Dipak Bhosale

Dipak Bhosale founded TextSight.ai in 2025 and runs it under Lacewing Technologies. He works on the detection models and the humanizer directly, and writes the guides on this site, so the how-it-works explanations come from inside the product rather than from the outside looking in.

His focus is making AI detection honest and useful: showing writers a clear authenticity signal instead of a black-box verdict, being upfront about where detectors get things wrong, and never claiming any tool is "100% accurate" or that rewritten text is "undetectable."

Published research

He is the author of TextSight's original research, which measures the company's own detector and publishes the dataset so the results can be checked by someone who does not work here.

• False positives on 1,180 human papers (August 2026) — every document published in 2018, years before ChatGPT, so every AI flag is a false positive by construction. The detector flagged 5.85% of them. The study also went looking for the well-known bias against second-language English writers and reported not finding a significant one.
• Text length and false positives (August 2026) — tests the common advice that submitting more text protects you from being wrongly flagged. Between 150 and 392 words it does not: 5.79% in the shorter half against 5.90% in the longer half, trend test p = 0.777.

Both are published with the full per-document dataset under CC BY 4.0. The research section also lists what has not been measured, including the absence of any head-to-head benchmark against other detectors.

Books

He has written three books, all free to read in full on this site with no email form and no paid edition. Together they run to roughly 38,000 words and take the same position as the research: say what the tools do, say where they fail, and show the working.

• Sounds Like You — 10 chapters, about 7,000 words. What “sounds like AI” means line by line, how detectors reach a verdict, and a twelve-minute edit that puts a person back into a draft.
• Nobody Wrote That Paper — about 16,000 words on citations that models invent rather than retrieve: the three shapes a fake reference takes, how to settle one with a DOI check, what to do when you find one in your own bibliography, and the point at which it reaches a judge.
• Flagged in Translation — about 15,000 words on why careful second-language English gets read as machine output. What Turnitin actually checks, how Google Translate and DeepL each give themselves away, and how to make a translation read like a person wrote it.

All three sit together on the books shelf, and all three are free to print and free to share whole and unchanged, which teachers and writing centres are welcome to do without asking.

Editorial standards

Everything published under this byline follows the same rules, and they are checkable against the pages themselves:

• No single headline accuracy number. One figure across every length, genre and language is misleading. The reasoning is set out in the methodology.
• No "undetectable" claims. The humanizer is described as an editing tool, not an evasion tool.
• Competitor comparisons state features, pricing and privacy, not measured accuracy — because no head-to-head benchmark has been run, and publishing accuracy figures nobody measured would be worse than leaving a gap.
• Limitations are on the page, not in a footnote. Each study carries a section saying what it cannot support.

What he writes about

• How AI content detectors actually work — and where they fail
• False positives and how often detectors are wrong
• What to do when you are wrongly accused
• Humanizing AI drafts without stripping the meaning
• Turnitin, GPTZero and other tools, compared on what they publish

Corrections

If something here is wrong — a figure, a competitor's pricing, a claim about how detection works — say so and it gets corrected with a date rather than quietly edited.