In Silicon Valley, artificial intelligence is often discussed like a promise—something that can be engineered, optimized, regulated, and eventually delivered at scale. In everyday life, though, AI is experienced less like a promise and more like a presence: a voice in your phone, a set of suggestions in your feed, an answer that appears when you ask a question, and sometimes a refusal when you ask the “wrong” thing. Campbell Brown—once Meta’s news chief—has been pointing to a gap between those two worlds, and the gap matters because it shapes what people trust, what they fear, and what they believe AI is “for.”
Brown’s core observation is simple but uncomfortable: the conversation about AI among builders and policymakers can be fundamentally different from the conversation among consumers. The first is dominated by capability—what models can do, what systems can be built, what safety frameworks might prevent, and what policy could require. The second is dominated by outcomes—what AI actually says, how it behaves in messy real-world contexts, whether it seems fair, whether it seems honest, and whether it helps or harms in ways that are visible to non-experts.
That difference sounds abstract until you consider how AI is deployed. Most people don’t interact with “AI” as a concept. They interact with products that wrap AI inside interfaces designed to keep them engaged: search results, recommendation engines, chatbots, summarizers, moderation tools, and ad targeting systems. Those interfaces don’t just deliver answers; they also deliver framing. They decide what gets surfaced, what gets emphasized, what gets omitted, and how quickly the system moves from uncertainty to certainty. In other words, the consumer experience is not only about the model’s intelligence—it’s about the editorial layer around it.
Brown’s perspective is especially relevant now because the industry is simultaneously racing ahead and trying to reassure the public. Companies want to demonstrate progress: better reasoning, faster responses, more personalization, more automation. At the same time, they’re trying to manage reputational risk: misinformation, bias, hallucinations, privacy concerns, and the broader question of whether AI is becoming a new kind of gatekeeper. But reassurance tends to be technical and procedural, while consumer skepticism tends to be experiential and emotional. People don’t feel “safe” because a company has a policy document; they feel safe when the system behaves consistently, corrects itself when it’s wrong, and doesn’t manipulate them.
This is where Brown’s “two conversations” framing becomes more than a rhetorical flourish. It suggests that the industry may be optimizing for the wrong audience. If Silicon Valley debates focus on what AI can do, but consumers focus on what AI says, then the metrics of success diverge. A model can be impressive in benchmarks and still be frustrating in daily use. A system can be compliant on paper and still feel untrustworthy when it refuses to answer, misrepresents sources, or confidently produces content that sounds plausible but isn’t.
The question “Who decides what AI tells you?” is therefore not just about who owns the technology. It’s about who controls the translation from raw model output into the final message you see. That translation includes training data choices, fine-tuning objectives, safety filters, retrieval systems, ranking algorithms, and product design decisions. It also includes the human and organizational incentives behind those choices. When the stakes are high—political information, health advice, financial guidance—those incentives can collide with the public interest.
Brown’s background gives weight to this point. As Meta’s news chief, he operated at the intersection of technology, journalism, and platform governance. That role is a reminder that “news” is not merely content; it’s a social function. It shapes public understanding, influences elections, and sets the agenda for what people think is important. When AI enters the news ecosystem—summarizing stories, recommending articles, generating commentary, or answering questions about current events—it doesn’t just add convenience. It changes the way information is curated and interpreted.
And curation is power. Even when AI is not explicitly “editing,” it can still act like an editor. Summaries compress complexity. Answers select which details matter. Explanations can smooth over uncertainty. Tone can imply authority. And when AI is integrated into platforms that already have strong incentives to maximize engagement, the system’s “editorial” tendencies can align with what keeps users clicking rather than what keeps them informed.
Consumers notice these tendencies quickly, even if they can’t name the mechanisms. They may not know whether a chatbot is using retrieval augmented generation, or how its safety classifier works, or what portion of its responses are constrained by policy. But they do know when the system seems to avoid certain topics, when it repeatedly steers them toward particular narratives, or when it provides answers that feel too confident. They also notice when the system fails in ways that are obvious: incorrect facts, missing context, or citations that don’t match what was asked.
That’s why Brown’s framing resonates: the consumer experience is shaped by trust. Trust is not a single variable; it’s built through repeated interactions. If AI is inconsistent—sometimes helpful, sometimes evasive, sometimes wrong in ways that look careless—users learn to treat it as unreliable. If AI is consistent but biased—always emphasizing one perspective, always downplaying another—users learn to treat it as manipulative. Either way, the relationship between user and system becomes adversarial, even if the user never intends to “fight” the technology.
Silicon Valley’s conversation often treats these issues as solvable engineering problems. Improve the model. Add guardrails. Better training data. More evaluation. More transparency. Those steps can help, but they don’t fully address the deeper issue: AI outputs are not neutral. They are shaped by objectives and constraints. Even “safety” can become a form of editorial control, deciding what is allowed to be said and under what conditions. Even “helpfulness” can become a form of persuasion, deciding what kind of answer is most likely to satisfy the user and keep them engaged.
Brown’s comments also highlight a timing mismatch. In tech circles, AI is discussed as a future transformation. In consumer life, AI is already here, and it’s already shaping behavior. That means the public is forming opinions based on real interactions, not on roadmaps. When companies talk about long-term governance, consumers are dealing with short-term consequences: a misleading summary that spreads, a chatbot that refuses to answer a legitimate question, a recommendation that amplifies harmful content, or a generated response that looks authoritative enough to be shared.
There’s another layer to the gap: language. Silicon Valley often uses terms like “alignment,” “robustness,” “evaluation,” and “mitigation.” Consumers experience AI in plain language: “It told me X,” “It refused Y,” “It recommended Z.” The same system can be described as “aligned” by engineers and described as “biased” by users. The difference is not only technical; it’s interpretive. People judge AI by whether it matches their expectations of truth, fairness, and respect.
This is why the “who decides” question is so loaded. If the answer is “the company,” consumers will ask: which part of the company? Product managers? Policy teams? Engineers? External auditors? Regulators? And if the answer is “the market,” consumers will ask: whose preferences are being optimized? Their own, or the platform’s business goals? If the answer is “society,” consumers will ask: which society, and through what process?
Brown’s unique contribution is to connect these questions to the lived reality of information consumption. News is a domain where people expect accountability. They expect corrections. They expect sourcing. They expect that if something is wrong, there is a mechanism to fix it. AI systems, by contrast, often behave like they are improvising. Even when they cite sources, the citations may not be verifiable in the way people expect. Even when they provide disclaimers, the overall presentation can still feel like a definitive answer. That mismatch between expectation and behavior is a trust problem.
So what does it mean to close the gap between the Silicon Valley conversation and the consumer conversation? It likely requires more than technical improvements. It requires a shift in how AI systems are evaluated and communicated.
First, evaluation needs to reflect the consumer experience, not just model performance. Benchmarks can measure accuracy on curated tasks, but consumer interactions are messy. Users ask follow-up questions. They ask ambiguous questions. They ask in emotionally charged contexts. They ask when they’re tired, distracted, or misinformed. They also ask when they’re trying to make decisions that affect their lives. If AI is evaluated only on clean inputs, it will still fail in the situations that matter most.
Second, transparency needs to be meaningful. “We have safety filters” is not the same as “here’s how the system behaves when it’s uncertain.” Consumers don’t need a full technical audit trail, but they do need signals that help them calibrate trust. That could include clearer uncertainty handling, better explanations for refusals, and more consistent citation practices. It could also include user controls that let people understand and adjust how the system responds.
Third, accountability must be operational, not symbolic. If AI makes mistakes, there should be a path to correction. If AI refuses, there should be a reason that is understandable and reviewable. If AI generates content that influences public discourse, there should be governance that treats that influence as consequential. In other words, the system should behave like it belongs in a society, not like it belongs in a lab.
Fourth, product design should stop treating AI as a magic oracle. The interface is part of the decision-making process. If the system presents answers in a way that implies authority, users will treat it as authoritative. If the system instead presents answers as drafts, hypotheses, or summaries with clear provenance, users will treat it accordingly. This is not about dumbing down AI; it’s about aligning presentation with epistemic reality.
Finally, the industry needs to listen to consumers