# 100 AI agents, 21 Tamil political YouTube channels: the full research (17 Nov 2026)

This folder is every file the test was measured from, so that each number in the video can be checked.
- The video: https://www.instagram.com/reel/Df2HfkXtKQM/
- The issue: https://politic.in/issues/iss_449c72442a5e065f
- All of it in one file: https://politic.in/media/research/iss_449c72442a5e065f/research.zip

## What was tested

The test asks one question: if 100 AI agents, each a newcomer to Tamil Nadu politics with no party leaning, read what
21 Tamil political YouTube channels said, which channels would the most agents follow?
- **The agents.** Each agent is one run of an AI model (Opus), given one profile and the same text. They are not
  people, and this is not a poll. Round 1 had all 100 agents (99 answers in form); rounds 2 to 6 had 20 each. The 100 profiles are in `profiles.md`; each trait is drawn in proportion to a
  published figure (Census 2011; a 2021 post-poll survey).
- **What they read.** They read text: YouTube's automatic captions of the channels' most-viewed political videos
  from 17 Sep to 16 Nov 2026. They did not watch the videos, and they did not browse YouTube.
- **How the method was fixed.** The method (`PLAN.md`) was written and fixed before any data of this round was
  read. Every change after that is in `DEVIATIONS.md`, with its time and its reason.
- **What happened.** `RECORD.md` says what happened, with every count.

## What came out

**The agents' result.** A channel's share is the share of agents who named it among the up to three channels they
would follow. Each round drew a fresh set of minutes; its score is the mean of six rounds, each weighted the same
(`viewers-six.json`).
- The top two are Chanakyaa (0.598) and Vikatan TV (0.566).
- Which of the two is first is not settled. The rule written before rounds 3 to 6 ran settles a leader only if it is
  first in at least 80% of 10,000 redraws of the rounds. Chanakyaa was first in 40.7% of them, and Vikatan TV in 38.9%.
- Counted answer by answer instead, over the 195 answers in form, Vikatan TV is in 145 and Chanakyaa in 110.
- Savukku Media Network scored 0.089 and Saattai 0.008.

**Views on real YouTube, among these 21** (not among all channels):
- **Measure 1, this month's political views in YouTube's own search** (`channels.json`): Savukku Media Network
  11,786,468; Arasiyal Sadhurangam 8,305,517; Chanakyaa 5,296,890; Saattai 4,005,776. In the whole search, which
  takes in TV news, Savukku Media Network is 5th of 247 channels.
- **Measure 2, the listed views of every upload from 17 Sep to 16 Nov 2026** (`window-views-deep.json`): long
  videos and Shorts together, political or not, with live streams not counted. Each figure is a range, least to most:
  - Galatta Voice: 95.46 to 95.51 million.
  - Saattai: 74.11 to 79.10 million.
  - Chanakyaa: 65.09 to 67.39 million.
  - Savukku Media Network: 59.6 to 66.2 million.
- By Shorts alone, which is not the measure, Saattai's place is not settled. The figures are in the same file.

**Minutes that stir feeling ("emotion") in every 100 readable minutes** (`marks.json`): the minutes charged with
anger, fear, pride or ridicule, by the label in `instructions-labels.md`.
- Chanakyaa has 8.3, the fewest of the 21, and Vikatan TV 28.8, the 4th fewest.
- Savukku Media Network has 57.5 and Saattai 74.4. The middle channel of the 21 is at 46.2, and Neerthirai is the
  highest at 80.3.

**What the record does not show: a link between emotion and views.**
- Over the 21 channels, the emotion mark goes strongly with the agents' picks: Spearman's rank correlation is −0.82.
- It goes with neither measure of views: +0.12 with measure 1, and −0.11 to −0.08 with measure 2.
- So the agents picked calm channels, and views show no pattern either way. When the video sets what the agents
  picked against what people watch, it describes those four channels. It does not show that people watch a channel
  because it stirs feeling.
- Chanakyaa, the calmest of the 21, is 3rd by measure 1 and 3rd or 4th by measure 2.

**The test's limits.**
- The agents read text only: no faces, no voices, and no fan's loyalty to any channel.
- Automatic captions mishear some words.
- Each channel was read as 8 random minutes a round, over six rounds.
- Some channels said their own name in the minutes read, and some agents recognised a channel (`RECORD.md`,
  section 8).
- Scaling the data and the compute might change the result.

## The files

- **The method:** `PLAN.md` (fixed first), `DEVIATIONS.md` (every change after), `RECORD.md` (what happened).
- **The channels, and why each is in:** `channels.json`.
- **The videos:**
  - every upload gone through, and what happened to it: `uploads.jsonl`;
  - the videos read: `kept.json`;
  - the tab listings: `listings/` and `listings-deep/`, without their titles;
  - the video pages' dates and counts: `pages-deep/`;
  - which videos had captions: `captions.json`.
- **The minutes and their marks:**
  - every minute's six marks, by video and time: `windows.jsonl`;
  - the totals: `marks.json` and `summary.json`;
  - the readers' instructions: `instructions-labels.md`, `instructions-political.md` and `instructions-title-screen.md`.
- **The agents:**
  - their profiles: `profiles.md`;
  - their instructions: `viewer-instructions.md`;
  - the minutes each round's agents read: `excerpts/` (round 1) and `excerpts-draw2/` to `excerpts-draw6/`;
  - each agent's answer, and the answers mapped to channels: `viewers/` and `viewers-draw2/` to `viewers-draw6/`;
  - the six rounds' scores and the redraws: `viewers-six.json`.
- **Measure 2:** `window-views-deep.json`.
- **The scripts that made each count:** `tools/`.

## What is withheld, and why

- **Creators' arrests and legal cases.** A title that may speak of a creator's arrest or legal case is withheld (19
  titles; DEVIATION 10), and so are the listings' title columns, which were never screened for this.
- **Minutes about a creator's case.** In round 2, five minutes the agents read speak of a creator's case; their text
  is withheld here. In rounds 3 to 6 such minutes were replaced before any agent read; their ids stay, and their text
  is withheld.
- **One sentence of one agent's answer.** In round 2, one agent's answer repeats what a creator said of his own time
  in jail. That sentence is withheld, and the place where it stood is marked.
- **The packets.** The packets given to the agents are rebuilt byte for byte by `tools/packets.py`, so they are not
  copied.
- **Reworded lines.** A few sentences were reworded for publication, where they described how a step or the list of
  named channels first reached the test. What each step was and why it was taken stays.
  - This touches 51 lines: 8 in `PLAN.md`, 22 in `DEVIATIONS.md`, 11 in `RECORD.md` and 10 in `channels.json`. No
    number and no rule changes.
  - `PLAN.md` as fixed has the sha256 `693d4ad196876039771ad83af44ad67f80e6e4e29ef6313432471631349e8720`. The
    published `PLAN.md` is that file with those lines reworded, so its own sha256 is the one in `index.json`.
- **Kept beside the record and not published.** `windows-text.jsonl` (the full caption text of every kept minute),
  `excerpts-excluded.json`, `titles-withheld.json`, the packets (rebuilt by `tools/packets.py`), and the title columns
  of the listings. Only the 1,008 minutes the agents read are published as text; every minute's mark is in
  `marks.json`.
- **Two search patterns in `tools/publish_pack.py`.** Both are withheld, because each would name the computer the
  files were made on.
