A reader sent us an essay a few weeks ago. They'd used ChatGPT to help them draft a 900-word piece on the ethics of algorithmic hiring, and then edited it for about 20 minutes. They felt like they'd done enough. They ran it through GPTZero out of curiosity — 78% AI.
Panic.
They sent it to us with one question: "Is this salvageable, or do I need to start over?"
It was salvageable. Here's everything we did, step by step, including the specific sentences that changed and why.
The Starting Draft
Here's a representative excerpt from the original draft — the part that came back most heavily flagged:
"It is important to consider the significant ethical implications of algorithmic hiring systems. These systems have the potential to perpetuate existing biases, as they are trained on historical data that may reflect societal inequalities. Furthermore, it is crucial to note that the lack of transparency in these algorithms makes it difficult for candidates to understand why they were rejected. This highlights the need for greater accountability in the development and deployment of such technologies."
Before we even ran it through TextSight, that paragraph had three problems visible to the naked eye:
- "It is important to consider" — classic AI opener
- "Furthermore, it is crucial to note" — double-stacked AI transition phrases
- "This highlights the need for" — the textbook AI summary sentence
The whole paragraph reads like a model that was asked to "write an academic paragraph about AI hiring bias." Which is, of course, exactly what happened.
Round 0 score: 78% AI on GPTZero / 28/100 on TextSight Humanization Score.
What TextSight Flagged
When we ran the full essay through TextSight, the AI Vocabulary Highlighter lit up consistently in four areas:
1. The introduction. Three opening sentences all starting with formal subject-verb constructions and hedging language: "It is important to..." / "This demonstrates that..." / "In order to fully understand..."
2. The hiring bias paragraph (excerpted above). Every sentence flagged.
3. Two transition sentences between sections. Both used additive transitional phrases that AI produces constantly: "Moreover, it should be noted that..." and "In addition to the above concerns..."
4. The conclusion. Almost entirely flagged. It started: "In conclusion, the ethical considerations surrounding algorithmic hiring are multifaceted and require a holistic approach to address effectively." That sentence has four words from the TextSight banned vocabulary list in one sentence: "multifaceted," "holistic," and two forms of the AI hedging register.
The clean sections? Three paragraphs in the middle where the reader had added their own thoughts — a specific example about Amazon's abandoned hiring algorithm, a personal observation about applying for jobs, and a paragraph they'd written entirely from scratch. Those were either not flagged or lightly flagged.
The pattern was clear: the skeleton of the ChatGPT draft was pulling the score down. The original additions were fine.
Round 1: Rewrite the Introduction and Transitions
The reader's first instinct was to rewrite everything. We told them not to. Touch only what's flagged.
The introduction went from:
"Algorithmic hiring systems have become increasingly prevalent in the modern job market. It is important to consider the significant ethical implications these systems present, as they have the potential to impact millions of job seekers globally."
To:
"Algorithmic hiring is everywhere now. Amazon, Goldman Sachs, Unilever — companies at every scale are running applications through AI screeners before a human ever sees a resume. That might seem efficient. It's also producing some uncomfortable outcomes that aren't going away on their own."
Changes made:
- Cut "it is important to consider" entirely
- Added a specific short list of real companies (Amazon, Goldman Sachs, Unilever)
- Broke the second sentence into a short statement and a longer one
- Replaced hedged language ("have the potential to impact") with direct language ("producing some uncomfortable outcomes")
The two mid-essay transitions went from:
"Moreover, it should be noted that..." "In addition to the above concerns..."
To just cutting them. The paragraphs connected fine without them. AI transitions are often just noise — removing them makes the writing tighter, not looser.
Round 1 score: 51% AI on GPTZero / 54/100 on TextSight.
Better. Still in the grey zone. The introduction and transitions were clean now. The flagged paragraph about training data bias and the conclusion were still problems.
Round 2: Attack the Flagged Paragraph
This was the hardest round because the content of the paragraph was actually correct — the ideas were sound. The problem was purely the phrasing.
Original:
"These systems have the potential to perpetuate existing biases, as they are trained on historical data that may reflect societal inequalities. Furthermore, it is crucial to note that the lack of transparency in these algorithms makes it difficult for candidates to understand why they were rejected."
Rewritten:
"These systems learn from historical hiring data — which means they learn from every biased decision that data contains. Amazon's 2018 scrapped hiring tool had been trained on 10 years of male-dominated hiring. It taught itself that women were worse candidates. The company caught it before it went live. Most companies don't audit this carefully. And when a candidate gets rejected by an algorithm, they typically get no explanation at all — just a form email."
What changed:
- "Have the potential to perpetuate" became "learn from every biased decision" — specific and active
- Added the Amazon 2018 example with a specific date (this is a real, documented case)
- "It is crucial to note" is gone
- "Makes it difficult for candidates to understand" became "they typically get no explanation at all" — concrete and direct
- Added "just a form email" — a specific, human detail that no AI model adds
The Amazon example did a lot of work here. Specificity is the enemy of AI-like text. Real dates, real company names, real documented incidents are the easiest way to make a paragraph sound like a person wrote it, because a person who researched the topic would naturally include those details.
Round 2 score: 31% AI on GPTZero / 74/100 on TextSight.
Major improvement. One section still dragging — the conclusion.
Round 3: Rewrite the Conclusion
Conclusions are where AI really shows its patterns. They tend toward summarizing what was just said (redundant), using resolution language ("it is clear that," "ultimately," "in conclusion"), and ending with vague calls for further action.
Original conclusion:
"In conclusion, the ethical considerations surrounding algorithmic hiring are multifaceted and require a holistic approach to address effectively. Stakeholders across sectors must collaborate to develop transparent, accountable systems that prioritize fairness and mitigate bias. Only through such collective efforts can we ensure that algorithmic hiring serves as a tool for equality rather than a mechanism for perpetuating inequality."
That's three sentences. Every word of it is flagged. It reads like a model completing an assignment.
Rewritten conclusion:
"Algorithmic hiring isn't going away. The question is whether companies will be required to audit their systems — and whether candidates will ever have a right to understand why an algorithm rejected them. Right now, neither of those things is consistently true. That's the problem worth solving. Not whether to use these tools, but whether anyone's actually watching what they do."
Changes:
- Cut "In conclusion" entirely — started directly with the subject
- Removed "multifaceted" and "holistic" (banned AI vocabulary)
- Cut the redundant summary sentences entirely
- Ended on a specific question/tension rather than a vague call to action
- Wrote in shorter, declarative sentences with one slightly longer one for rhythm
Round 3 score: 12% AI on GPTZero / 88/100 on TextSight.
Done.
What Made the Difference
Looking across all three rounds, the changes that moved the score the most were:
Specificity. The Amazon 2018 example, the real company names in the intro, the "form email" detail. Specific facts are the fastest way to sound like a human who actually thought about the topic.
Cutting AI transition phrases. "Furthermore," "Moreover," "In addition," "It is important to note," "It is crucial to consider" — every one of these is a signal. They don't add meaning. Cut them or replace them with a short, direct sentence.
Short sentence variation. "That's the problem worth solving." Four words. No AI model drafts four-word standalone sentences in academic essays. Throwing one in breaks the pattern.
Cutting the conclusion summary. Don't summarize what you just wrote. End with a question, an implication, or a specific thing that's unresolved. Summaries are for AI; endings are for writers.
Total rewriting time across three rounds: about 35 minutes. The essay went from "would trigger an academic integrity review" to "passes everything."
Try It on Your Draft
The process works because TextSight shows you exactly what's flagged. You're not guessing — you're editing specific sentences until the highlights go away.
Five free scans a day, no account needed. Paste in your draft, see your Humanization Score, look at what's highlighted, make changes, scan again.
Try the same process on your draft → textsight.ai — 5 free scans daily, no account needed.
Related reading:
- How to Avoid AI Detection as a Student
- How to Humanize ChatGPT Text
- Why Sentence-Level AI Feedback Changes Everything
- Can GPT-5 Be Detected?
Patterns That Show Up in Almost Every AI-Assisted Draft
After working through a few dozen of these editing processes, the same problems appear consistently. It's worth naming them so you can spot them in your own drafts without waiting for a scan to tell you.
The double-stacked hedge. "It is important to note that it is crucial to consider..." AI models love stacking hedges. They make the model sound measured and academic. They make writing sound like a compliance document. Cut one, cut both, replace with a direct claim.
The "this" topic sentence. "This demonstrates..." "This highlights..." "This underscores..." "This suggests..." Starting three paragraphs in a row with "this" followed by a verb is a GPT signature. Vary the structure: start with the subject, start with the evidence, start with a question, start with a short observation.
The significance coda. Ending paragraphs with a sentence about why the topic matters: "This is significant because..." "The implications of this are far-reaching..." "Understanding this is crucial for..." AI models add significance codas because they were trained on academic writing that explains its own importance. Real writing usually lets the significance be obvious. Cut these.
The three-part listing without specifics. "This affects students, educators, and policymakers." That sentence sounds complete but doesn't say anything specific. Any list of three should have something concrete about each item, or you should cut the list and say the specific thing.
Conclusion as summary. We covered this above. AI conclusions tell you what was just argued. Human conclusions tell you what's unresolved, what comes next, or what the argument changes. If your conclusion could be replaced by "in summary, what we discussed was..." then rewrite it.
Recognizing these patterns in your own drafts is faster than waiting for a scan. You'll start catching them at the writing stage, which means fewer editing rounds and better first drafts over time.
Why This Process Beats the "Just Rewrite Everything" Approach
The instinct when you get a high AI score is to trash the draft and start over. That's usually wrong.
The clean paragraphs in that original essay — the ones the reader had written themselves, with the Amazon example, the personal anecdote about job applications — were genuinely good. They were specific, grounded, and didn't sound like AI. Throwing them away and starting fresh would have meant rewriting the good parts alongside the bad.
Sentence-level feedback lets you keep what works and fix what doesn't. That's faster, and it usually produces a better final piece because the original observations — the ideas you actually had — stay in. You're editing a draft, not starting from scratch.
The 35 minutes we spent on that essay covered three targeted rounds on about 40% of the total word count. The other 60% was untouched. That's the efficiency gain from knowing where to look.