Who Decides What Your AI Can Say? Inside Model Censorship and Alignment

When you ask an AI to help with something and it refuses, that refusal didn't happen by accident. Someone, or more precisely, a team of researchers, lawyers, and ethicists at a major AI lab, made a deliberate choice to build that boundary into the model. Today, major model providers (e.g. OpenAI, Anthropic, Google, Meta, Mistral, and Cohere) each maintain their own alignment teams, each with distinct values, risk tolerances, and commercial pressures shaping what their models will and won't do. Those decisions affect every downstream application built on those models.

Understanding who makes these decisions, and how, matters if you're building AI products, fine-tuning models, or simply trying to understand why your AI behaves the way it does.

How alignment gets built into models

Alignment isn't a single switch. It's layered in across the training pipeline. The two primary mechanisms are Reinforcement Learning from Human Feedback (RLHF) and Constitutional AI (Anthropic's variant). In practice, both involve humans rating model outputs according to guidelines that encode the provider's values.

Those guidelines are not public. OpenAI's model spec became partially visible only after a leak in 2023. Anthropic publishes high-level principles but not the detailed rating rubrics given to contractors. Google's guidelines for Gemini are similarly opaque. What you see as a user is the output, not the instructions that produced them.

The practical result is that every closed model you deploy carries invisible constraints you didn't negotiate. A model might refuse to generate marketing copy for certain legal-but-controversial industries, add unsolicited warnings to medical questions, or systematically hedge on politically sensitive topics. You find out at runtime, often in production.

The table of control: Closed vs. open models

Whether you have any recourse depends heavily on the model you chose.

The divide is clear. Closed API models give you partial system-prompt control and no access to underlying alignment decisions. Open-weight models (e.g. Llama, Mistral, and their derivatives) let you fine-tune alignment out entirely, replace it, or layer your own on top.

This is not a minor technical distinction. If you're building a platform for security research, legal analysis, medical documentation, or any domain that closed models systematically hedge around, the difference between "full control" and "partial override" is the difference between a product that works and one that doesn't.

Real-world cost: Running your own aligned model

Open-weight models require compute. That cost is real, but it's often smaller than teams assume when they first price it out.

Worked example — Llama 3.1 70B inference for a 10-person dev team:

• Model: Llama 3.1 70B (fits in ~40 GB VRAM with 4-bit quantization)
• Hardware needed: 1x NVIDIA A100 80GB or 2x A40s
• VOLT spot price: $1.49/hr for A100 PCIe
• Assumed utilization: 8 hrs/day, 22 working days/month
• Monthly compute: 8 × 22 × $1.49 = $262.24/month

Compare that to OpenAI GPT-4o at $5.00 per 1M input tokens. A team sending 50M tokens/month pays $250/month just in inference fees. And this comes with zero ability to modify alignment, and all prompts and outputs logged to OpenAI's servers.

For $262/month on VOLT, you get a model you fully control, running on hardware you're renting, with no data leaving your inference stack.

Why VOLT

VOLT's decentralized GPU marketplace is directly relevant to the model censorship question because the ability to run open-weight models depends on having affordable, accessible compute.

Censorship-sensitive deployments typically have two requirements that commodity cloud providers make expensive: cost and data isolation. VOLT addresses both. Spot pricing for H100s starts at $1.99/hr, A100s at $1.49/hr, and A40s (adequate for 13B–34B models) at $0.79/hr. None of those rates require enterprise contracts or reserved instance commitments.

More importantly, VOLT doesn't inspect or log your model's inputs and outputs. You're renting GPU time, not subscribing to a managed AI service with its own content policies layered on top. What you run is your business. For teams building in legal tech, security research, or any other domain that closed-model providers have decided to restrict, that separation matters.

---

Related Questions

Q: Can you remove alignment from a closed model like GPT-4 with jailbreaks?

No, not reliably or durably. Jailbreaks exploit gaps in RLHF coverage, and providers patch them continuously. More importantly, using jailbreaks likely violates the provider's terms of service, exposing your application to account termination. If you need consistent behavior that closed models refuse to provide, open-weight fine-tuning is the only durable path.

Q: What's the difference between alignment and censorship?

Alignment is the broad project of making AI systems behave according to specified values including safety, honesty, and helpfulness. Censorship is a subset, refusing to produce specific outputs. The distinction matters because well-aligned models should also refuse, but the refusal criteria should be principled and transparent. When a model refuses things that are legal and clearly benign, that's alignment implemented badly.

Q: How do model providers decide what to restrict?

Through a combination of policy (legal team input), safety research (red-teaming for harmful outputs), and commercial risk assessment (what topics create liability). The weighting varies. OpenAI's decisions reflect their consumer product exposure. Anthropic's reflect their published Responsible Scaling Policy. Meta, distributing open weights, restricts far less because they're not operating an inference service. There is no neutral arbiter. Every decision is made by an organization with particular incentives.

Q: Can I fine-tune a model to remove restrictions without losing capability?

Generally yes, with caveats. Fine-tuning Llama 3.1 on domain-specific data without RLHF constraints will remove most alignment behaviors. Capability is largely preserved at the base model level. The risk is that you lose beneficial behaviors alongside the restrictions you wanted to remove. Most teams fine-tune to add domain knowledge and constraints specific to their use case, rather than to remove all guardrails entirely.

Q: Is running open-weight models on rented GPU actually production-viable?

For inference workloads under ~100 concurrent users, yes. A single A100 PCIe handles roughly 10–30 requests/second on a 13B model with 4-bit quantization, depending on context length. At VOLT's $1.49/hr spot rate, that's workable for most SMB-scale applications. Scaling beyond that requires multi-GPU setups or batching strategies, but the economics still beat managed API fees once you're past roughly 30M tokens/month.