When users prompt an AI for a summary of a political conflict or a medical explanation, the response feels like a neutral retrieval of facts. However, beneath the interface lies a complex layer of human-designed constraints, known as guardrails, that determine not just what the AI says, but how it says it and what it is forbidden from mentioning.
This invisible curation has sparked a growing debate over who decides what AI tells you, as the gap between the technical goals of developers and the expectations of the public widens. Campbell Brown, the former head of news at Meta, suggests that the industry is currently operating in two different realities.
According to Brown, the internal logic driving AI development in Northern California is fundamentally disconnected from the experience of the people using the tools. She notes that the conversation in Silicon Valley is often focused on “alignment” and “safety”—technical terms for ensuring a model doesn’t produce toxic content or violate corporate policies—while consumers are more concerned with accuracy, transparency, and the perception of ideological bias.
The mechanics of AI alignment
To understand how an AI’s “worldview” is formed, one must look at Reinforcement Learning from Human Feedback (RLHF). This is the process where human reviewers rank various AI-generated responses based on specific guidelines. These reviewers essentially act as the first editors of the AI, rewarding answers that are polite and safe while penalizing those that are provocative or contradictory to the company’s established policies.
The result is a model that is “aligned” with the values of the organization that built it. While this prevents the AI from providing instructions on how to build a bomb or using hate speech, it can also lead to “refusals”—where the AI declines to answer a factual question because it deems the topic too sensitive or “controversial.”
For those with a background in journalism or software engineering, this process looks less like objective programming and more like an editorial desk. The challenge is that unlike a traditional newspaper, where an editor’s name or a publication’s mission statement is public, the “editorial guidelines” for Large Language Models (LLMs) are often proprietary and opaque.
A disconnect in priorities
The tension Brown identifies stems from a misalignment of risk. For a company like Google or OpenAI, the primary risk is a “hallucination” or a biased response that goes viral and creates a PR crisis or legal liability. The “safety” conversation in the Valley is often about risk mitigation.
Consumers, however, view the tool as a utility for truth. When an AI steers a user away from a specific viewpoint or provides a sanitized version of a historical event, the user does not perceive “safety”—they perceive censorship or manipulation. This creates a friction point where the tool’s attempt to be “helpful and harmless” is interpreted as being biased.
| Perspective | Primary Objective | View of Guardrails | Key Risk |
|---|---|---|---|
| Silicon Valley | Model Alignment | Essential safety measures | Reputational/Legal damage |
| Consumers | Information Accuracy | Potential censorship/bias | Loss of objective truth |
Who holds the pen?
The actual decision-making power resides with a relatively small group of policy leads, ethics researchers, and engineers. These individuals define the “system prompts”—the hidden instructions the AI receives before the user ever types a word—and the reward functions used during training.
This concentration of power raises significant questions about the democratization of information. If a handful of companies in one geographic region define the boundaries of “acceptable” truth for a global user base, the potential for cultural and political hegemony is high. The “truth” provided by an AI is not a reflection of a global consensus, but a reflection of the consensus reached by a specific set of employees at a specific company.
The stakes are particularly high in the realm of news and current events. As AI becomes a primary interface for information discovery, the way these models summarize news can shift public perception. If the guardrails are tuned to avoid “controversy,” the AI may omit the nuances of a conflict, presenting a flattened version of reality that satisfies corporate safety requirements but fails the test of journalistic integrity.
The path toward transparency
Addressing the question of who decides what AI tells us requires a shift from “black box” alignment to transparent governance. Some industry advocates suggest the following steps to bridge the gap between the Valley and the consumer:
- Publicly available system cards: Detailed documentation explaining the specific guidelines used to train the model’s guardrails.
- User-adjustable settings: Allowing users to choose between different “alignment profiles” (e.g., a “strictly factual” mode versus a “creative/exploratory” mode).
- Third-party auditing: Independent reviews of model biases by academic or non-profit organizations to ensure the AI isn’t steering users toward specific political or social conclusions.
Without these measures, the AI’s role as an information intermediary remains a point of vulnerability. The goal of “safety” is necessary, but when safety is defined solely by the entity selling the product, it can inadvertently become a tool for steering the narrative.
The next significant checkpoint for this debate will be the continued implementation of the EU AI Act, which aims to establish stricter transparency requirements for high-risk AI systems, potentially forcing companies to be more explicit about how their models are steered.
Do you feel that AI responses have become too sanitized, or are guardrails necessary for a safe digital ecosystem? Share your thoughts in the comments below.
