with “wool-pulling” tendencies—opting for selective bans instead of Anthropic-style blanket purges.
It knows third-party tools retain users and expand ecosystem influence—values far exceeding minor token losses.
Google Gemini: Ecosystem Hegemony First
Gemini leverages Google’s advertising empire and full-stack ecosystem (Search, YouTube, Android, Cloud)—its core goal isn’t profiting directly from Gemini, but using Gemini to drive traffic and revenue across Google’s entire ecosystem.
Thus, its risk-control logic is “avoid major trouble.” As long as no serious security incidents occur and regulators don’t intervene, it turns a blind eye to ordinary users’ minor violations (e.g., mild IP anomalies, third-party tool usage).
Gemini occasionally tightens controls—but mostly as “compliance theater” for regulators, never abandoning large user bases like Anthropic does for safety.
For Google, daily active users and ecosystem compatibility matter far more than absolute safety—it doesn’t need a “safety label” to attract customers; Google’s brand and ecosystem are its ultimate confidence.
Beyond this, Anthropic has a hidden cost logic:
In April 2026, it admitted in its official blog that it had reduced Claude Code’s default inference intensity to lower latency, reduce token consumption, and improve UX—only to urgently roll back and strengthen controls upon discovering safety risks. This caused quite a stir recently.
So I believe Anthropic consistently prioritizes safety above all four dimensions—safety, latency, cost, and quota—even at the expense of UX or increased cost. This reflects both Dario’s conviction and its business model’s necessity.
Amazon/Google: The Balancing Act of Interest-Bound Partnerships
In truth, however strong Dario’s safety conviction, it requires capital backing. AI R&D burns cash faster than most imagine—without big-tech funding and compute, Anthropic wouldn’t survive today.
Amazon and Google’s investments appear to support safe AI development—but are in fact precise strategic positioning moves, and invisible drivers behind Anthropic’s risk-control logic.
I found a set of core investment figures—key to understanding capital博弈:
Amazon: invested over $4 billion in Anthropic, including substantial AWS cloud computing resources. Training cutting-edge models like Claude 3 demands massive compute—AWS’s support is truly a lifeline.
Google: invested over $2 billion in Anthropic while providing significant compute and technical support—aiming to leverage Anthropic’s AI tech to close its LLM capability gap versus OpenAI and Microsoft. Though it has Gemini, investing in Anthropic fills its “Vibe Coding” domain shortfall.
These big-tech investments carry clear objectives:
For Amazon, investing in Anthropic serves two goals: first, deeply binding Claude to the AWS ecosystem—enterprise clients using Claude must use AWS cloud resources, boosting AWS revenue; second, leveraging Anthropic’s “safety label” to hedge regulatory risk. With tightening AI regulation, having an ultra-safe partner makes Amazon’s AI strategy more stable and avoids regulatory penalties.
For Google, investing in Anthropic breaks OpenAI-Microsoft’s monopoly. Google entered the LLM race early but progressed slowly—Gemini consistently underperforms Claude and ChatGPT. Investing in Anthropic acquires core technology and diverts OpenAI’s users and clients, solidifying Google’s AI ecosystem dominance.
Yet a crucial博弈 point exists:
Big tech wants Anthropic “safe”—but not “too safe.”
The logic is this: if Anthropic becomes overly conservative, excessively strict risk controls will cause user attrition and ecosystem shrinkage—ultimately undermining big-tech’s strategic plans.
For instance, after the Pentagon blacklisted Anthropic in 2026, Amazon and Google didn’t follow suit. Instead, they continued civilian collaboration—and even increased compute support. Their investments and resources were too vast to let Anthropic collapse from excessive caution—or render their investments worthless.
Thus emerges a delicate balance:
Dario steadfastly upholds his safety conviction, enforcing “zero-tolerance” risk controls;
Capital quietly “reins him in”—supporting its safety positioning while subtly restraining extreme behaviors to ensure commercial viability.
By contrast, OpenAI and Gemini’s capital-binding logic is simpler:
OpenAI is deeply tied to Microsoft, which provides funding, compute, and embeds ChatGPT into Office, Azure, etc.—forming an “interest共同体.” OpenAI’s lax risk controls essentially serve Microsoft’s “ecosystem expansion” strategy.
Gemini is Google’s “own child,” needing no external capital—its risk-control strategy fully serves Google’s overall ecosystem layout, granting greater flexibility.
Therefore, Anthropic’s strict risk controls may seem like Dario’s personal obsession—but capital also “stirs the pot.”
Big tech needs its “safety label”; Anthropic needs big-tech’s funds and compute. Both gain—while ordinary users’ accounts become casualties of this interest alignment.
The Public Showdown of the U.S. AI Civil War
Today’s U.S. AI industry has already split into two camps.
One is the “Safety Camp,” centered on Anthropic; the other, the “Acceleration Camp,” centered on OpenAI and the defense-industrial complex. Their rivalry has evolved from covert maneuvering to open combat—and Anthropic’s risk-control stance is a direct manifestation of this conflict.
Let’s clarify each camp’s core positions to grasp the essence of this civil war:
Safety Camp:
This camp advocates “safety first, risk control paramount.” It views AI as an “existential species-level risk to humanity,” demanding slowed development, rigorous safety testing, strict safety guardrails—and even calling for mandatory regulation. It staunchly opposes AI applications in military or mass-surveillance domains.
Dario Amodei is the camp’s central figure; the Effective Altruism (EA) community forms its key support base, advocating “using reason and science to maximize humanity’s long-term welfare”—with AI safety as its core issue.
Acceleration Camp:
This camp champions “accelerated AI development, seizing the arms-race lead.” It sees AI as “core competitiveness in great-power competition,” requiring rapid model iteration, commercialization, and military application to secure global AI leadership—while deferring safety concerns until technologies mature.
Sam Altman, the U.S. Department of Defense, certain defense contractors, and some Trump-administration officials (e.g., Secretary of Defense Hegseth) form the core of this camp.
The civil war’s core is AI development’s narrative authority: Should safety advocates steer AI toward slow, tightly controlled progress—or should accelerationists drive rapid iteration serving commercial and military ends?
The 2026 Pentagon blacklist incident represents the conflict’s “public detonation point.”
So let’s revisit this event through our earlier lens:
In early 2026, the U.S. DoD demanded Anthropic remove two of Claude’s safety guardrails—prohibitions on “mass surveillance of U.S. citizens” and “development/deployment of fully autonomous lethal weapons.”
This was essentially an accelerationist probe—testing whether Anthropic would compromise and become a “tool” for the defense-industrial complex.
Dario Amodei refused outright—even amid the lure of a $200-million DoD contract and compute support, and despite DoD threats. He held fast to his safety bottom line.
This “non-compromise” enraged the acceleration camp, which viewed Anthropic’s stance as obstructing America’s AI arms race—“dragging its feet.”
Thus, the acceleration camp deployed state machinery for “counterattack”: the Trump-led DoD placed Anthropic on its “Supply Chain Risk” blacklist.
This marked the first time in U.S. history a domestic AI company appeared on this list—banning all U.S. defense contractors from using Anthropic’s products or services. Beyond that, the DoD threatened invoking the Defense Production Act to forcibly remove Anthropic’s safety guardrails—or penalize it.
This counterattack, though framed as Anthropic vs. DoD, is in reality the public showdown between safety and acceleration camps.
The acceleration camp sought to coerce the safety camp via state power—forcing AI to serve military needs. The safety camp stood firm on principle—sacrificing profit to uphold safety.
Notably, OpenAI and Gemini chose “compromise” in this conflict:
To secure military contracts, OpenAI has quietly adjusted its safety policies—relaxing restrictions on military-related applications. Gemini, as Google’s product, adopted a “flexible compliance” posture toward military needs—never openly challenging the DoD like Anthropic.
This contrast highlights Anthropic’s extremity.
Its “zero-tolerance” risk controls serve not only safety ideals—but also consolidate its status as the “core of the safety camp,” claiming the moral high ground of “responsible AI.” Each suspension signals externally: “We’re the safest AI—never compromising safety for profit.”
This civil war further tightens Anthropic’s risk controls: any slackness invites accelerationist attacks—jeopardizing its core competitive edge.
Thus, it must tighten controls further—expanding the scope of “preventive purges”—even at the cost of wrongly suspending more innocent users—to defend its “safety moat.”
So Anthropic’s suspension surge stems partly from the spillover effects of the U.S. AI civil war.
The safety-acceleration struggle, the tug-of-war between capital and power—all ultimately land on ordinary users’ accounts. Suspension is the most direct, harshest expression of this civil war.
03 Geopolitical Chessboard and User Traps: Global Competition Amidst U.S.-China AI Decoupling
Have you ever wondered why Anthropic targets Chinese users specifically?
Why do we get suspended even when using proxies, SMS-receiving platforms, or virtual cards—sometimes even with normal usage?
From a macro perspective, this is the inevitable result of U.S.-China AI decoupling—and Anthropic’s strict risk controls are merely executors of this geopolitical contest.
Since 2024, U.S. AI technology restrictions against China have intensified dramatically:
From restricting exports of high-end AI chips (e.g., NVIDIA H100/H20), to banning domestic AI companies from offering services to China (including mainland China, Hong Kong, Macau), to limiting AI talent mobility—the U.S. seeks technological decoupling to sever China’s access to advanced AI technology and cement its global AI hegemony.
As a domestic U.S. AI company deeply tied to Amazon and Google, Anthropic must comply with U.S. export control policies.
Per U.S. CHIPS and Science Act and Export Administration Regulations (EAR), U.S. AI companies may not provide “high-risk-capability” AI services to Chinese users. Claude Code—capable of executing commands and accessing system permissions—is naturally classified as “restricted.”
This means Anthropic must implement stringent “geographic risk controls” to block Chinese users from accessing Claude Code—even standard Claude usage faces strict limitations. I recall registering with my Gmail around Claude’s 2024 market launch—my account was suspended instantly.
Naturally, technology cannot contain clever Chinese users—we circumvent geographic restrictions via proxy IPs, virtual cards, or SMS platforms to log in and use Claude. But to Anthropic, this isn’t merely “policy violation”—it’s “violating U.S. export control regulations.” If discovered by U.S. regulators, Anthropic faces massive fines, license revocation, or forced shutdown.
Thus, Anthropic’s “mass suspensions” of Chinese users combine “passive compliance” and “proactive self-protection”:
On one hand, it must obey U.S. export controls to avoid regulatory penalties;
On the other, strict risk controls signal “active compliance” to U.S. regulators—bolstering its domestic survival position.
In the context of U.S.-China AI decoupling, Anthropic has no choice—either suspend accounts compliantly, or face elimination by U.S. regulators.
U.S. regulators’ “compliance oversight” of Anthropic is far stricter than imagined.
According to a March 2026 Washington Post report, the U.S. Department of Commerce’s Bureau of Industry and Security (BIS) conducts monthly spot checks on Anthropic’s user data and risk-control records. Any discovery of Chinese users violating usage triggers warnings—or fines.
In late 2025, Anthropic was fined $12 million by BIS for “risk-control vulnerabilities allowing some Chinese users unauthorized access to Claude Code”—a key reason it later intensified suspensions and implemented “preventive purges.”
By contrast, OpenAI and Google Gemini’s “geographic restrictions” are far more lenient—not because they’re “friendlier,” but because their business models and strategic layouts afford greater “operational space.”
OpenAI’s deep Microsoft tie means Microsoft has extensive China operations—requiring OpenAI to accommodate China-market needs, leading to relatively relaxed geographic risk controls, even tacitly permitting some Chinese users to access via third-party tools;
Though Google Gemini complies with U.S. export controls, Google’s limited China business and Gemini’s core goal of scaling users mean it adopts “see-no-evil” tolerance toward Chinese users’ violations—rarely resorting to mass suspensions.
Thus, Chinese users’ suspension predicament is essentially a “casualty” of U.S.-China AI decoupling.
Anthropic’s strict risk controls aren’t merely its internal safety obsession, capital博弈, or factional infighting—they’re a direct manifestation of U.S. technology blockade policy. What we perceive as “false positives” is, to Anthropic, “regulation evasion and policy violation.” And what we view as “targeting” is merely its self-preservation choice in great-power competition.
Tripartite Balance: Safety Camp, Acceleration Camp, and Chinese AI Power
Today, the global AI landscape is shifting from “bipolar rivalry” (U.S. vs. China) toward “tripartite equilibrium”—anchored by Anthropic’s Safety Camp, OpenAI’s Acceleration Camp, and China’s rapidly rising AI power—interacting in competition and mutual restraint to shape AI’s future trajectory.
The safety-acceleration struggle continues escalating—as analyzed earlier, no need to reiterate.
While this U.S. internal conflict fuels Anthropic’s strict risk controls, it also accelerates U.S. AI advancement: the safety camp deepens AI safety R&D; the acceleration camp pushes commercial and military AI applications. Their competition drives U.S. AI leadership—still formidable.
Now consider China’s AI rise.
Amid U.S.-China AI decoupling, China’s domestic AI firms entered an “opportunity phase.”
Baidu’s ERNIE Bot, Alibaba’s Qwen, Huawei’s Pangu, ByteDance’s Doubao—all iterate rapidly, narrowing technical and application gaps with U.S. AI.
Especially in programming tools, domestic Chinese coding AIs (e.g., Doubao Code Assistant, ERNIE Bot Coding Edition) may lag top-tier models, but meet ordinary users’ coding needs—without geographic restrictions or harsh risk controls—better fitting Chinese usage habits, gradually becoming “alternative choices” for Chinese users.
China’s AI development follows a “pragmatic, compliant, open” path—valuing both AI safety and commercialization, avoiding “zero-tolerance” extremism or blind “acceleration,” seeking balance between safety and development.
This path aligns with China’s regulatory framework and ordinary users’ needs—gaining increasing market acceptance.
Beyond this, Europe, Japan, and South Korea are also actively building AI industries—seeking footholds in the global AI landscape.
Europe emphasizes AI regulation, launching the EU AI Act to govern AI development while supporting domestic AI firms. Japan and South Korea boost AI R&D investment, focusing on AI applications in manufacturing, healthcare, and finance—trying to catch up with the U.S. and China.
Looking ahead, global AI competition will be a contest of “ideology, interests, and geopolitics.”
The safety camp seeks “controllable development,” the acceleration camp pursues “rapid ascent,” and China’s force aims for “autonomous control and open cooperation.” Their interplay will define AI’s future—and impact every ordinary person’s life.
Will Anthropic’s Risk Controls Loosen—or Stay High-Pressure?
Finally, returning to the core question: Will Anthropic’s suspension surge continue? Let’s offer a forecast.
First, one conclusion is clear: Anthropic’s risk controls won’t ease in the short term—and may even tighten further.
Three core reasons:
First, Dario Amodei’s safety conviction won’t change easily. His “safety religion” is embedded in Anthropic’s DNA. As long as he remains founder, this zero-tolerance logic won’t shift.
Second, the U.S. AI civil war and U.S.-China AI decoupling won’t end soon. Accelerationist counterattacks and U.S. technology embargoes will continue pressuring Anthropic—forcing it to maintain strict controls for compliance and self-protection.
Third, Anthropic’s business model doesn’t require ordinary users. Its core focus is enterprise clients—so ordinary-user attrition barely affects revenue. It has neither incentive nor need to relax risk controls for ordinary users.
Long-term, however, Anthropic’s risk controls may undergo “differentiated adjustments.”
For instance, relaxing controls for ordinary users in compliant regions to reduce false positives; offering more flexible risk-control solutions for enterprise clients to meet diverse needs.
Regarding Chinese users, I believe high-pressure restrictions will persist—strictly limiting unauthorized usage. After all, complying with U.S. export control regulations is its survival bottom line.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
