The Proof of Thought network
AI has broken the once (admittedly tenuous) link between fluency and proof-of-human-thought. While a fluent piece of writing can still mean a lot of thought went into the content, it can also mean a one-line prompt did. A reader has no (easy) way to tell whether that fluency is born of extensive thought or not.
The Proof of Thought network is a way to help solve this problem. It is a directory of writers whose long-form work has been verified (to the best of our abilities) as their own thinking, as at the date shown. There are still readers out there who want to find a human mind behind the writing and writers who do the thinking deserve a way to be found.
To become a verified PoT writer, there are two tiers of verification:
- Clear signal (first filter): Verified that recent writing is not obscured by AI
- Proof of Thought (second filter): Verified live thinking. Once a clear signal has been established (again, using the only tools we currently have available), any recent live demonstration of thinking linked to that writer will garner them the Proof of Thought qualification.
Clear signal: Verified that recent writing is not obscured by AI
As soon as AI detection is registered on a piece of writing, the signal between fluency and proof of thought is immediately obscured. For this reason, the baseline for being listed in the PoT network is that a writer's work passes independent AI-detection scans on four out of their five most recent free-to-view long-form pieces of content, published by the date shown. (It's not a perfect system, and the 20% room for error is to handle the rare cases where an essay is falsely flagged as being partially assisted by AI.)
The criteria for verification are:
- Articles clear: Four out of five most recent long-form articles are scored as fully human-written by external Pangram scans (no affiliation). If a writer doesn't have five recent long-form articles, that's recorded as an insufficient long-form sample (neither pass nor fail). Only the body text is scanned: no images, captions and quoted or third party text is included (so block quotes can't trip a false flag against the writer).
- Notes clear: Eight out of ten of their most recent notes over 100 words are scored as fully human-written by external Pangram scans. If a writer doesn't publish notes, or has fewer than ten notes over 100 words, this criterion is not applicable.
Limitations
- Pangram is one of the best-validated AI detectors, but not infallible. In a recent study (Graphite 2026), 1.8% of pre-ChatGPT articles were wrongly judged mostly AI-written. Any writer who falls into this rare category need only message @hereisyourbrain on Substack and their next five long-form posts will be scanned instead.
- This network in no way wants to bias against non-native writers, which is another reason Pangram was selected as the AI detector of choice. The 2023 study that showed a bias against non-native writers is now three years old and tested older detectors, not Pangram. Pangram's own testing on that study's essays found no false positives (Emi & Spero 2024).
Proof of Thought: Verified live thinking
The higher level is a casual conversation by the writer about some of their published work and the thinking behind it. It can be a one-on-one question-and-answer or a conversation with another person. Most Substack Lives are great examples of this.
What the directory is (and isn't)
- This directory of writers whose work is a clear signal of thought (some of whom have also demonstrated this thinking, live) is incomplete and non-exhaustive and is constantly being added to. It contains only writers who have been verified.
- It is a time-stamped snapshot, not a permanent certificate. Each entry reflects checks and any live conversation completed on the dates shown. Writers are not continuously rescanned, but random writers are selected for “spot checks,” with their PoT date stamp being updated accordingly (or in rare cases revoked, if the signal is found to be obscured).
- Absence does not mean AI authorship. It could mean a writer has not been nominated yet, the signal is obscured, or the published work is multi-authored.
(If you're listed here and would rather not be, please message @hereisyourbrain on Substack from that account and I'll remove you. No reason needed.)
The donation pot
The donation pot pays the curators who find writers not yet listed in the network whose work is still a clear signal of thought. It sits untouched until the network has passed a thousand listed writers. (The rule for dividing it will be published in advance. The founder never receives any of it.)
Calls
A call is a bet that an essay will go on to become a hit, and possibly even a diamond. A hit is an essay that has done better than a writer's usual, whereas a diamond is an essay that has beaten all but their recent best, and gone at least 20% past the hit line. Some key notes:
- A post that has already beaten a writer's usual at the time of calling must still grow meaningfully to be successful, and is marked as having kept climbing.
- No call can be made if the essay has already beaten the writer's best at the point of attempting to call it.
- Calls must be made within three days of an essay being published, and are resolved when the essay is a week old.
- Once a curator has called an essay, no other curator can make a call on it.
- Curators can make up to 30 calls per week.
- Calling is in beta. The lines will be recalibrated as the data grows; a change applies to calls made after it, and past calls keep the lines they were made under.
Taste score
Around half of all essays beat their writer's usual by chance, so a hit rate above chance (over enough calls) suggests the curator is seeing something others aren't – they have “taste.” (It takes approximately 20 calls to calibrate a curator's taste, so any score should be considered noisy until this point.)