The Challenges of Humanizing AI

By Matt Watson, CEO of Full Scale

An AI humanizer rewrites AI generated content so it sounds like it was written by a human. We use StealthGPT on every article we write for Full Scale, and it's worth it. But it also changes things you didn't ask it to change. I ran 39 of our long-form articles through it, and every single one needed work afterwards.

This post covers what a humanizer does well, the specific ways it breaks real content (using some of our long-form articles as examples), and how a second AI pass fixed it. Basically, the two tools do different jobs, and using both is better than just using one.

Why we use a humanizer in the first place

There's a cadence to AI writing. Sentences come out roughly equal in size, paragraphs end at the same point, and everything feels stitched together rather than spoken aloud. An editor can smooth that out, but it takes a lot of hours to do it for 39 articles ranging from 2,000 to 3,800 words each.

Of the humanizers we tested, StealthGPT did the best job. It trimmed verbose sentences, disrupted the rhythm, and brought the language closer to speech. For one article, it replaced "No badge required" with "No swipe card necessary," which was much better than our line. It also maintained a more consistent reading level than the others, which is why we picked it.

So the tool works. The problem arises when you treat its output as finished.

If your content is markdown, check the output format first

We write our articles in markdown, so all the links, headers, tables, and bold text are marked up, and our FAQ sections generate structured data. The default output of StealthGPT's API is plain text, which cannot represent any of that information. We ran one article through the API this way and lost every link and image. By enabling markdown output we eliminated most of that loss before writing any repair code.

We submit our articles to StealthGPT in chunks of approximately 230 words, with headings as hard boundaries. After each chunk we compare the returned text against the original and record every fix. This isn't necessary if you're producing plain text, but if your articles contain links, tables, or other structured data you should verify the output format before submitting even a single document. Even with markdown enabled, the problems below still came through.

Where it fits in our publishing process

It isn't an experiment or a pet project. At Full Scale, Claude handles the whole content pipeline, from drafting to publishing.

The StealthGPT pass is just one part of that workflow, an automated step that immediately triggers a repair step. I'll describe the full process later. Running it on every article without fail is what exposed the problems.

All of it is logged, and since every single run produces a log, that's where the patterns below came from.

Challenge 1: It modifies quotes and facts

This is the highest severity. Humanizers view each sentence as something to be smoothed out or rephrased, without distinguishing between regular prose and things like quotes or data that should not be touched!

In our article on fintech app costs, our draft said that one vendor guide estimates annual compliance costs after year one at 40 to 60% of the first-year investment. StealthGPT changed that to:

One vendor's FAQ states that "on average, the cost to maintain compliance in subsequent years will be approximately 40–60% of year one."

What was once an unquoted paraphrase has been turned into a quote, which implies the vendor literally said those words. In another post on HIPAA, StealthGPT converted a paraphrase of government guidance into a literal quote, and added a bracketed "[T]he" to make it look like a legal citation. It also modified actual quotes. One had a word inserted within its quotation marks, and another lost its quotation marks completely, so a line quoted directly from a law firm is now presented as the author's own paraphrase.

And numerical claims were affected, too. Our original line, "It sorts most vendors in about an hour," came back as "It will eliminate 90% of vendors in one hour." That 90% figure is not in our source, and the claim it creates is much larger than the one we actually made.

Across all 39 articles, our review found altered or made-up content in 36 of them. Whenever you have a quote or a specific number in your draft, be sure to verify that the tool didn't change it. Sometimes it does, and anyone reading that quote has no way to know it was reworded.

Challenge 2: It rewrites sentences that are company policy

There are only a few sentences in every company's copy that act as a sort of promise. Since a humanizer can't tell which ones those are, it treats them all equally.

These are the changes StealthGPT made to the terms we publish at Full Scale:

All four revisions read well, and each one describes something different from what we actually offer. Across the 39 articles, roughly two in three had the money-back guarantee reworded, about 17 turned "as little as 7 days" into a flat claim, and about 30 distorted the $35 starting rate.

It was also the case for a more general statement about our working relationship with clients. The original sentence, "Our team follows whatever processes and procedures you already have," was rewritten as "Our people do whatever you want us to do." No company would ever intentionally publish that second sentence. The tool had no way to know that.

It can even change who someone is employed by. AMC Theatres is one of our clients, and in the original sentence, we talked about the Full Scale developers who work full-time on AMC's software. On one of our test runs, the tool rephrased it to say "full-time AMC employees." Both sentences refer to AMC. But the employer has been swapped.

We employ them, putting them on the AMC team, and the tool's rephrasing says that AMC hired them directly, which is the reverse of how our business model operates. Nothing was omitted, nothing was misspelled. If a program was looking for either of those errors, it wouldn't have found it. Someone would have had to know who really employs those developers.

It's stuff like that, minor details, that makes humanizing text hard. You swap out one word, and either a promise or an employer has changed, but the sentence still reads well.

Challenge 3: It breaks links, tables and structure

Humanizers have difficulty dealing with any kind of markup, be it HTML, markdown, or something else. Since they're designed to rewrite words, any formatting attached to those words usually gets erased. Bold type, hyperlinks, tables, and headings are typical victims. This is the easiest category to detect and the most annoying to fix.

• Hyperlinks. 36 out of 39 articles lost hyperlinks, 191 in all. Another test run dropped 33 of 33.

• Link anchors. 37 of 39 had link anchors rewritten. For example, "managing distributed teams" was replaced by a nine-word, title-cased string that doesn't appear anywhere else on the website.

• Tables. 37 of 39 had their tables flattened, the pipes removed, and the table cells rewritten.

• FAQ sections. All 39 had FAQ answers either rewritten or broken into multiple paragraphs, creating a conflict between the visible page and its structured data.

• "Quick Answer" boxes. 33 of 39 had a "Quick Answer" callout converted to regular text.

• Headings and bolding. 15 articles received headings no one wrote, while 34 lost bold callouts.

Additionally, the tool sometimes invents links to pages that don't exist, and in a handful of cases merges a legitimate anchor into the middle of a different sentence. An actual quote from a consulting firm ended up with a hyperlink pasted in the middle.

Challenge 4: The meaning flips that nobody sees

It's the sort of mistake a spell checker will never catch. Every word is right, but the sentence now means something different.

For example, we wrote that if you hire a contractor halfway around the world, "you may be their second job, or their third." The tool returned "they may be your second or third client." The risk we identified became completely benign.

In another instance, we wrote that a problem was more common than most people realize, and that "it isn't an offshore problem at all." The tool returned "it's not just a developing-world phenomenon," which suggests that it actually is one.

Then there's the one where the subject got switched. "A model bill is the one invoice in software that grows when customers like the feature" came back as "the one we charge for as a model," suggesting that our company charges for models.

In the set of 39 articles I ran through the tool, our review found meaning inversions in 30 of them, and garbled logic in 30 as well.

Even jokes suffered. A line about an AI deleting a production database was reduced to a lifeless sentence with no punchline. It retained the subject but lost its purpose.

How our full process works in Claude

We run this whole process with Claude. We have built our content process in Claude Code in the form of "skills," where each skill is a saved set of instructions for a specific task: create a brief, write a first draft, check voice and sourcing, publish. Both the humanizing and the repairing steps are part of the workflow, so Claude executes them identically across articles without anyone having to remember. So here's how it works once we have a draft that's approved:

1. Claude hands off the draft to StealthGPT. It calls the API with markdown output, chunks the article into segments of roughly 230 words, and treats every heading as a hard boundary. We run this once per draft and never loop.

2. A script compares each segment with the original draft. It detects deleted and rewritten links, changed anchors, flattened tables, stripped bold, dropped numbers, lost brand mentions, unrequested headings, and altered quotes. It catches a lot, but it only uncovers roughly one third of the damage.

3. A second script repairs anything that's objectively broken. It brings back tables, callouts, FAQs, alt text, link anchors, bold, and straight quote marks, and it removes invented headings. No decisions are needed here, and no model's attention should be spent on it.

4. Claude reviews each segment against the source of truth. This is what the scripts can't do. Claude compares the humanized version side by side with the original, with our approved facts, our canonical company numbers and our style rules loaded. This is how it finds the employer change, the rephrased 90%, the softened notice term and the swapped subject. A model that only sees two versions of a sentence will usually go with the nicer one, and that's the entire reason we use reference material.

5. Claude corrects the issue inside StealthGPT's own prose. Rather than pasting our original sentence back, Claude fixes the defect while keeping StealthGPT's surrounding language. Dropping in the original sentence brings back the cadence that made it look AI-generated, and defeats the purpose of running StealthGPT.

6. Claude logs every change. For each article, there is a record of what the tool did wrong, what Claude fixed, and why. All the examples in this post came from those records.

7. Claude builds the article and places it in the publishing queue. A final build pass ensures the page renders, and the publish skill adds it to our release queue.

Order matters. Cheap mechanical repairs run first so the AI read can focus its attention on meaning, and the final check runs last, since any step that writes a file, including a repair, could introduce new errors.

What the combined process gets you

All of the articles retained their links, tables, FAQs and answer boxes after the repair pass, and all of the policy statements accurately reflected what we actually provide. The humanizer's real value was also preserved. We captured 63 actual improvements across the entire run, from simpler language to better flow and brevity, and every article received at least one.

This is why they need to be paired together. StealthGPT performs the tedious task of making a long written draft sound more human. The audit and repair pass does what the humanizer can't: it knows what shouldn't be changed. If you skip the former, the draft will still be stilted. If you skip the latter, you'll end up publishing a quotation no one ever made, or a promise you don't actually make.

Whenever you use AI to create any form of content, you should always perform a second quality check, especially with a humanizer, since the purpose of a humanizer is to alter your language. You need to ensure it didn't also alter your meaning. Provided that check is performed, we've found that StealthGPT brings significant value.

Matt Watson is the CEO of Full Scale, a software development staffing company that connects US tech companies with engineers in the Philippines. He is a four-time technology entrepreneur and the author of Product Driven.