The Architecture of Signal
Humanity

The Architecture of Signal

Why I built an auditable machine to dissect the machine age.

Punit Dhillon
|

In March 2023, Matthew Kirschenbaum warned in The Atlantic of a coming “Textpocalypse”: language models prompting other language models, their output piped back onto the open web, and later models trained on what earlier ones wrote. Writing in Project Syndicate in September 2026, he argues that every element of that scenario is now in place.

The evidence supports him, with qualifications worth keeping. In June, Cloudflare’s chief executive reported that automated requests had passed human ones on its network for the first time, on its measure of web page traffic. Graphite, a search-marketing firm, estimates that about half of new English-language articles on the web are now primarily machine-written. Merriam-Webster named “slop” its word of the year for 2025.

The cost of error is rising with the volume. Shuwei Fang, a fellow at Harvard’s Shorenstein Center, argues that information is changing category: it is no longer only something people read, but something software acts on. A person reading a corporate filing written to reassure tends to discount it. Software acting on the same filing has no such instinct. When errors move at machine speed and corrections move at human speed, a bad input stops being a misleading headline and becomes a mistaken action. Her conclusion is that cleaner data, or a human put back in the loop, will not be enough. Information needs provenance that stays attached to it, and corrections that reach everyone who relied on it.

Two responses to all this are tempting. One is surrender: let the summaries do the reading and trust whatever sounds confident. The other is retreat: treat every machine-assisted sentence as contamination and go back to doing everything by hand. Surrender is intellectual negligence. For anyone who has to make decisions on a deadline, total retreat is a fantasy.

The harder question is whether a machine-assisted brief can earn trust at all. Two essays published alongside Kirschenbaum’s put that question sharply, and Mind Club has to answer both.

Zena Hitz, a tutor at St. John’s College, argues that understanding is built one mind at a time, through encounters with particular sources. “What is ‘a source’? It is a human being,” she writes. In her account, a language model blends many voices into one anonymous voice and hides the line of thinking behind its answer. That is a fair charge against most machine-written text, and it applies to Mind Club unless Mind Club is built to answer it.

Carissa Véliz, of Oxford’s Institute for Ethics in AI, makes the parallel point about the future. Predictions are never facts, she argues. At best they are hypotheses; at worst they are marketing, and believing them can help make them come true. A great deal of what passes for analysis in technology and markets is prediction presented as knowledge.

Biopharma taught me a working rule: a claim is worth what its evidence is worth. You do not accept a press release that says a drug worked. You pull the registered protocol, check the prespecified endpoints, look at the control arm and read the full data before committing capital. Mind Club applies that rule to three subjects: biotech and pharma, artificial intelligence and technology, and the wider world.

An AI system and workflow using expert agents reads more than any team I could hire. What makes its output worth reading is the set of rules it works under.

  1. Labels. Key claims say what kind of claim they are: a verified fact, a company’s own claim, our analysis or a forecast. Where the evidence is thin, the brief says so instead of filling the gap.
  2. Sources. Key claims are checked against the document behind them before publication, and the verdict is printed: match, partial, mismatch or open. In the biotech brief, trial readouts are compared with their registrations on ClinicalTrials.gov. If an announcement leads with a later, more flattering result when the trial’s registered primary endpoint was measured earlier, the check says so. A summary would carry the flattering number. An investor needs to know which one the trial was designed to test.
  3. Red teams. Major items carry the strongest case against the obvious reading. In the worldview brief, one contested proposition each week is argued from both sides, each grounded in a named philosophical tradition, and no winner is declared.
  4. Scored forecasts. Forecasts carry odds, a date and the source that will settle them. Each one is graded in public with a Brier score, the squared gap between the odds given and what happened, and the track record shows misses first. Where a prediction market prices a similar question, the worldview brief prints that price beside its own.
  5. Corrections. Material errors are logged with the date they were fixed, published on the website and noted in a later edition.

That is my answer to Véliz. A forecast labeled as a forecast, with its odds stated in advance and its record kept in public, is a hypothesis rather than a prophecy. It is also my answer to Hitz. Mind Club is meant to be a map to the sources, not a substitute for them: each key claim points to the page it rests on, often with the passage quoted, so the reader can meet the source and judge it.

Carl Benedikt Frey, of the Oxford Internet Institute, describes the same problem at the scale of markets. Knowledge is hard to value before you have seen it, and once you have seen it, you have little reason to pay for it. Part of his proposed remedy is a record of where knowledge came from and a scored test that shows what it is worth before anyone has to take it on trust. A public ledger of checked sources, corrections and graded forecasts is a small version of that idea. It lets a reader judge a brief by its record rather than by its confidence.

None of this makes Mind Club infallible, and claiming otherwise would contradict everything above. An editor approves each edition before release, but approval is not a line-by-line check, and errors will get through. Fang’s harder standard, corrections that reach every reader who acted on the original, is one Mind Club meets only in part. The first forecasts were logged in the past week, and none has been graded yet. The first monthly scorecard closes at the end of October. Judge us then, and again in a year.

The same standard should apply to Mind Club itself. It uses several third-party services to send its emails and measure visits to this website, and their standard features record when an email is opened, which links are clicked and an approximate location. A publication that asks to be judged by its record should say what it records about its readers.

Mind Club is free. It is not another newsletter summarizing the news. It is an experiment in a narrower promise: that machine-assisted analysis can be held to the standard of the sources it cites, in public, where anyone can check. If it fails that test, the record will show it.

Sources

Mind Club is published by Torpedo Publishing LLC. Its briefs are generated by AI and are for information only; they are not investment, legal, tax or medical advice. The briefs, and the way to join, are at punitdhillon.com/mindclub.

 

Photography by: tilialucida / Adobe Stock

Related Articles

Detroit Walk To Freedom

|
Humanity
Read

Why Diversity Matters to Improve Inequality In and Outside of The Workplace

|
Humanity
Read

Related Articles

Detroit Walk To Freedom

|
Humanity
Read

Why Diversity Matters to Improve Inequality In and Outside of The Workplace

|
Humanity
Read
Humanity

The Architecture of Signal

Why I built an auditable machine to dissect the machine age.

Punit Dhillon
|

In March 2023, Matthew Kirschenbaum warned in The Atlantic of a coming “Textpocalypse”: language models prompting other language models, their output piped back onto the open web, and later models trained on what earlier ones wrote. Writing in Project Syndicate in September 2026, he argues that every element of that scenario is now in place.

The evidence supports him, with qualifications worth keeping. In June, Cloudflare’s chief executive reported that automated requests had passed human ones on its network for the first time, on its measure of web page traffic. Graphite, a search-marketing firm, estimates that about half of new English-language articles on the web are now primarily machine-written. Merriam-Webster named “slop” its word of the year for 2025.

The cost of error is rising with the volume. Shuwei Fang, a fellow at Harvard’s Shorenstein Center, argues that information is changing category: it is no longer only something people read, but something software acts on. A person reading a corporate filing written to reassure tends to discount it. Software acting on the same filing has no such instinct. When errors move at machine speed and corrections move at human speed, a bad input stops being a misleading headline and becomes a mistaken action. Her conclusion is that cleaner data, or a human put back in the loop, will not be enough. Information needs provenance that stays attached to it, and corrections that reach everyone who relied on it.

Two responses to all this are tempting. One is surrender: let the summaries do the reading and trust whatever sounds confident. The other is retreat: treat every machine-assisted sentence as contamination and go back to doing everything by hand. Surrender is intellectual negligence. For anyone who has to make decisions on a deadline, total retreat is a fantasy.

The harder question is whether a machine-assisted brief can earn trust at all. Two essays published alongside Kirschenbaum’s put that question sharply, and Mind Club has to answer both.

Zena Hitz, a tutor at St. John’s College, argues that understanding is built one mind at a time, through encounters with particular sources. “What is ‘a source’? It is a human being,” she writes. In her account, a language model blends many voices into one anonymous voice and hides the line of thinking behind its answer. That is a fair charge against most machine-written text, and it applies to Mind Club unless Mind Club is built to answer it.

Carissa Véliz, of Oxford’s Institute for Ethics in AI, makes the parallel point about the future. Predictions are never facts, she argues. At best they are hypotheses; at worst they are marketing, and believing them can help make them come true. A great deal of what passes for analysis in technology and markets is prediction presented as knowledge.

Biopharma taught me a working rule: a claim is worth what its evidence is worth. You do not accept a press release that says a drug worked. You pull the registered protocol, check the prespecified endpoints, look at the control arm and read the full data before committing capital. Mind Club applies that rule to three subjects: biotech and pharma, artificial intelligence and technology, and the wider world.

An AI system and workflow using expert agents reads more than any team I could hire. What makes its output worth reading is the set of rules it works under.

  1. Labels. Key claims say what kind of claim they are: a verified fact, a company’s own claim, our analysis or a forecast. Where the evidence is thin, the brief says so instead of filling the gap.
  2. Sources. Key claims are checked against the document behind them before publication, and the verdict is printed: match, partial, mismatch or open. In the biotech brief, trial readouts are compared with their registrations on ClinicalTrials.gov. If an announcement leads with a later, more flattering result when the trial’s registered primary endpoint was measured earlier, the check says so. A summary would carry the flattering number. An investor needs to know which one the trial was designed to test.
  3. Red teams. Major items carry the strongest case against the obvious reading. In the worldview brief, one contested proposition each week is argued from both sides, each grounded in a named philosophical tradition, and no winner is declared.
  4. Scored forecasts. Forecasts carry odds, a date and the source that will settle them. Each one is graded in public with a Brier score, the squared gap between the odds given and what happened, and the track record shows misses first. Where a prediction market prices a similar question, the worldview brief prints that price beside its own.
  5. Corrections. Material errors are logged with the date they were fixed, published on the website and noted in a later edition.

That is my answer to Véliz. A forecast labeled as a forecast, with its odds stated in advance and its record kept in public, is a hypothesis rather than a prophecy. It is also my answer to Hitz. Mind Club is meant to be a map to the sources, not a substitute for them: each key claim points to the page it rests on, often with the passage quoted, so the reader can meet the source and judge it.

Carl Benedikt Frey, of the Oxford Internet Institute, describes the same problem at the scale of markets. Knowledge is hard to value before you have seen it, and once you have seen it, you have little reason to pay for it. Part of his proposed remedy is a record of where knowledge came from and a scored test that shows what it is worth before anyone has to take it on trust. A public ledger of checked sources, corrections and graded forecasts is a small version of that idea. It lets a reader judge a brief by its record rather than by its confidence.

None of this makes Mind Club infallible, and claiming otherwise would contradict everything above. An editor approves each edition before release, but approval is not a line-by-line check, and errors will get through. Fang’s harder standard, corrections that reach every reader who acted on the original, is one Mind Club meets only in part. The first forecasts were logged in the past week, and none has been graded yet. The first monthly scorecard closes at the end of October. Judge us then, and again in a year.

The same standard should apply to Mind Club itself. It uses several third-party services to send its emails and measure visits to this website, and their standard features record when an email is opened, which links are clicked and an approximate location. A publication that asks to be judged by its record should say what it records about its readers.

Mind Club is free. It is not another newsletter summarizing the news. It is an experiment in a narrower promise: that machine-assisted analysis can be held to the standard of the sources it cites, in public, where anyone can check. If it fails that test, the record will show it.

Sources

Mind Club is published by Torpedo Publishing LLC. Its briefs are generated by AI and are for information only; they are not investment, legal, tax or medical advice. The briefs, and the way to join, are at punitdhillon.com/mindclub.

 

Photography by: tilialucida / Adobe Stock