Last month, I wrote about “context laundering” in AI summarization. Over the last few weeks, this issue has taken on some new dimensions.

In June, TIME began delivering Markdown versions of their website. This was done in response to the increase in bots and agents in scraping their website and is part of their larger, ongoing GEO project (generative engine optimization). GEO is like the old SEO (search engine optimization) but for AI bots. In July, TIME began including advertisements in the Markdown version. The theory is that these ads will affect the way that the advertised products are represented in AI search. In somewhat predictable fashion, Perplexity (an AI company) has announced that it is blocking these ads, proclaiming that they are “deceptive advertising.” Meanwhile, TIME isn’t the only site doing this, either.

This is just one skirmish in our emerging media environment. Several web security companies and experts (e.g. Cloudflare) now report that automated traffic amounts to more than half of the web requests, leading to a new, less conspiratorial version of “dead internet theory.” These are the kinds of searches that trouble media companies, who rely on humans visiting their sites, reading their articles, and clicking on their ads. Beyond this one encounter, the interaction among agents and media providers is multi-layered, with different internal and external agents accessing different slices of the media stack.

So how many AI agents will there be?

In his most recent “manifesto,” Zuckerberg imagines each of us having an AI assistant. What would the web look like with millions of AI assistants “assisting” Americans and interacting with the many corporate and institutional agents out there (and who knows what else)? To be clear, I think his story is AI hopium fan fiction, but even then I can’t see how it would end well. In this AI revolution, will corporations decide they no longer want to have advantage over their workers and customers? In the negotiations between “my assistant” and HR’s AI, am I going to get my raise? More likely the interactions will remain asymmetrical. HR will have more authority and access to more relevant data than my assistant will.

None of these developments initiated our post-truth condition. We already recognize the harms algorithmic social, mobile media has done to our world. In short, we know quite plainly that computational media technologies can do tremendous harm to us. AI agents increase the risks of these harms. At minimum they are another computational variable in our media environment whose operation produces outputs with indeterminate provenance, even when sources are cited. And that “truth” problem becomes much more complicated when we include other agents with competing aims. AI agents can be influenced by other agents in ways that are consequential but generally invisible to humans. We can seek to secure these connections, but in the end I think we are still left with this question.

How did this network of models, agents, retrieval systems, publishers, advertisers, LLMs, and humans cause this output to my prompt to be generated?

Unreliably. That would be my one-word answer. More to the point, AI generation has become networked output. We cannot think of our interactions with chatbots as prompting a model that generates an output. That oversimplification no longer serves us. And I think that may become a challenge for the workplace deployment of AI.

I’m thinking of a university with thousands of employees each with agents roaming the intranet for various tasks. This is not 5000 employees with Co-Pilot. First, there are also institutional agents. Those agents may report to a particular person/reverse centaur, but they have been instructed with institutional goals. My agent interfaces with institutional agents for scheduling, grading, budget requests, and so on: all the ways I once communicated for myself. If my agent is one provided by the university, then that is simply another institutional agent. I then require my own agent, programmed and paid for by me, that will interact with this institutional agent. The result is a new computational layer of governance and continuous inference.

However, as I have been saying here, this new layer is a zone of obscured nonhuman negotiation and conflict. In this context, output happens.