AI humanizers compared directly
How does StealthGPT Heavy perform in an academic AI humanizer comparison? The archived run also selected PhD, but StealthGPT's current API documentation says that tone is ignored when rewriting existing text.
Internal benchmark · July 28, 2026. The same source texts, blinded review, and separate checks for meaning, language, length, and citations.
The result in one sentence
+15.4StealthGPT maintained the language and many citation sets, but shortened texts substantially. Certentis led clearly on overall quality, semantic fidelity, and Human-plus-Mixed.
Overall quality
Quality scale 40–100 · truncated axis
Semantic fidelity
Exact citations
Distance from source quality
Best version in blinded review
Product profile and terms
The benchmark measures output quality. This section adds the current product context: modes, limits, languages, platforms, and publicly documented safeguards.
Certentis in everyday use
A controlled humanizer for academic and professional texts, with separate AI analysis and a traceable history.
Certentis does not optimize for a scanner classification alone. Automatic checks cover language, length, rewriting, and recognized protected content. Meaning is assessed in the archived benchmark and still requires your final review.
Product profile
A detector-oriented writing suite with a humanizer, writer, and research agent. Heavy is the humanizer model; PhD is a tone for newly generated essays and, according to the API, is ignored for rewrites.
StealthGPT focuses on substantial rewriting and high volumes. Its own test shows that formatted citations can work well, while structured citeproc cases require particularly careful review.
| Features and terms compared | Certentis | StealthGPT Heavy PhD |
|---|---|---|
| Price and trial access | Credits; Humanizer uses 1 credit per input word. Internal retries within one process do not consume additional credits. | Pricing page: Essential $1.00, Pro $1.45, Business $2.15, and Enterprise $7.15 per day as the displayed plan equivalents. |
| Input limit | Up to 30,000 characters and 3,000 words per input in the current web app. | Depending on the plan, 1,000, 1,500, 2,000, or 20,000 words per request; the API still recommends short outputs for optimal results. |
| Languages | Automatic language detection; the rewrite remains in the original's dominant language. | The current pricing page advertises 100+ languages; other official pages give different figures. |
| Modes and workflow | One controlled mode with local checks followed by an analysis report. | Heavy or Lite plus Quality or Fast mode. PhD applies to newly generated essays, not to the humanization of existing text. |
| Academic texts and citations | Existing citation blocks are protected and checked for completeness after the rewrite. | Writer and Agent provide separate research and citation features. Byte-exact protection in the Heavy humanizer is not documented. |
| Platforms and API | Web app with history, AI analysis, and plagiarism checking in the same workspace. | Web suite and API. The API costs $0.20 per 1,000 calculated words and counts both input and output. |
The site offers weekly, monthly, and annual plans, but the table displays daily equivalents. According to the pricing table, consumer subscriptions do not include API access.
Official product sources
Provider information verified on August 3, 2026
Prices and features may change. Provider claims were not treated as evidence of quality; only the separate Certentis benchmark is used for that assessment.
The comparison at a glance
A single AI detector score cannot decide the result. A useful rewrite must also remain natural, faithful to the source, stable in language, and dependable in its citations.
| Metric | Certentis | StealthGPT Heavy PhD | Stronger result |
|---|---|---|---|
| Overall quality | 85.4 | 70.0 | Certentis |
| Semantic fidelity | 91.7 | 72.8 | Certentis |
| Exact citations | 8/10 | 6/10 | Certentis |
| Correct language | 10/10 | 10/10 | Tie |
| Word ratio | 0.98 | 0.82 | Certentis |
| Human + Mixed | 92.96% | 68.75% | Certentis |
Certentis: 10 usable cases · StealthGPT Heavy PhD: 10 usable cases
Overall quality
The blinded review scored every output separately. These profiles show whether a tool merely sounds fluent or also preserves structure, precision, evidence, and meaning.
StealthGPT produced an unusually split profile: very strong grammar and academic style, but markedly weaker coherence and semantic fidelity across the full run.
| Metric | Certentis | StealthGPT Heavy PhD |
|---|---|---|
| Grammar & language | 93.9 | 91.6 |
| Clarity & readability | 91.8 | 82.6 |
| Coherence & structure | 89.4 | 73.0 |
| Academic style | 80.3 | 85.9 |
| Precision | 84.6 | 78.0 |
| Naturalness | 87.5 | 79.8 |
| Citation integration | 94.2 | 82.0 |
| Semantic fidelity | 91.7 | 72.8 |
Citeproc keys / Formatted citations
Five source texts used structured citeproc keys and five used already formatted references. That prevents format-specific weaknesses from disappearing inside an average.
On formatted citations, StealthGPT reached 86.2 quality points and narrowly exceeded Certentis. In citeproc, incomplete and heavily shortened outputs pulled quality down to 53.8.
| Metric | Certentis | StealthGPT Heavy PhD |
|---|---|---|
| Quality | 85.0 | 53.8 |
| Fidelity | 90.6 | 56.0 |
| Citation integration | 92.8 | 70.2 |
| Citations exact | 3/5 | 2/5 |
| Human + Mixed | 97.4% | 69.53% |
| Word ratio | 0.96 | 0.63 |
| Metric | Certentis | StealthGPT Heavy PhD |
|---|---|---|
| Quality | 85.8 | 86.2 |
| Fidelity | 92.8 | 89.6 |
| Citation integration | 95.6 | 93.8 |
| Citations exact | 5/5 | 4/5 |
| Human + Mixed | 88.51% | 67.98% |
| Word ratio | 0.99 | 1.02 |
Observation from the benchmark
These observations come from individual cases in the archived run. They explain the metrics, but they are not blanket claims about every future output.
Several outputs with standard references were fluent, terminologically precise, and dependable in their citation handling.
Large parts were missing in the German and Spanish citeproc cases; average length for this suite was only 63 percent of the source.
The review found foreign script fragments or inaccurate terminology in individual cases, even though the target language check passed.
What the numbers mean in practice
The right tool depends on whether you prioritize the highest detector classification, a particular writing style, or a controlled academic rewrite.
Certentis
different citation types must be processed in one stable workflow without major shortening.
StealthGPT Heavy PhD
the document mainly contains formatted citations and every output will be checked manually for completeness.
The 0.82 average word ratio means StealthGPT shortened the texts by about 18 percent on average.
Shorter outputs are not inherently worse, but in this run they coincided with lower semantic fidelity.
For academic text, preserving intended meaning and formal evidence matters more than maximizing the amount of rewriting.
Internal benchmark
Every provider received the same source texts. Outputs were scored against a fixed rubric without visible provider labels. Missing results, language drift, and changed citations remained visible.
Ten identical source texts with standard and structured citations
Blinded quality review across eight dimensions
Separate comparison of facts, qualifications, and intended meaning
Additional checks for language, word ratio, and complete citations
The results describe a dated run and do not guarantee every future output.
Provider models, prices, and product versions may change after the test.
Ten cases are a practical product test, not a complete representation of every text type.
Human + Mixed is an external document classification, not a percentage of human-written sentences.
Yes. Both produce more natural rewrites. Certentis additionally focuses on controlled preservation of meaning, language, and citations.
The measured outputs averaged 82 percent of the input word count. The benchmark describes that effect without claiming to know the competitor's internal cause.
Certentis preserved eight of ten complete citation sets; StealthGPT preserved six of ten.
No. This run is a dated sample of ten demanding multilingual texts. It reveals concrete strengths and risks, but it does not replace testing your own document.
The classification describes how an external scanner categorizes a document. It does not assess factual correctness, citation preservation, language stability, or argument quality.
Every reference present in the source had to remain complete and unchanged in the rewrite. A missing, reformatted, or damaged citation meant the case did not pass as exact.
Compare all seven tools
Every page uses the same benchmark run, so differences can be interpreted without changing the underlying dataset.
Overall quality: 82.4 · Semantic fidelity: 85.4
Read the full comparisonOverall quality: 42.7 · Semantic fidelity: 62.0
Read the full comparison