Who decides what AI tells you: the gap between Silicon Valley and real users

95% of users who interact daily with AI tools take no part in any forum where the rules governing what those systems can or cannot tell them are decided. That distance is precisely the centre of the argument developed by Campbell Brown, former head of news policy at Meta, in an interview published by TechCrunch this week.

Brown sums it up with a phrase worth pausing over: "The conversation is sort of happening in Silicon Valley around one thing, and a totally different conversation is happening among consumers." It's not a new criticism, but it carries particular weight coming from someone who spent years on the other side of that table.

The core problem: who has a voice in these decisions?

When Anthropic publishes an update to Claude's usage policy, or when OpenAI adjusts the filters on its models, the process typically unfolds internally, with some consultation from safety groups, alignment researchers and, at best, external red teamers. The end user, the one using the system to search for medical information, draft legal documents or simply resolve everyday questions, has no clear channel to influence those decisions.

Brown knows that dynamic well. At Meta she oversaw editorial policies that determined which news stories were amplified and which were suppressed in the newsfeed. The analogy to AI isn't perfect, but it's not forced either: in both cases, a small group of people at a handful of companies decides what information reaches, and how it reaches, hundreds of millions of people.

What has changed with large language models is the scale and opacity. A recommendation algorithm at least has the constraint of working with pre-existing content. An LLM generates new responses in each conversation, turning every interaction into an implicit editorial decision: what nuance to use, what sources to mention, what perspectives to omit.

Why this matters now

May 2026 is no ordinary moment for this conversation. Claude Opus 4.7 already handles context windows of a million tokens and is routinely used in professional workflows that previously required human teams. Claude Code manages complete development pipelines through sub-agents and hooks. Real dependence on these tools has grown far faster than the accountability frameworks that should accompany it.

In that context, Brown's question, who decides?, shifts from philosophical to operational. If a company configures an agent based on Claude to answer customer inquiries, and that agent sidesteps certain topics by design, who bears responsibility for that omission? The developer writing the system prompt? Anthropic, which trains the model with its own values? The marketplace distributing the plugin?

The chain of accountability fragments precisely where it matters most that it remain clear.

Who this debate affects

This discussion impacts very different groups:

• Developers building products on Claude or any other LLM who need to understand what real margin they have to configure model behaviour against what's imposed by provider policies.
• Companies deploying agents in regulated sectors (healthcare, law, finance) who assume legal risks from responses they didn't fully control.
• Journalists and regulators who have spent months trying to build oversight frameworks without real access to the internal moderation criteria of major labs.
• End users, most of whom don't know that the responses they receive have passed through layers of filtering whose design was never subject to public consultation.

Brown doesn't offer concrete solutions in the interview, something that can be read as intellectual honesty or as a missed opportunity depending on your perspective. What she does do is point out that the industry has spent years promising participation mechanisms that aren't materializing at the pace the products themselves are being deployed.

---

At ElephantPink we've been observing this asymmetry for some time in the integration projects we support: decisions about what an agent can and cannot do are made in layers the end client never sees. The merit of voices like this one, even when they come from within the system, is keeping focus on a question the industry would prefer to push into the background.