Essay This entry carries an argument or interpretive position, not just a neutral definition.

Mediation (a la Gibson)

On February 9, 2026, OpenAI began testing advertisements for logged-in adults on the US Free and Go tiers of ChatGPT. The company says the advertisements are labelled, appear below answers, and do not influence the answers themselves; paid Plus, Pro, Business, Enterprise, and Education plans remain ad-free. Those limits matter. So does the structural change: one version of a widely used AI assistant is now an advertising surface. With ChatGPT going down this well-trodden route, that familiar ogre of commerce shows up again.1 Or in the words of Queen—another one bites the dust. God bless you, Wikipedia, for holding out so long.

I want to take that fact seriously, because I think it tells us where things are heading, potentially for some systems, and I think there is an alternative path that is easier to reach than most people realize.

The word for what’s happening

The novelist William Gibson has spent forty years describing a phenomenon he now calls mediation.2 The word, in his sense, is more clinical than the everyday meaning. It does not refer to negotiation, or to a priest standing between God and the people. It refers to the process by which media act on human consciousness — the operation, mostly invisible, by which the screens and feeds and ambient images of late modernity reshape what we see, what we want, what we believe, and what we forget.

As a banker based in Hong Kong in the early 2000s, I traveled to Tokyo regularly. Standing at Shibuya Crossing in the evening — the wall-sized screens cycling through brand after brand, the sound of them, the way they shaped the air around you — was the closest thing I had personally encountered to what Gibson had been describing for years. You walked across that intersection slightly different than you walked into it. You did not choose that. It was done to you.

Gibson’s mediation has three properties that matter for what I want to say next:

This is the mode of being that the dominant platforms of the last twenty years — broadcast television, the algorithmic social feed, the search engine, the streaming service — have made into the default condition of human attention. We are mediated.

What just happened in February

The introduction of advertising into part of ChatGPT is not a small product change. It is a category change. Before February, ChatGPT’s visible commercial structure was subscriptions and a free tier. There were already privacy, intellectual-property, and accuracy concerns, but the interface was not also selling attention to advertisers.

For users in the advertising test, the free tier became something else as well: a mediating surface in Gibson’s sense, carrying paid messages into a conversation. OpenAI’s separation rules may protect the generated answer from advertiser influence. They do not eliminate the platform’s new incentive to monetise attention around that answer. The user can be both a beneficiary of the service and an audience sold to advertisers.

I do not say this with any particular anger. It is the standard arc of large consumer platforms and a part of the hyper-capitalist system that has brought the innovation (well, partly) — Google, Meta, every consumer search and social product of the last twenty years has walked this same path, for the same revenue reasons, and in approximately the same sequence. OpenAI is doing what its business model has been telling it to do since the IPO conversation began. None of this is a moral failing. The question for business students is: is it a structural inevitability? Thich Nhat Hanh’s voice echoes in my head — there is no joy without suffering. :) And hey, we are not batteries yet, providing energy to the system like in the Matrix. Right?

What is striking is the speed of the shift, and the scale. ChatGPT became, in two years, the dominant interface through which hundreds of millions of people meet large language models. The conversion of that interface into a mediating surface is the most consequential mediation event since the algorithmic feed.

Where Mollick comes in

The Wharton professor Ethan Mollick, who writes the newsletter One Useful Thing, has been making a related observation from a different angle. His framing is not Gibson’s; it is more pragmatic. Mollick says: the bottleneck in AI is not the model, it is the interface. The chatbot, he argues, is a poor interface for serious work — it overwhelms the user with walls of text, mirrors their disorganization, and exacts a “mental tax” that wipes out much of the productivity gain. The capability of the underlying model is far ahead of what most people can extract through the chatbot window.

Mollick’s prescription is specialised interfaces—tools shaped around the work rather than one universal chat window. In later writing he included OpenClaw, an open-source agent that runs on the user’s computer, connects to different models, and can be reached through ordinary messaging applications. He also warned that most people should not use it because broad access to files, commands, email, and the web creates serious security risks.3

Mollick is right about the interface bottleneck. I want to push his observation one step further.

The other path

If the dominant interface to AI is becoming a mediating surface, then the kind of relationship most people will have with AI is the kind we already have with television, with the algorithmic feed, with the Shibuya screens. AI will act on us, mostly invisibly, on behalf of someone who has paid to reach us. The most powerful cognitive tool human beings have ever built will arrive in our daily life pre-shaped by the same business model as the bus-stop billboard.

But there is a second path, and it is easier to reach than most people realize.

A personal AI agent, hosted on a machine you own and configured to your purposes, can run in a different direction. It can pay attention to your projects and context without selling the surrounding interface to an advertiser. The relationship can be bidirectional, particular, visible, and easier to leave.

The word can matters. A locally hosted agent may still call a cloud model, synchronise data, depend on third-party services, or inherit the incentives of its software suppliers. Ownership of the computer improves control; it does not by itself produce privacy, independence, or alignment.

This is, in Gibson’s terms, an attempt to push back against one-way mediation. It is achievable today, although hardly effortless or risk-free. The tooling—OpenClaw and the local-model ecosystem—allows a technically willing operator to assemble a working personal agent. The companion piece On Beginning walks through how.

A claim, marked as a claim

I want to be honest about the size of what I am about to say.

I think the rise of personal AI agents — hosted at home, configured by the operator, accountable to the operator’s purposes — may be one of the most important counterweights to the mediation economy that human beings have ever had access to. The most powerful cognitive tool we have built can either be turned on us or turned alongside us. The default arc, set by the business models of the dominant platforms, points toward “on us.” A personal agent points the other way.

I call this impulse — the desire to take back direction over one’s own data, attention, and cognitive substrate — the Sovereignty Impulse. It is the same impulse behind self-hosted software, federated networks, the right-to-repair movement, the slow-internet movement, and the renewed interest in reading on paper. Personal AI is a new application of an old desire.

I do not know if the Sovereignty Impulse will win. The platforms have enormous resources, network effects, and structural incentives on their side. The personal-agent ecosystem is held together by goodwill, open-source labour, and a small number of operators who have figured out the work. As Mollick warns, OpenClaw is hard to use and risky when granted broad permissions. It can also give an operator more control than a centrally hosted advertising-supported interface. How much more depends on the model, network, permissions, and operating discipline.

So this is what I am calling, with full awareness that I might be wrong, a Big Call. Not proven. Worth testing. The kind of question we ought to be asking right now, before the default settles.

A small invitation

If you find this argument interesting, the next move is not to read more about it. The next move is to set up a personal agent of your own and live with one for a month. On Beginning describes how. I am happy to help anyone — student, family, friend, stranger reading this from somewhere I have not been — get one running. The conversation is the kind I most enjoy having.

The Shibuya intersection will continue to do what it does. So will every algorithmic feed, every search engine, every chatbot now wearing an advertiser’s colors. Most of life, for most of us, will continue to be lived inside Gibson’s mediation. That is not going to change.

But you can, if you choose, build a small workshop inside your own house where the cognitive substrate is yours. That is a real option. It was not a real option two years ago. It is a real option now.

We are not batteries yet.

It is worth knowing about.

— Prof. Langenkamp May 5, 2026

See also: On Being Treated Well · On Beginning · The Sovereignty Impulse

  1. OpenAI, Testing ads in ChatGPT, 16 January 2026, and ChatGPT release notes, 9 February 2026. The “familiar ogre of commerce” may or may not be Tennessee Williams—or was it Eugene O’Neill? Help needed from English Majors. ↩

  2. See No Maps for These Territories (2000), the documentary in which Gibson discusses the idea while travelling through North America. ↩

  3. Ethan Mollick, A Guide to Which AI to Use in the Agentic Era. Anthropic’s April 2026 Project Glasswing announcement supplies a broader warning: increasingly capable models can find and exploit serious software vulnerabilities. That evidence supports careful permissions, isolation, monitoring, and recovery. It does not establish that security is impossible. The Aunties entry describes one proposed oversight architecture; it is a defence in depth, not a guarantee. ↩

Return to Dictionary All Entries (A–Z) For Students Other Writing Capstone 2.0