to/human

Blog

From the ToHuman team

Research, guides, and analysis on AI text humanization, detection tools, and the evolving landscape of AI-generated content.

Research

What Changes When AI Text Is Humanized: 138 Pairs, Measured

Every study of machine-written prose compares human text to AI text. There is a third category almost nobody has measured: the same text before and after a rewrite. We scored 138 of our own pairs with the public AI Tells Scanner word lists. Median tell density falls to zero — and the statistics got worse for up to 9.7% of texts, rhythm flattened for three in ten, and most text the scanner flagged before is still flagged after. No detector was run.

· 9 min read

Explainer

Can You Humanize AI-Generated Code? The Honest Answer

Mostly no. GPTZero describes its classifier as finetuned for prose; Turnitin will not score a submission without 300 words of long-form prose. Code is not what they measure. What is actually readable as machine-written in a code submission is the English around it, which is the only layer a rewriter should touch. Includes what our own model does to fenced code blocks, and what it refuses to do.

· 7 min read

Research

How to Appeal an AI Detection Accusation: The Procedure, Step by Step

We read 21 published US university appeal procedures. Every one of the 19 that publishes a deadline gives you 3 to 10 days, and 12 of them exactly 5. The grounds are a closed list: procedural error is accepted at 16 of 16, but only 7 of 16 let you argue the finding wasn't supported by the evidence. What that means for where you make the factual case, and what to put in the letter.

· 11 min read

Quarterly Tracker

Do AI Detectors Work? The AI Detector Accuracy Tracker (Q3 2026)

Every quarter we push the same 861 verified human-written sentences through whatever detectors we can reach. Q3 2026: three of six open-weight detectors scored at or below a coin flip, the two that passed validation still flagged ~10% of human sentences as AI, and 23.3% of the corpus was flagged by at least one detector while only 2.7% was flagged by all three.

· 15 min read

Research

Falsely Accused of Using AI: The Full Checklist Nobody Publishes in One Place

We coded the six student-facing advice pages that rank for this question against ten concrete defence actions. Mean coverage: 5.0 out of 10, best page 6. Tool vendors document the evidence half, law firms document the rights half, and the three most-omitted actions are the three that cost nothing. Here is the merged checklist — plus how to prove you wrote it, and why detectors flag human writing.

· 12 min read

Product

GPTZero Bypass: What That Search Actually Means, and What Actually Helps

Six of the ten results for "gptzero bypass" promise a guaranteed pass. None of them can deliver one — the model belongs to somebody else. What the demand actually reflects: GPTZero flagged 13.8% of 861 verified human-written sentences in our own study, against a stated ≤1% false-positive rate. Plus what genuinely helps, and where a humanizer honestly fits.

· 8 min read

Product

ChatGPT Humanizer: Humanize ChatGPT Text — and the Honest Limits

Paste ChatGPT text in, get it back reading human — then read the part the other tool pages skip. A before-and-after, the four approaches ranking for this query compared, and our production data: across 179 accounts the median humanization is 49 words, not a 2,000-word essay. Plus the failure rates in every window and when you don't need one at all.

· 13 min read

Research

Does Turnitin Detect AI-Generated Content in 2026? An Honest Explainer

We pulled the live US search results for the four biggest Turnitin questions and classified every result by who owned it. Page one of "how to bypass Turnitin AI detection 2026" has 17 organic results — zero from a university, zero from Turnitin. Plus a first-party census of 34 institutional guidance pages on what those universities actually say.

· 10 min read

Guide

How to Humanize AI Chatbot Responses and Email Replies (2026 Guide)

Support chatbots and AI-drafted email replies have a signature stiffness — corporate greetings, tidy hedges, paragraphs that keep going. Three methods for fixing it (prompt engineering, API humanization, manual editing) compared with worked before/after examples for a support chat and a Gmail AI draft.

· 9 min read

Comparison

Best AI Humanizer: 11 Tools Tested in 2026

Eleven humanizers run against five detectors, scored on free tier, readability, and true cost per word — with our own conflict of interest stated up front. The finding that matters isn't the ranking: six purpose-built tools sit inside a 12-point band, so price and free allowance decide it.

· 14 min read

Developer Guide

How to Humanize AI Text with an API: n8n, Zapier & MCP Integration Guide (2026)

Every humanizer API is one HTTP call. Here's the exact REST request shape for n8n's HTTP Request node, Zapier's Webhooks action, and an MCP tool for Claude Desktop — with sync vs async endpoints, the four intensity levels, common failure modes, and which of the six providers actually has a real free tier.

· 11 min read

Original Research

We Ran 861 Human-Written Sentences Through GPTZero. 13.8% Were Falsely Flagged as AI.

A controlled benchmark on four corpora of verified pre-LLM human writing — Wikipedia, PubMed, news, and Reddit ESL. GPTZero flagged 13.8% of sentences as AI. News and journalism prose, the most-published category of human writing, was flagged the most. Full per-corpus breakdown, top falsely-flagged examples, and a downloadable raw CSV.

· 12 min read

Guide

How to Pass GPTZero in 2026

GPTZero is the AI detector educators use most, and it returns a percentage score that can sit on your essay before a human reads it. How the classifier actually works, why it flags writing that was never AI-generated, and the step-by-step methods that reduce your AI score before submission.

· 11 min read

Original Research

Is Google Punishing AI-Written Content? We Checked 258 Top-Ranking Pages.

We pulled 50 Google SERPs, fetched 258 ranking pages, and ran every one through GPTZero. Result: 96.5% scored human. Only 1.9% flagged as AI — a fraction of the 17% Originality.ai reports. We then compared against five industry studies spanning 1.5 million pages. The 45x gap between tools is the finding.

· 12 min read

Analysis

GEO Gets the Credit. SEO Did the Work.

Search interest in "AI search optimization" terms grew 2-6x in twelve months. Three pieces of evidence published in March-April 2026 — Ahrefs' 1.4M-prompt study, Lily Ray's GEO/SEO breakdown, and Google's March 2026 Core Update — all point the same direction: AI search visibility is downstream of established Google rankings, not a separate playbook.

· 8 min read

Policy

What Universities Get Wrong About AI Detection Policies in 2026

Most campus AI detection policies still treat probability scores as evidence — from tools whose own vendors now disclaim that use. A policy-level look at what academic integrity and compliance teams should actually do in 2026.

· 12 min read

Review

Does Undetectable AI Work? An Honest Review for 2026

With a 2.1/5 Trustpilot rating and growing user complaints, people are asking whether Undetectable AI still delivers. An honest look at what works, what doesn't, and what to consider instead.

· 9 min read

Research

AI Detection Is Broken: The False Positive Crisis of 2026

NBC News is covering it. Students are filing lawsuits. 25+ universities have banned AI detectors. A deep dive into the data behind AI detection's false positive problem — and what it means for writers.

· 10 min read

Analysis

Universities Are Banning AI Detection -- Here's Why (2026)

12+ universities including Yale, Vanderbilt, and Curtin have disabled Turnitin's AI detection. False positive rates up to 61% for ESL students. A comprehensive look at the institutional rejection of AI detection technology.

· 10 min read