writing-style on S Anand

Beating AI detectors by reading aloud

Wed, 01 Jul 2026 17:43:49 +0800

Ranjeeta asked me for an article for Built In. I went straight to ChatGPT and said:

Write an article for Built In.
Read the section below for context on Built In's audience, style, and content preferences.
Read the pitch that explains what the article should be about.
Then use my blog posts, talk content, transcripts, TIL, etc.
Write an article using my writing style.

… and gave it all related information.

Here is ChatGPT’s article. The article itself was pretty good, content-wise, but it wasn’t exactly in my style and I iterated once. Still… not exactly there.

One of the criteria is that “Final drafts must score ‘human-written’ or less than 20% on ZeroGPT and Pangram.” ZeroGPT is easy to fool but Pangram is harder. Pangram said:

AI Generated. 100% of this text is AI Generated

“No Juniors, No Experts? Ankor runs a …” (384 words) - We believe the segment is fully AI generated. Confidence: High.

“In March 2026, I pointed a coding …” (383 words) - We believe the segment is fully AI generated. Confidence: High.

“Make it generate rare failures. Make it …” (196 words) - We believe the segment is fully AI generated. Confidence: High.

Also: “AI-Generated indicates text produced by an AI system with minimal human input or revision.”

Sigh… Just to cross-check, I pasted a 2021 blog post and Pangram said:

Human Written. 100% of this text is Human Written.

“Picking gifts is hard. Gift-giving feels …” (331 words) - The segment is fully human-written. Confidence: Low.

Well, maybe Pangram does do a good job of detecting AI-generated text. It managed to by-pass my carefully crafted writing style which includes several LLM smell avoidance techniques.

Next attempt: I checked if Claude’s new Sonnet 5 could do a better job.

Here is Claude’s article. Though the content was, again, spot-on, the style had clear LLM smells despite all my instructions.

Pangram said:

AI Generated. 100% of this text is AI Generated

“No Juniors, No Experts? AI is cutting …” (359 words) - We believe the segment is fully AI generated. Confidence: High.

“He’s three times faster than someone …” (132 words) - We believe the segment is fully AI generated. Confidence: High.

At this point, inspiration struck. (Actually, not quite. I had this idea for a few days, maybe weeks, ago… but the opportunity struck.)

I opened the Claude article on the left, ChatGPT on the right (just to transcribe, nothing else - ChatGPT has the best transcription right now), and read out the article. Not word-for-word, but in my style. For example:

Claude said …	I read it out as …
AI is cutting the entry-level jobs that used to train tomorrow’s architects.	These days, AI is reducing the number of entry-level jobs that we have.
AI makes senior architects more productive and cuts the need for junior engineers.	He said, AI makes senior architects more productive and reduces the need for junior engineers.
The data backs his worry.	The data supports his concern.
A free phone app now beats any grandmaster.	Any free phone application today can beat every grandmaster.
A broken deployment, a weak first draft: you catch it and move on.	Broken deployments and prototypes and quick POCs aren’t so important that you can’t live with them.

The beauty of this is that it I was able to read it out almost live. Just read a sentence and narrate it like I’m talking to an audience. I’m used to doing this.

Here is the version I read out. Pangram said:

Human Written. 100% of this text is Human Written.

“No Juniors, No Experts? These days…” (381 words) - The segment is fully human-written. Confidence: High.

The beauty of it is that it has High confidence - even more so than my earlier blog post which was truly human written.

Still, good to know that there’s an efficient way to use AI that doesn’t smell like AI.

Update: 22 Jul 2026. Publicize suggested some feedback:

Overall, the piece is conversational, which fits Built In’s tone. That said, we’d suggest a few adjustments:

Structure: Our other published articles tend to use clear subheadings and a defined structure, and we think this piece would benefit from the same. It currently reads a bit loosely by comparison.
Depth and takeaways: The current draft feels like it’s missing some of the concrete, educational insights that give readers a clear takeaway. Our original draft had more of that, and we’d suggest reintroducing some of those points.
Title: The current title is a bit more open-ended than what we’ve typically used. Our past titles (and the one in our original draft) tend to be more specific about what the reader will get from the piece. This one’s more of a preference than a hard rule, so happy to discuss if you feel strongly about it.

I passed this to Claude Fable 5 make a few edits, and here is the revised version following the same process. Pangram said:

Human Written. 100% of this text is Human Written.

“These days, AI is reducing…” (380 words) - The segment is fully human-written.

“Chess engines have become the coaches…” (350 Words) - The segment is fully human-written

AI makes me a better person

Thu, 21 Mar 2024 04:23:00 +0000

Every time I get annoyed at people, I remind myself to be more like ChatGPT. Specifically:

Don't get annoyed. Be patient.
Encourage them.
Step back and show them the big picture.

(Then I get annoyed at myself for getting annoyed.)

Today, I analyzed how exactly ChatGPT is different from me. So, I took a pitch document I co-authored with ChatGPT.

Section A: Authored by Anand

WHAT DO WE NEED?

We are looking for API access to (SYSTEM) via the REST API as an Agent role (read/respond to emails). Specifically, access via a bearer token.

This would be accessed by a single application developed by a team of 3 developers and 1 business analyst. None of them have access to (SYSTEM) today.

WHY DO WE NEED THIS, AND WHY SO SOON?

We need this to classify emails automatically, as they arrive, into categories such as “non-value-add” (e.g. Thank you, Out-of-office, etc.)

We’d line access today, please. Currently, we are processing XX,XXX non-value-add emails per month. Each day of delay leads to a processing waste of ~XX emails per day. At current volumes, this will save ~$XX,XXX per annum.

IS THERE ANY OTHER WAY?

To access emails in real-time, this is the only option. (We’re ruling out web scraping.)

(SYSTEM) rule-based closure won’t work. It’s based on keyword search, not semantic matches. For example, “Thank you” is present in non-value-add emails as well as follow-up questions. Multi-lingual matches are a challenge. So, though (SYSTEM) currently closes emails with rules, 1 out of 6 emails that are NOT closed already are non-value-add.

This cannot be done via (SYSTEM) or any other BI system because it involves accessing the data in real-time as well as responding via (SYSTEM). This needs to be steady-state access, not one-time, since we want to classify emails real-time.

HOW LONG WILL IT TAKE?

Based on our past experience, this would take about X weeks to prototype, X weeks to develop, and X weeks to test. We should have the application deployed in X months.

Section B: Authored by ChatGPT

WHAT’S THE IMPACT IF WE DON’T MOVE FORWARD?

Quantitatively, continuing without this solution wastes resources on XXX non-essential emails daily, costing $XX,XXX yearly. Delays in customer responses will ensue.

Qualitatively, lack of automation strains customer service, risking employee burnout and diminished customer experience.

WHAT’S THE PRIORITY (MOSCOW)?

(SYSTEM) API access via REST as an Agent role with a bearer token is a “MUST HAVE”. Alternative methods like scraping (SYSTEM) are impractical due to high effort and maintenance.

WHAT’S THE OUTCOME IF WE MOVE FORWARD?

Financially, securing API access saves $XX,XXX annually by cutting manual email sorting. Efficiency gains reduce operational costs and free resources for strategic tasks.

Strategically, automating email classification boosts customer satisfaction and response times, establishing leadership in service innovation. Long-term, it enables data-driven insights into customer behaviour, improving decisions and customer service scalability.

Analysis

I fed these to ChatGPT and asked: "You are an English professor evaluating an essay. Compare the writing styles of the 2 sections. They're written by different authors. List the major differences in writing style."

Here's what it said (roughly):

Area	Anand	ChatGPT
Structure	Direct, concise, with specifics	Broader perspective and implications
Focus	Detailing the process	Benefits and consequences
Detailing	Specific, providing clear picture	General, providing implications
Tone	Conveys urgency	Reflective rather than urgent
Approach	Problem-solving oriented	Outcome-oriented

Yeah, that's definitely me on the left.

I like the direct, concise, specific part. I plan to add the "broader perspective", "implications", and "outcome-orientation" to my life.

Postscript

Jaidev pointed out that this is confirmation bias. He asked ChatGPT for the similarities in the writings. It said both are clear, direct, structured, specific, quantitative, and strategic. So, if you ask for differences, you'll get them, even if they're marginal.

I now need to learn (and write about) framing questions well!

Lord of the Rings by other authors

Thu, 07 Feb 2002 12:00:00 +0000

The Lord of the Rings, by various other authors.