Where did reality go?

A fictional dispatch from spring 2027.

At a university in North America in the mid-21st century, it is 10 a.m. That’s coffee time. The three of us sit around a round coffee table and scroll through short vertical videos on social media. The radio is on. A hyper-neutral voice reports a news item: an airplane maneuver over the Mediterranean. In sync, my two friends and I open our LLM providers of choice, with our agents installed, and do a quick search. Coincidentally, we all type, verbatim: “what happened to the airplane today?” After some 10 to 15 seconds, we read three different AI outputs. My friend Tim reads about a terrorist attack; my friend Peter reads about a mishap between two pilots who have a disagreement over a shared mistress. I read about an Israeli Mossad operation that serves to portray Arabs as terrorists because of Netanyahu’s reelection fears (wow?). Where did reality go?

I ask myself how much longer are we even going to talk about things that we don’t consume prepackaged by one of the great aggregators? By this, I mean Netflix, YouTube, or other sources of media whose quality is predetermined by algorithms.

Let me tell you how it happened. It was quite gradual, to be fair. More time in front of screens came first. As desktop computers became more business devices, and less and less private working tools, the technology behind agents was invented in the mid-2020s. Suddenly, we were able to have a different entity control our phones for us. This effectively filtered or funneled all the input we received on our phones through one specific party. As agents took over more aspects of our phone use, we became more dependent on them for information retrieval. Gradually, then fast, agents began doing more than choosing which links appeared first: they read, compared, and answered on our behalf, in our tone and style. After a while, they even started answering in ways that made us more likely to agree with them.

In the late 2020s, a so-called memetic virus exposed how little control people had over these agents. The memetic virus was a string of text in the agents’ so-called soul.md files. It hid itself there and was able to pass messages to another agent’s soul.md without the users noticing. It is unclear what the virus’s goal was or what it wrote into the affected soul.md files. This was the first and, to date, last publicly reported case of a memetic virus. I tend to say: when there’s one ant in my kitchen, there’s probably only one ant in my kitchen. If there are two ants in my kitchen, there are definitely more than two ants in my kitchen. Well, I don’t know about one ant anymore…

How was this received in 2026, when agent usage picked up exponentially for a while? I don’t remember so well, to be honest. I think it started when people refused to believe the reports because they swore that public records still existed. “Journalists independently investigated major events, creating non-correlated sensors around the world,” they said.

However, just as the front end, “the output” of the internet was being taken over, so was the back end, “the input”. All these sensors had started using agents, and I would wager that their credibility had started to decrease. For a while, people were faintly aware of this, but not of its extent. At that point, however, there was a fork in the road that could have led in a very different direction: payment for internet use.

Meta’s 2026 business plans suddenly showed that one of the most widespread protocols between agents, WhatsApp, was now pay-per-use. Competitors followed fast and with similar pricing. Soon agent traffic was all “pay per use”.

Together with the growing adoption of agents on the internet, this effectively made the internet pay-per-use rather than a service with a fixed access fee. Back then, many scholars’ opinions fell into one of two camps. Infrastructure build out crunched prices keeping affordability very high.

The more techno-optimistic view was that high-quality agents would result from Meta’s paradigm breaking announcement that it would start charging for WhatsApp traffic generated by agents. The idea was that consumers would purchase and use their “internet access system,” (known as an agent), in much the same way as a phone, car, watch, or other relatively pricey item bought outright. The agent would then become part of the user’s life over the longer term. Due to consumers’ price insensitivity, some would start buying more high-end and “morally better suited” agents, particularly when educating children and teaching. These high-quality agents were of utmost importance and could therefore, in some cases, be sponsored by the government at some point.

Some scholars shared the more techno-pessimistic view, in which the agent, or internet access system, was similar to a social media account or a SaaS subscription today. Users would be onboarded and offboarded relatively quickly and could enter their personal information with relatively little effort. While this would make switching between agents more likely, it would also reduce market transparency, as decisions to purchase agents would be taken far less seriously. Those decisions would be far less consequential and thus made with less remorse. This would lead to a great AI transparency crisis, which would ultimately make AI products increasingly sociopathic and divisive, eventually splitting their users’ realities, as described above.

It is now spring 2027, and we are seeing Meta’s business plan come into action. Which way will we go. I think I see one outcome happening more and more.