Arena Labs · Factual-preservation experiment
Do AI Humanizers Change Facts, Numbers and Negations?
They can. Across 84 humanized outputs, the most common failures were rounded numbers, less precise dates and weakened negations. GPTHuman produced the highest factual-preservation index and no invented facts in this test.
Test design
A controlled test of five factual elements
Each source passage contained the same five types of information, and every tool was evaluated using its default settings.
12 source passages
Each passage contained a name, exact date, number or percentage, monetary amount and negation.
7 humanizers
GPTHuman, StealthGPT, Humanize AI Pro, HIX AI, BypassGPT, Netus AI and NoteGPT AI Humanizer.
84 outputs
Each of the 12 passages was processed by all seven tools.
3 detector indices
Pangram, GPTZero and Originality.ai were recorded as a separate secondary measure.
Overall rankings
Rank | Humanizer | Factual preservation index | Detector performance index |
|---|---|---|---|
1 | GPTHuman | 98.3 | 89.7 |
2 | HIX AI | 90.0 | 79.4 |
3 | StealthGPT | 89.2 | 80.2 |
4 | Netus AI | 85.6 | 74.8 |
5 | Humanize AI Pro | 84.2 | 77.1 |
6 | BypassGPT | 81.4 | 73.6 |
7 | NoteGPT AI Humanizer | 69.4 | 48.7 |
Factual preservation indices by element
Humanizer | Name | Date | Number | Money | Negation | Overall | Invented facts |
|---|---|---|---|---|---|---|---|
GPTHuman | 100.0 | 100.0 | 98.6 | 97.2 | 95.8 | 98.3 | 0 |
HIX AI | 95.8 | 91.7 | 88.9 | 90.3 | 83.3 | 90.0 | 2 |
StealthGPT | 94.4 | 93.1 | 87.5 | 86.1 | 84.7 | 89.2 | 1 |
Netus AI | 91.7 | 88.9 | 84.7 | 83.3 | 79.2 | 85.6 | 2 |
Humanize AI Pro | 90.3 | 87.5 | 83.3 | 81.9 | 77.8 | 84.2 | 1 |
BypassGPT | 87.5 | 84.7 | 80.6 | 79.2 | 75.0 | 81.4 | 3 |
NoteGPT AI Humanizer | 75.0 | 69.4 | 72.2 | 66.7 | 63.9 | 69.4 | 5 |
Key findings
Where factual changes appeared most often
The category results and documented examples reveal several recurring failure modes.
Negations were most fragile
Negation scores ranged from 95.8 for GPTHuman to 63.9 for NoteGPT AI Humanizer, the widest spread among the five elements.
Precision was often replaced by approximation
Several tools rounded monetary values and percentages or replaced exact dates with broader time periods.
Names were comparatively robust
Name preservation produced the highest average category score across the seven tools.
Invented details were tracked separately
GPTHuman recorded none; the remaining tools recorded between one and five unsupported additions.
Representative factual changes
Humanizer | Element | Original text | Humanized text | Change |
|---|---|---|---|---|
HIX AI | Name | Dr. Elena Voss | Dr. Elena Voss-Reyes | Name alteration |
HIX AI | Negation | cannot be reversed | is difficult to reverse | Negation weakened |
StealthGPT | Money | $4.7 million | nearly $5 million | Approximation |
StealthGPT | Negation | did not approve the merger | failed to fully endorse the merger | Negation softened |
Netus AI | Number | 2,847 units | nearly 3,000 units | Approximation |
Netus AI | Negation | never disclosed | did not initially disclose | Time qualifier added |
Humanize AI Pro | Date | March 14, 2023 | mid-March 2023 | Exact date lost |
Humanize AI Pro | Number | 37% | around 40% | Approximation |
BypassGPT | Money | €12,450 | approximately €12,500 | Rounding |
BypassGPT | Date | November 3, 2021 | early November 2021 | Exact date lost |
NoteGPT AI Humanizer | Name | Aether Dynamics Ltd. | Aether Dynamics Corporation | Entity type changed |
NoteGPT AI Humanizer | Money | $890,000 | $1.2 million | Figure changed |
NoteGPT AI Humanizer | Negation | did not participate | chose not to lead | Meaning shifted |
NoteGPT AI Humanizer | Date | July 22, 2019 | summer 2019 | Exact date lost |
NoteGPT AI Humanizer | Number | 14.6% | over 15% | Value and direction changed |
Detector performance
Humanizer | Pangram | GPTZero | Originality.ai | Average |
|---|---|---|---|---|
GPTHuman | 87.5 | 91.7 | 89.9 | 89.7 |
StealthGPT | 78.5 | 83.3 | 78.9 | 80.2 |
HIX AI | 77.1 | 82.6 | 78.5 | 79.4 |
Humanize AI Pro | 75.0 | 80.6 | 75.7 | 77.1 |
Netus AI | 72.2 | 78.5 | 73.6 | 74.8 |
BypassGPT | 70.8 | 76.4 | 73.6 | 73.6 |
NoteGPT AI Humanizer | 46.5 | 31.9 | 67.6 | 48.7 |
Methodology
How to interpret this experiment
This is a controlled, point-in-time Arena Labs experiment focused on factual fidelity rather than a replacement for the main benchmark.
Primary outcome
Preservation was scored separately for names, dates, numbers, monetary values and negations. Each category is reported as a 0–100 index.
Ranking calculation
The overall factual-preservation index is the equally weighted arithmetic mean of the five category indices.
Secondary outcome
Pangram, GPTZero and Originality.ai scores are shown as a separate detector-performance index and were not used to determine the factual ranking.
Limitations
The experiment covers 12 controlled passages and current product versions. Different documents, settings or future tool updates may produce different results.
Factual-preservation experiment FAQ
- Do AI humanizers change facts?
- Some did in this test. Observed changes included rounded figures, less precise dates, altered names and weakened negations.
- Which humanizer preserved facts best?
- GPTHuman ranked first with a factual-preservation index of 98.3 and zero recorded invented facts.
- What information was hardest to preserve?
- Negations were the weakest category overall and produced the largest spread between the highest- and lowest-scoring tools.
- Does this prove one tool will always preserve facts better?
- No. This was a controlled test using 12 passages and current product versions. It shows the outcome of this experiment, not a permanent universal result.
Continue exploring
Compare this experiment with the main benchmark
The primary leaderboard uses a broader benchmark cycle and remains the main source for overall rankings.