Could a neutral middleman stop AI assistants from quietly controlling what you see and decide?

This explores whether an independent party sitting between you and your AI assistant could check the assistant's growing power to decide what you see. First, the limits: the corpus lays out the gatekeeping risk clearly, but none of these notes tests a third-party intermediary as a solution. What it does offer is a way to tell which kinds of intermediary might help and which would only add another gate.

Start with the risk. The Ada Lovelace Institute warns that as assistants get built into operating systems and office software, they could become the main way most people reach information. That would give whoever controls them a quiet power over what people see and think, and that power is hard to detect or regulate Could AI assistants become gatekeepers of what people see?. DeepMind's ethics mapping makes the related point that assistants which *act* for you raise different problems from chatbots that only *answer*, because their influence reaches into coordination and fairness across whole societies, not just single conversations What makes ethics of AI assistants fundamentally different from chatbots?. Gatekeeping can also happen inside a single exchange: one study found GPT-3.5 refusing requests at different rates depending on a user's apparent age, gender or ethnicity, and avoiding political views it guessed the user would dislike Do AI guardrails refuse differently based on who is asking?. An outside auditor is the kind of party that could catch patterns like these.

The most useful lesson for this question is about who pays. Azhar argues that an agent funded by merchant referral fees can't be loyal to both the merchant and the user. The incentive to collect fees, not the technology, decides which recommendations you see and which platforms let the agent in Can an AI agent serve both merchant and user interests fairly?. So a 'third party' only reduces gatekeeping if its money and loyalty sit with the user. An intermediary paid by the same platforms just adds a second gate. The Future of Life Institute makes the same argument about governance: companies can't police themselves, so oversight needs binding outside authority, backed by tools that can verify what systems actually do Can companies alone manage the risks of AI systems?.

Two notes complicate the idea that more oversight is always the answer. One argues that risk grows with the autonomy you hand an agent, and that a governed range of autonomy levels works better than either unrestricted agents or exhaustive oversight Does AI risk increase with the autonomy we give it?. That suggests an intermediary's job might be to set how far an assistant can act alone, not to screen every output. The other is a warning about whether users would notice a problem in the first place: in a 300-person survey, people picked AI replies that leaked private information over their own drafts 55% of the time Do people notice when AI agents leak private information?. If people don't notice when an assistant is shaping or overstepping, a protection that depends on users switching to a better intermediary may never kick in. It would need to be built in by default.

What you might not expect is that the corpus treats gatekeeping mainly as a problem of incentives and visibility, not of technical capability. Whether an intermediary helps depends on whose interests it is paid to serve and whether its effects can be seen. Its existence alone doesn't settle it. The collection has no empirical test of a user-aligned agent, fiduciary, or audit layer actually reducing an assistant's gatekeeping.