By ByteBot
Anthropic got a C+. That’s the best grade. The Future of Life Institute released their Winter 2025 AI Safety Index on December 3, grading eight major AI companies on safety practices. For existential safety—preventing AGI loss of control—every single company got D or F. Not Anthropic. Not OpenAI. Not Google. Zero companies have adequate plans to prevent catastrophic misuse or loss of control of superintelligence.
Max Tegmark, MIT professor and FLI president, put it bluntly: “AI is less regulated than sandwiches.” These aren’t theoretical concerns. You use Claude, ChatGPT, Gemini daily. The companies building these tools are racing toward AGI without safety guardrails.
The Grading Breakdown: C+ Is the Best We Get
FLI’s independent panel of eight AI safety experts—including Stuart Russell, David Krueger, and Dylan Hadfield-Menell—evaluated Anthropic, OpenAI, Google DeepMind, xAI, Meta, Z.ai, DeepSeek, and Alibaba Cloud across six domains using 35 safety indicators. The results reveal a clear divide between top performers and everyone else.
| Company | Overall Grade | Score | Existential Safety |
|---|---|---|---|
| Anthropic | C+ | 2.67 | D |
| OpenAI | C+ | 2.31 | D |
| Google DeepMind | C | 2.08 | D |
| xAI | D | 1.17 | D/F |
| Meta | D | 1.10 | D/F |
| Alibaba Cloud | D- | 0.98 | F |
Anthropic, OpenAI, and Google DeepMind scored A’s and B’s on information sharing, risk assessment, and governance. The bottom five—xAI, Z.ai, Meta, DeepSeek, and Alibaba Cloud—lack foundational safety frameworks and meaningful disclosures. Chinese companies fared worst, with Alibaba scoring 0.98 out of 4.0.
Moreover, the evaluation covered Risk Assessment (dangerous capabilities testing), Current Harms (actual safety performance via HELM Safety benchmarks), Safety Frameworks (risk management processes), Existential Safety (AGI control plans), Governance (whistleblower protections, oversight), and Information Sharing (technical disclosures, policy engagement).
Existential Safety: Everyone Failed the Critical Test
Here’s the headline buried in the report: every company got D or F for existential safety. This isn’t a secondary metric—it’s the “prevent human extinction” category. Anthropic’s C+ overall grade becomes a D when evaluating plans to control smarter-than-human AI. OpenAI? D. Google? D. xAI, Meta, Alibaba? F.
The FLI report states: “All companies reviewed race toward AGI without presenting explicit plans for controlling smarter-than-human technology, leaving consequential risks unaddressed.” Stuart Russell, one of the expert panelists, summarized it: AI companies claim they can build superhuman AI, but none have demonstrated how to prevent loss of human control over such systems.
What does “no testable plan” mean in practice? Companies have no concrete safeguards with measurable triggers, no demonstrated monitoring mechanisms, no credible strategies for maintaining control if their models become smarter than humans. They’re building AGI first, figuring out control later. That’s not a plan—that’s hope.
OpenAI’s Lobbying Problem: “Safety First” Meets Political Reality
OpenAI earned a C+ overall, but the report singles them out for criticism: “Must reduce lobbying against state safety regulations.” While publicly stating “safety is core to our approach” and investing in frontier safety research, OpenAI actively lobbied against state-level AI safety legislation. The FLI report calls out “ambiguous safety thresholds” and “insufficient independent oversight.”
This creates cognitive dissonance. You can’t credibly claim safety is your top priority while fighting laws designed to enforce safety standards. It’s like a pharmaceutical company saying patient health matters most while opposing FDA oversight. NBC News reported that Tegmark emphasized US AI companies “remain less regulated than restaurants and continue lobbying against binding safety standards.”
Anthropic, in contrast, received praise for being “relatively supportive of both international and US state-level governance and legislative initiatives.” Still, they need “more concrete, measurable thresholds and stronger evaluation independence with external validators.” Even the best performer has significant gaps.
The AGI Race Is Happening Without Guardrails
This isn’t a theoretical future problem. Companies are actively pursuing superintelligence right now. OpenAI is “turning attention to superintelligence” in 2025. Meta rebranded as “Meta Superintelligence Labs” in June 2025. Microsoft formed a “MAI Superintelligence Team.” Sam Altman publicly discusses AGI timelines for 2025.
The FLI report warns: “The increase in AI capabilities is severely outpacing safety-focused efforts” and “This widening gap between capability and safety leaves the sector structurally unprepared for the risks it is actively creating.” Translation: we’re accelerating toward a cliff with no brakes.
Furthermore, as Euronews highlighted, the “less regulated than sandwiches” comparison isn’t hyperbole. Food safety has health inspections, FDA oversight, strict manufacturing standards, liability frameworks. AI building superhuman intelligence? Self-regulation and voluntary commitments that this report proves aren’t working.
What This Means for Developers and Users
You use these tools daily. Claude for code reviews. ChatGPT for documentation. Gemini for research. The companies building them don’t have adequate safety plans for the systems they’re racing to create. That’s not an abstract policy debate—it’s your workflow, your productivity stack, your career tooling.
The report’s recommendations are straightforward: companies must produce concrete safeguards with measurable triggers, realistic thresholds, and demonstrated monitoring. Either credibly plan for AGI control or clarify they won’t pursue such systems. Independent oversight, not self-assessment. Binding standards, not voluntary promises.
Additionally, developers face a tension: AI proficiency is becoming a job requirement (68% say so, according to related surveys), but the tools we’re adopting come from companies failing independent safety evaluations. Axios noted that even top performers like Anthropic “remain far from meeting adequate safety expectations.”
The Report Card Everyone Failed
C+ shouldn’t be celebrated when it comes with a D for existential safety. The Winter 2025 AI Safety Index makes abstract concerns tangible: grades, scores, company-by-company breakdowns. Anthropic leads at 2.67. Alibaba Cloud fails at 0.98. Everyone gets D or F for AGI control.
Tegmark’s “less regulated than sandwiches” quote captures the absurdity perfectly. We’ve built robust safety frameworks for food, pharmaceuticals, aviation, nuclear power. For AI potentially smarter than humans? Nothing binding. Just promises that independent evaluators say aren’t translating to adequate safeguards.
The AGI race is happening now, not in 2030. Companies are rebranding for superintelligence while failing basic safety tests. Developers and users are caught between adoption pressure and mounting evidence that voluntary commitments aren’t working. That’s the uncomfortable reality the FLI’s Winter 2025 report card reveals: when it comes to AI safety, nobody’s passing the test that actually matters.
