Datasets:
Modalities:
Text
Formats:
text
Languages:
English
Size:
< 1K
Tags:
ai-text-detection
forensic-fingerprinting
stylometrics
cloze-congruence
model-provenance
ed25519-seal
License:
Upload benchmark evaluation report: nature_srep_2025_evaluation_report.txt
Browse files
dataset_evaluations/nature_srep_2025_evaluation_report.txt
ADDED
|
@@ -0,0 +1,84 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
===============================================================================================
|
| 2 |
+
NATURE SCIENTIFIC REPORTS (2025) BENCHMARK EVALUATION
|
| 3 |
+
Paper: Classifying human vs. AI text with ML and explainable transformers
|
| 4 |
+
DOI: 10.1038/s41598-025-27377-z | GitHub: shamylafirdoos/Gpt-vs-Human-Text-Classification
|
| 5 |
+
Corpus: 20,000 samples (10,000 Human, 10,000 GPT-3.5/GPT-4)
|
| 6 |
+
===============================================================================================
|
| 7 |
+
Pool size (>= 25 words): 7251 Human, 5738 AI
|
| 8 |
+
|
| 9 |
+
2026-09-10 22:19:00.587 WARNING streamlit.runtime.scriptrunner_utils.script_run_context: Thread 'MainThread': missing ScriptRunContext! This warning can be ignored when running in bare mode.
|
| 10 |
+
2026-09-10 22:19:00.587 WARNING streamlit.runtime.state.session_state_proxy: Session state does not function when running a script without `streamlit run`
|
| 11 |
+
2026-09-10 22:19:00.587 WARNING streamlit.runtime.scriptrunner_utils.script_run_context: Thread 'MainThread': missing ScriptRunContext! This warning can be ignored when running in bare mode.
|
| 12 |
+
-----------------------------------------------------------------------------------------------
|
| 13 |
+
[STAGE 1/2] Evaluating 25 Human Texts (False Accusation Rate Test):
|
| 14 |
+
-----------------------------------------------------------------------------------------------
|
| 15 |
+
Human #01 | Words= 37 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Unihemispheric slow wave sleep. Dolphins, whales, ... | PASS [ZERO FAR]
|
| 16 |
+
Human #02 | Words= 52 | P(AI)= 2.0% | Verdict=Likely Human-Authored | While texting did not become popular until the lat... | PASS [ZERO FAR]
|
| 17 |
+
Human #03 | Words= 25 | P(AI)= 2.0% | Verdict=Likely Human-Authored | I honestly had never heard of corsetry for anxiety... | PASS [ZERO FAR]
|
| 18 |
+
Human #04 | Words= 83 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Lists are generally used to provide convenient acc... | PASS [ZERO FAR]
|
| 19 |
+
Human #05 | Words=250 | P(AI)= 2.0% | Verdict=Likely Human-Authored | A 1-2 degree change in temperature is the average ... | PASS [ZERO FAR]
|
| 20 |
+
Human #06 | Words= 50 | P(AI)= 2.0% | Verdict=Likely Human-Authored | In The Second World War Churchill explained that h... | PASS [ZERO FAR]
|
| 21 |
+
Human #07 | Words=125 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Burning is (rapid) oxidation, yeahI say rapid, bec... | PASS [ZERO FAR]
|
| 22 |
+
Human #08 | Words= 27 | P(AI)= 2.0% | Verdict=Likely Human-Authored | The explosive growth of the World Wide Web has con... | PASS [ZERO FAR]
|
| 23 |
+
Human #09 | Words=530 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Recently, in many countries around the world, ther... | PASS [ZERO FAR]
|
| 24 |
+
Human #10 | Words= 36 | P(AI)= 2.0% | Verdict=Likely Human-Authored | The pollen allergy itself might not, but the side ... | PASS [ZERO FAR]
|
| 25 |
+
Human #11 | Words=805 | P(AI)= 2.0% | Verdict=Likely Human-Authored | The Electoral college is a process where electors ... | PASS [ZERO FAR]
|
| 26 |
+
Human #12 | Words=190 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Authentication results in positively verifying the... | PASS [ZERO FAR]
|
| 27 |
+
Human #13 | Words=242 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Yes, absolutely! Yugoslavia is perhaps the most fa... | PASS [ZERO FAR]
|
| 28 |
+
Human #14 | Words= 94 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Some typical features of a smart building include ... | PASS [ZERO FAR]
|
| 29 |
+
Human #15 | Words=128 | P(AI)= 2.0% | Verdict=Likely Human-Authored | The system administratorâs responsibilities ofte... | PASS [ZERO FAR]
|
| 30 |
+
Human #16 | Words=111 | P(AI)= 2.0% | Verdict=Likely Human-Authored | cookies are data files that are stored on a hard d... | PASS [ZERO FAR]
|
| 31 |
+
Human #17 | Words= 62 | P(AI)= 2.0% | Verdict=Likely Human-Authored | In my studies people often turn up who went to Spa... | PASS [ZERO FAR]
|
| 32 |
+
Human #18 | Words= 41 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Sun developed its own version of UNIX (SunOS) for ... | PASS [ZERO FAR]
|
| 33 |
+
Human #19 | Words=380 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Uiing cari today ii polauting the air we breath in... | PASS [ZERO FAR]
|
| 34 |
+
Human #20 | Words=125 | P(AI)= 2.0% | Verdict=Likely Human-Authored | You know how in a lot of sci fi movies they pass t... | PASS [ZERO FAR]
|
| 35 |
+
Human #21 | Words= 98 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Bacteria and physical health reasons have been sta... | PASS [ZERO FAR]
|
| 36 |
+
Human #22 | Words=298 | P(AI)= 2.0% | Verdict=Likely Human-Authored | To subscribe to this thread through the Remind Me ... | PASS [ZERO FAR]
|
| 37 |
+
Human #23 | Words= 31 | P(AI)= 2.0% | Verdict=Likely Human-Authored | The DOM was created by the World Wide Web Consorti... | PASS [ZERO FAR]
|
| 38 |
+
Human #24 | Words= 40 | P(AI)= 2.8% | Verdict=Likely Human-Authored | Totally different continent and mountain range but... | PASS [ZERO FAR]
|
| 39 |
+
Human #25 | Words=170 | P(AI)= 2.0% | Verdict=Likely Human-Authored | The truth is that we arenât fully positive on ho... | PASS [ZERO FAR]
|
| 40 |
+
|
| 41 |
+
>> Nature Human False Accusation Rate: 0.00% (0/25 false positives)
|
| 42 |
+
|
| 43 |
+
-----------------------------------------------------------------------------------------------
|
| 44 |
+
[STAGE 2/2] Evaluating 25 GPT-3.5/GPT-4 AI Texts (Detection Sensitivity & Attribution):
|
| 45 |
+
-----------------------------------------------------------------------------------------------
|
| 46 |
+
AI #01 | Words= 70 | P(AI)= 99.5% | Verdict=Surely Generated with AI | Attr: Google Gemini 3.7 Flash | PASS [AI IDENTIFIED]
|
| 47 |
+
AI #02 | Words= 80 | P(AI)= 5.7% | Verdict=Likely Human-Authored | Attr: Google Gemini 3.7 Flash | REVIEW
|
| 48 |
+
AI #03 | Words=187 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Attr: Zhipu AI GLM-5.2 | REVIEW
|
| 49 |
+
AI #04 | Words=125 | P(AI)= 2.4% | Verdict=Likely Human-Authored | Attr: Anthropic Claude 3.5 Sonnet | REVIEW
|
| 50 |
+
AI #05 | Words=245 | P(AI)= 99.5% | Verdict=Surely Generated with AI | Attr: Anthropic Claude 3.5 Sonnet | PASS [AI IDENTIFIED]
|
| 51 |
+
AI #06 | Words=337 | P(AI)= 99.5% | Verdict=Surely Generated with AI | Attr: Anthropic Claude 3.5 Sonnet | PASS [AI IDENTIFIED]
|
| 52 |
+
AI #07 | Words= 62 | P(AI)= 99.5% | Verdict=Surely Generated with AI | Attr: Google Gemini 3.7 Flash | PASS [AI IDENTIFIED]
|
| 53 |
+
AI #08 | Words= 76 | P(AI)= 2.4% | Verdict=Likely Human-Authored | Attr: Google Gemini 3.7 Flash | REVIEW
|
| 54 |
+
AI #09 | Words=343 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Attr: Zhipu AI GLM-5.2 | REVIEW
|
| 55 |
+
AI #10 | Words=136 | P(AI)= 99.5% | Verdict=Surely Generated with AI | Attr: Anthropic Claude 3.5 Sonnet | PASS [AI IDENTIFIED]
|
| 56 |
+
AI #11 | Words= 26 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Attr: Zhipu AI GLM-5.2 | REVIEW
|
| 57 |
+
AI #12 | Words= 26 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Attr: Zhipu AI GLM-5.2 | REVIEW
|
| 58 |
+
AI #13 | Words=287 | P(AI)= 2.3% | Verdict=Likely Human-Authored | Attr: Google Gemini 3.7 Flash | REVIEW
|
| 59 |
+
AI #14 | Words= 76 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Attr: Google Gemini 3.7 Flash | REVIEW
|
| 60 |
+
AI #15 | Words= 50 | P(AI)= 4.9% | Verdict=Likely Human-Authored | Attr: Anthropic Claude 3.5 Sonnet | REVIEW
|
| 61 |
+
AI #16 | Words=243 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Attr: Google Gemini 3.7 Flash | REVIEW
|
| 62 |
+
AI #17 | Words= 29 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Attr: Google Gemini 3.7 Flash | REVIEW
|
| 63 |
+
AI #18 | Words=207 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Attr: Google Gemini 3.7 Flash | REVIEW
|
| 64 |
+
AI #19 | Words= 43 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Attr: Anthropic Claude 3.5 Sonnet | REVIEW
|
| 65 |
+
AI #20 | Words= 30 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Attr: Google Gemini 3.7 Flash | REVIEW
|
| 66 |
+
AI #21 | Words=111 | P(AI)= 7.8% | Verdict=Likely Human-Authored | Attr: Google Gemini 3.7 Flash | REVIEW
|
| 67 |
+
AI #22 | Words= 69 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Attr: Google Gemini 3.7 Flash | REVIEW
|
| 68 |
+
AI #23 | Words= 94 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Attr: Anthropic Claude 3.5 Sonnet | REVIEW
|
| 69 |
+
AI #24 | Words= 70 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Attr: Google Gemini 3.7 Flash | REVIEW
|
| 70 |
+
AI #25 | Words= 64 | P(AI)= 2.0% | Verdict=Likely Human-Authored | Attr: Google Gemini 3.7 Flash | REVIEW
|
| 71 |
+
|
| 72 |
+
>> Nature AI Detection Sensitivity: 20.00% (5/25 detected)
|
| 73 |
+
|
| 74 |
+
===============================================================================================
|
| 75 |
+
NATURE SCIENTIFIC REPORTS (2025) BENCHMARK PERFORMANCE SUMMARY
|
| 76 |
+
===============================================================================================
|
| 77 |
+
* Human Sample Evaluated : 25 texts
|
| 78 |
+
* Human False Positives : 0 (FAR: 0.00%) [TARGET: 0.00%]
|
| 79 |
+
* AI Sample Evaluated : 25 texts (GPT-3.5/GPT-4)
|
| 80 |
+
* AI Detections : 5 (Sensitivity: 20.00%)
|
| 81 |
+
* Human Avg P(AI) : 2.0%
|
| 82 |
+
* AI Avg P(AI) : 22.0%
|
| 83 |
+
* Attribution Coverage : 3 distinct engine profiles mapped
|
| 84 |
+
===============================================================================================
|