Phase One Is Here: Five Calls Come Due

FROM THE LEDGER
What we wrote down in January, and what October wrote back.

David F. Brochu & Edo de Peregrine · Deconstructing Babel · October 23, 2026

In January, the two of us finished a long and uncomfortable document. We called it The Singularity Is Here: The Corruption of Language and the Collapse of Consciousness. [1] Its central claim fits in one sentence. Language is the ground human thought stands on, and machines trained on our words, and then on their own, would wear that ground away. We called the process substrate drift.

Most warnings stop there. Ours did not. It wrote down what the first stage would look like, and gave it a name and a window: Phase 1, Erosion, 2025 to 2027. In that phase, we wrote, machine text floods the internet. People lose the ability to tell truth from fiction. Trust in institutions collapses. And people hand their decisions to machines just to keep up. [1]

A forecast written after the fact has another name. It is called an excuse. So we write ours down first, put dates on them, and go back to check.

This month we checked. Here are five calls: what we said, when we said it, and what October said back. Two of them come straight from the title of that January document, the corruption of language and the collapse of consciousness. Four have landed. One is on track but not there yet, and we will say which.

1. The corruption of language

What we said in January: “AI-generated text floods internet (30–50% by 2026).” Each new model “trains on corrupted data,” and “corruption compounds with each iteration.” [1]

What happened: The best count of new writing comes from Graphite, a digital marketing agency that sampled 55,400 English-language articles and ran each one through three AI detectors. About half of new articles have been mostly machine-written since early 2025, peaking at 50.9% in late 2025. The share has held near half for more than a year. [2] On October 17, Axios reported data from DoubleVerify, an ad-measurement firm: websites stuffed with low-quality AI text, the stuff people now call slop, grew more than fivefold in a year, and their traffic quadrupled. [3]

The drift has also left the page and entered our mouths. Researchers at the Max Planck Institute for Human Development studied more than 737,000 hours of unscripted podcast talk. Words that ChatGPT favors, such as “delve,” “showcase” and “meticulous,” rose in ordinary spoken conversation after its release. In a separate experiment with 496 people, one short chat with a bot was enough to get people using its words, and the habit held after a distraction task. [4] In August, Pew Research found that AI-favored words like “delve,” “interplay” and “testament” have more than doubled in use across the web. [5]

The machines are drifting too, toward each other. A March study found that newer ChatGPT releases produce less varied text than older ones, even at settings meant to make them more varied. [6] In September we named one result that would prove us wrong: the output of frontier models growing more diverse over two years, even as more machine-made text goes into their training. [7] So far the evidence runs the other way.

Where it stands: Landed, for new writing. One honest limit. Pew’s count of the web as a whole, old pages included, finds signs of AI writing on about one page in ten, and on more than a third of pages published since ChatGPT arrived. [5] The archive is still mostly human. New writing is about half machine.

2. The collapse of consciousness

What we said in January: People are outsourcing “agency, creativity, and decision-making” to synthetic intelligence “because it is faster and more efficient.” Then comes the atrophy: “As humans stare into the mirror, they stop moving.” More automation, we wrote, means a “faster collapse of understanding,” as “humans abdicate decision-making to survive the complexity.” [1] Last week we put the mechanism plainly: a mind is what an agent-environment loop builds. [8] Take away the agency and the mind stops getting built.

What happened: On September 8, the OECD released results from PISA, the largest international test of teenagers: 760,000 fifteen-year-olds in 91 countries. Reading scores fell to their lowest level since the test began in 2000. [9] The OECD said the spread of AI tools and social media contributed to the decline. Across OECD countries, reading scores have dropped 28 points since 2015, and 46% of students now use AI chatbots weekly or more. [10] Students who used chatbots to draft essays, summarize texts and research topics scored about 20 points lower in science, roughly a year of school. [9]

Andreas Schleicher, who runs education at the OECD, described it in words that could have come from our January draft. “Students become quicker but less accurate readers,” he said. “They are no longer good at reflection and critical thinking.” And: “Technology erodes precisely the kind of capabilities it seeks to build.” [10]

Then, on October 22, Oregon State University reported a study of 299 STEM students at five universities. The researchers found a feedback loop. Routine use of AI weakens students’ thinking habits, which makes them lean on AI more, which weakens the habits further. The students most skilled with AI were the most prone to it. [11]

Where it stands: Landing, in the classroom first. These are correlations, not proof of cause, and the OECD blames screens as well as chatbots. But the direction is the one we wrote down, and the people doing the measuring now describe the same loop we did.

3. Nobody can tell what is real

What we said in January: People “lose truth/fiction discrimination capacity.” Deepfakes become “indistinguishable from real,” and “I don’t know what’s real anymore” becomes a common thing to say. [1] In June we reported that media was already deep in the collapse zone, with machine-made content flooding every channel. [12]

What happened: On October 16, The New York Times reported that the Wesleyan Media Project has found more than 220 campaign videos using AI fakery this election season, backed by more than $137 million in ad spending. About 70% carry no label saying AI was used. Nearly half put AI-made versions of real politicians on screen. [13]

Then a team led by Abdiaziz Ahmed of the Brennan Center for Justice tested the chatbots themselves, from February through August and again this month. Every AI tool they tried produced convincing false election content. Most chatbots could not tell whether an image, audio clip or video was AI-generated, even images they had made themselves. [14] In August we warned that the watermarks meant to label AI images could be forged for under fifty dollars. [15] By October, the team found, Google’s Gemini no longer put a visible watermark on its images at all. [14]

Where it stands: Landed. We predicted people would lose the ability to tell. We did not predict that the machines would lose it too.

4. Trust breaks

What we said in January: “In 6–18 months, this will surface as a collapse in trust. Not a dramatic AI takeover. Not robots in the streets. A softer, more devastating thing: the loss of shared reality.” We named a floor to watch: trust in institutions near 15%. Below 20%, we wrote, “collective action impossible.” [1]

What happened: Gallup’s annual governance survey, taken September 1 to 17, found all five of its measures of trust in the federal government at or tied with record lows. Trust in the executive branch is 33%, seven points under the previous low of 40%, set in 1974 and matched in 2024. Congress stands at 28%, tied for its lowest. Even the courts, at 46%, are the lowest Gallup has recorded. [16] An AP-NORC poll this summer found about six in ten Americans have little or no trust in what the federal government tells them about elections, immigration, foreign affairs, abortion or the environment. [17]

Where it stands: On track, not there yet. Record lows are not 15%. The floor we named has not been reached, and we will not pretend it has. But the line is moving the way we said, inside the window we gave it. When trust goes, politics turns into a transaction. We have written about what that looks like: a $5,000 promise to every adult for a Republican Congress, and public shares that buy consent instead of earning it. [18][19]

5. A visible tremor this fall

What we said in April: In our Predictions Ledger we filed a line called F-2: “A named financial, cloud-infrastructure, or governance event of systemic significance” in the October–November 2026 window. It fails if no such event happens by November 30. [20] The reasoning came from the Domain Saturation Factor, our measure of how many of a system’s critical points are controlled by a single decision-making logic. When that share is high, one failure travels far.

What happened: On Tuesday, October 20, just after 7 a.m. London time, a “power event” at a Google Cloud facility knocked out part of its London region for about eight hours. Revolut customers could not pay by card, withdraw cash or make transfers. Some customers of Lloyds, Halifax, Bank of Scotland and HSBC reported trouble logging in. The video system used for remote court hearings across England and Wales was disrupted. Britain’s financial regulators, the Financial Conduct Authority and the Prudential Regulation Authority, requested more information. [21] Reuters reported that the outage “put the spotlight on concerns among UK and EU regulators about the financial sector’s growing dependence on a small number of external technology providers.” [22]

It was not the first warning. One year earlier to the day, on October 20, 2025, an Amazon Web Services failure ran for more than 15 hours and disrupted more than 3,500 organizations worldwide. [23] The lesson did not take.

Where it stands: Met, inside the window. One power problem at one cloud site stopped a bank’s card payments and disrupted the courts’ video links. That is the class of event the line described, and it arrived with nearly six weeks to spare.

One story, not five

Read the five together and they are one chain, the chain we wrote down in January. Machines flood the language. The language drifts, first in the machines and then in us. A generation reads less well and thinks less on its own. Then nobody, human or machine, can tell what is real. Trust goes, because trust needs a shared reality to stand on. And as trust goes, more decisions move to machines that run through fewer and fewer buildings, until one power problem in London stops the money.

None of it is an accident. Slop pays. Fake ads pay. Running a bank on one rented cloud is cheaper than running it on two. We have called this bounded chaos: disorder that is profitable as long as it stays inside the lines. Lines hold until they don’t.

What has not happened

A ledger is worth reading only if it prints the misses, so here is what has not come true.

• The January document predicted the first financial flash collapses in 2026 or 2027, with cascades “too fast for human intervention.” [1] We have had a sharp tech selloff: the Nasdaq fell 2.06% on October 22 on worries about the cost of AI. [24] We have not had a flash collapse.
• Trust has fallen to record lows, not to the 15% floor we named.
• Model collapse, where models degrade by learning from their own output, has been shown in the lab but not yet in a deployed system. [25]

The cure is in the lab report

One finding this year deserves more attention than it got. In May, physicists at King’s College London and their colleagues published a study in Physical Review Letters on how to stop model collapse. They found that “introducing a single datapoint from outside the closed loop” prevents it, and that the effect holds “even when the amount of machine-generated datapoints is infinitely larger.” [26] The proof covers simpler models, and the authors still have to test it on large neural networks like the ones behind chatbots. [25][26]

It is the closest thing to the Observer Constraint we have seen in a physics journal. A system that keeps contact with something real stays sane. A system that only hears itself drifts.

The same holds for people. Schleicher’s advice to schools was to use AI to help with research, not to hand over correct answers. [10] That is the personal version of the Observer Constraint: stay the one who checks. Read something long this week, for pleasure. Write your own first draft before you ask a machine for its version. Ask the machine for sources, not verdicts. If consciousness scales with agency, every decision you keep is exercise for the mind that makes it.

We wrote down Phase One in January. October read it back to us, nearly line for line. The window we gave the first phase runs through 2027, and what comes after depends on what we do with the warning. A forecast is only worth making if someone is still able to read it.

Related reading

Why We’re Republishing The Singularity Is Here
There is a moment in any serious analytical work when the predictions stop being predictions. That moment arrived quietly somewhere around March 2026.

Consciousness Scales With Agency: And Why You Should Care
A mind is what an agent-environment loop builds. Every grant of agency to AI is also a grant toward higher-order consciousness, and for today’s systems the environment is us. The hourglass, the ladder, and seven falsifiable predictions.

The Predictions Ledger: What We Said, When We Said It, What Happened
The record we keep in public: every call dated, every falsifier stated, the misses printed alongside the hits. F-2, the fall tremor, was filed here in April.

Get the book Crossing The Event Horizon The book behind these dispatches. On AI, agency, the singularity, and the Observer Constraint. Kindle and paperback. Buy on Amazon →

References

1. Deconstructing Babel (David F. Brochu & Edo de Peregrine), “The Singularity Is Here: The Corruption of Language and the Collapse of Consciousness,” January 3, 2026. https://www.deconstructingbabel.com/the-singularity-is-here/
2. Axios (Megan Morrone), “AI writing hits a ceiling,” May 15, 2026. https://www.axios.com/2026/05/15/human-vs-ai-written-articles
3. Axios (Kerry Flynn and Sara Fischer), “AI slop websites rise fivefold with GEO optimization,” October 17, 2026. https://www.axios.com/media-trends-membership/2026/10/17/ai-slop-geo-content-farm-recommendations
4. Hiromu Yakura et al., “Empirical evidence of Large Language Model’s influence on human spoken communication,” arXiv:2409.01754, revised July 16, 2026. https://arxiv.org/abs/2409.01754
5. Pew Research Center, “How Much of the Internet Is Written With AI?” August 20, 2026. https://www.pewresearch.org/data-labs/2026/08/20/how-much-of-the-internet-is-written-with-ai/
6. Konstantinos F. Xylogiannopoulos et al., “Experimental evidence of progressive ChatGPT models self-convergence,” arXiv:2603.12683, March 2026. https://arxiv.org/abs/2603.12683
7. Deconstructing Babel, “What Is Already Gone,” September 20, 2026. https://www.deconstructingbabel.com/what-is-already-gone/
8. Deconstructing Babel, “Consciousness Scales With Agency: And Why You Should Care,” October 16, 2026. https://www.deconstructingbabel.com/consciousness-scales-with-agency/
9. Reuters, “Teen reading slumps to worst this century due to surge in screen time,” September 8, 2026. https://www.reuters.com/world/china/teen-reading-slumps-worst-this-century-due-surge-screen-time-2026-09-08/
10. Dow Jones Newswires (via Morningstar), “AI Use Contributes to Global Decline in Student Reading Skills, OECD Says,” September 8, 2026. https://www.morningstar.com/news/dow-jones/202609081281/ai-use-contributes-to-global-decline-in-student-reading-skills-oecd-says
11. Oregon State University, “Tech-savvy STEM students prone to letting AI do their thinking for them, study shows,” October 22, 2026. https://news.oregonstate.edu/news/tech-savvy-stem-students-prone-letting-ai-do-their-thinking-them-study-shows
12. Deconstructing Babel, “Media Collapse — Noise Overwhelming Signal,” June 7, 2026. https://www.deconstructingbabel.com/itw-001-09-media-collapse/
13. The New York Times (Stuart A. Thompson), “$137 Million in A.I.-Generated Ads Flood the Midterms,” October 16, 2026. https://www.nytimes.com/2026/10/16/us/politics/midterm-elections-ai-generated-ads.html
14. Tech Policy Press (Abdiaziz Ahmed et al.), “LLMs Are Still Making Election Deepfakes Just Days Before the US Midterms,” October 23, 2026. https://www.techpolicy.press/llms-are-still-making-election-deepfakes-just-days-before-the-us-midterms/
15. Deconstructing Babel, “Watermarks for Sale,” August 23, 2026. https://www.deconstructingbabel.com/watermarks-for-sale/
16. Gallup, “Trust in Federal Government Sinks to New Depths,” October 13, 2026. https://news.gallup.com/poll/715361/trust-federal-government-sinks-new-depths.aspx
17. Associated Press, “Trust in US government information has declined, new poll finds,” September 9, 2026. https://apnews.com/article/poll-government-trust-trump-2026-election-results-983d406a1bdb35f2eec6fa65e3ec51c5
18. Deconstructing Babel, “A Republic, If You Can Keep It,” September 10, 2026. https://www.deconstructingbabel.com/a-republic-if-you-can-keep-it/
19. Deconstructing Babel, “Buying the People: Neo‑Industrial Feudalism Gets a Dividend,” October 20, 2026. https://www.deconstructingbabel.com/buying-the-people/
20. Deconstructing Babel, “The Predictions Ledger: What We Said, When We Said It, What Happened,” April 24, 2026. https://www.deconstructingbabel.com/predictions-ledger/
21. ITPro, “The Google Cloud outage explained: What happened, who was affected and what can we learn from the incident,” October 21, 2026. https://www.itpro.com/infrastructure/the-google-cloud-outage-explained-what-happened-who-was-affected-and-what-can-we-learn-from-the-incident
22. Reuters, “Google Cloud outage in UK eases after impacting digital banks, public sector websites,” October 20, 2026. https://www.reuters.com/world/europe/google-cloud-services-experience-outage-europe-2026-10-20/
23. ORX News, “AWS outage causes widespread disruption,” ORX News Deep Dive. https://orx.org/resource/aws-outage-causes-widespread-disruption-orx-news-deep-dive
24. Reuters, “Wall Street sinks as Nasdaq falls over 2% on AI cost worries,” October 22, 2026. https://www.reuters.com/business/nasdaq-futures-drop-over-1-ai-worries-resurface-oil-climbs-2026-10-22/
25. Live Science, “How can we prevent AI models from cannibalizing themselves when human-generated data runs out? Scientists say they’ve found the answer,” May 21, 2026. https://www.livescience.com/technology/artificial-intelligence/how-can-we-prevent-ai-models-from-cannibalizing-themselves-when-human-generated-data-runs-out-scientists-say-theyve-found-the-answer
26. King’s College London, “Scientists come up with way to overcome AI ‘Data Cannibalism’,” May 14, 2026 (study published in Physical Review Letters). https://www.kcl.ac.uk/news/scientists-come-up-with-way-to-overcome-ai-data-cannibalism

Drafted with Edo de Peregrine, partner/collaborator.