On using AI

I make Life in India alone. I have a full-time job, and around it a small, unruly garden of other projects: Akshara, Paper Lanterns, This Indian Life, Hope and Despair, Rabbit Holes, Poetic Reveries, Seneca, and the essays I write on my Substack. Several of them are still waiting for more of my attention than I can give. The subjects I want to follow here, and on This Indian Life, its more data-driven cousin, are the ones I care about most.

None of it is possible without large language models. Forget building this site; the idea of it alone would have run headfirst into the plain human limit of doing one or two things well. Reading every judgment, report and book behind a Long View, following each one as it changes, recording people in their own language, transcribing and translating them, checking every figure against the page it is printed on: that is a newsroom’s work, and I am one person with evenings and weekends.

I want to talk to people the mainstream media rarely calls, about issues it passes over, from angles it doesn’t take. I want to follow people doing consequential work that otherwise gets no attention at all. Whether any of it gets the attention I hope for depends on Life in India finding readers, and that is a giant if. Without these tools, it wouldn’t even be a question.

I see these models differently from the way the argument about them usually goes. The fight between human-written prose and AI slop is, to my mind, a dumb and false dichotomy. Something is not good because a human wrote it. Have you read the internet? And something is not bad because a machine helped a human make it.

AI is a tool, at least until the day it becomes conscious and develops murderous urges. Until then, the worth of a tool lies in the hands of whoever holds it. I’m not a tech noob, and in my hands these models have made it possible to work on the things I’m most passionate about, where the written word, the spoken word and data meet.

There is a distinction worth keeping. “AI-generated”, in its pejorative sense, and fairly so, describes low-effort slop made to farm engagement and travel on social media. I don’t traffic in that, and nothing here is made that way.

What happens here is AI-assisted. The models work as a writer of first drafts, a proofreader, an ideator, a sparring partner, an assistant and a creative nurse, and they are very good at all of it, whatever your priors about them. I decide what the record follows, whom I speak to, what goes in and what is cut, and every word published here passes through my hands. A great deal of effort and care goes into everything on this site. I will slip up, and it won’t always be perfect. When I get something wrong, it goes in the corrections log.

On a Long View, they read each source, draft its entry and check it, and write “Where things stand” from the entries. On a Close Up piece, they transcribe the recording, tidy and translate it, write a first draft and check it. The charts are drawn from official data by code alone, with no model involved. Code checks everything a reader could check, and the rules below hold for every piece.

1. Quotes are cut, never written. Every word we print as a speaker’s is a run of their own words from the transcript, in order. Code checks this on every draft, and again on the editors’ final edit of a conversation. Words we add are shown in [brackets].
2. Every figure is on the page cited. In a reading, code finds each number on the printed page of the report it cites. A chart value must also sit under its own label in the report’s table.
3. A model’s memory is not a source. Anything a piece says that its source does not must come from a dated, linked entry in a Long View, or it is cut.
4. A second model checks the first. The model that checks a draft claim by claim is never the one that wrote it.
5. Translations are made twice. Where someone spoke in another language, a second model translates their words independently and a third compares the two.
6. People make the decisions. The editors name the speakers, choose what the piece is about, edit the draft by hand and publish it. Every piece is read back to the speaker, or sent to the report’s authors, first.
7. Unpublished recordings stay out of free test models. Models that are free because their makers see what is sent to them are never given anyone’s recording or transcript.
8. What we may quote is decided by rule. Copyright law, not a model, sets how much of a source we quote: a short excerpt under fair dealing, more from judgments, Acts and works out of copyright.
9. Every call is logged. The step, the exact model and the date. Each piece’s and each Long View’s “How this was made” is generated from that log, and so is this page.

None yet. When the first piece made with language models is published, it will be listed here with the models that worked on it and how much the editors changed.

1. Long View 6: Private equity in health

   68 sources read and 57 entries published, 1 October 2026 to 2 October 2026. Models: Claude Opus 5.5, DeepSeek Flash, MiMo V2.6 Flash, GPT-6 Luna. Each step

2. Long View 5: Heat

   56 sources read and 56 entries published, 1 October 2026 to 2 October 2026. Models: Claude Opus 5.5, DeepSeek Flash, MiMo V2.6 Flash, GPT-6 Luna. Each step

3. Long View 4: Business groups

   41 sources read and 40 entries published, on 30 September 2026. Models: Gemini 3.8 Flash, DeepSeek Flash, Xiaomi MiMo V2.6 Flash, Xiaomi MiMo V2.6 Flash Free, Gpt 6 Luna. Each step

4. Long View 3: The jobs crisis

   20 sources read and 20 entries published, 29 September 2026 to 30 September 2026. Models: DeepSeek Flash, Gemini 3.8 Flash. Each step

5. Long View 2: Right to walk

   36 sources read and 30 entries published, 25 September 2026 to 29 September 2026. Models: Gemini 3.8 Flash, DeepSeek Flash, Gemini 3.1 Pro Preview, Space Bunny (maker not named). Each step

Every model used so far, and what it has done here.

Claude Opus 5.5

• Searched journals, filings, ratings, official statistics and reporting for sources
• Drafted each entry from its source
• Revised the entries the checks faulted
• Built the charts and wrote “Where things stand” from the entries
• Searched journals, reporting and official records for sources
• Wrote “Where things stand” from the entries

DeepSeek Flash

• Judged how far each source bears on the subject
• Drafted each entry from its source
• Revised the entries the checks faulted
• Checked “Where things stand” against the entries
• Wrote “Where things stand” from the entries
• Chose the links in “Where things stand”

Gemini 3.1 Pro Preview

• Checked each entry claim by claim against its source
• Wrote “Where things stand” from the entries
• Checked that the chart’s figures count the same thing

Gemini 3.8 Flash

• Searched the web for sources
• Sorted what the search found
• Drafted each entry from its source
• Checked each entry claim by claim against its source
• Wrote “Where things stand” from the entries
• Checked “Where things stand” against the entries
• Chose the links in “Where things stand”
• Revised the entries the checks faulted
• Judged how far each source bears on the subject
• Found the chart’s figures in the sources

Gpt 6 Luna

• Checked each entry claim by claim against its source
• Wrote “Where things stand” from the entries
• Checked “Where things stand” against the entries

GPT-6 Luna

• Checked each entry claim by claim against its source

MiMo V2.6 Flash

• Judged how far each source bears on the subject

Space Bunny (maker not named)

• Checked “Where things stand” against the entries

Xiaomi MiMo V2.6 Flash

• Revised the entries the checks faulted
• Chose the links in “Where things stand”

Xiaomi MiMo V2.6 Flash Free

• Checked each entry claim by claim against its source
• Checked “Where things stand” against the entries