Explore independent research, real-world investigations, and the evidence behind our accuracy claims.
















































Research ›
Independent studies, internal tests and clearly stated limitations.
AI detection accuracy ›
Follow the model-development record.
Operational proof ›
Security, reliability, patents and NSF I-Corps.
In the press ›
Newsroom investigations powered by our research.
By use case ›
Evidence for publishing, education and enterprise teams.
Comparisons ›
Review Originality.ai alongside major alternatives.
API documentation ›
Bring consistent review into existing systems.
Data handling ›
Understand how information moves through the service.
01
Dataset
02
Threshold
03
Metrics
04
Version
05
Limits
AI detection accuracy
As generative models changed, our detectors evolved with them. Every result belongs to its model, dataset and threshold.
Nov 2022
1.1 Beta
Before ChatGPT
First commercial GPT-3-trained detector.
Feb 2024
1.1 Beta
98.8%
Accuracy on the toughest internal test; 2.8% false-positive rate.
Sep 2025
Lite · Turbo · Academic
Three models
Specialized models for lightweight, robust and academic review.
Jul 2026
AI Allowance
99.4%
At 15% allowance on Benchmark V6. Policy now shapes interpretation.
* Accuracy figures are not directly comparable across releases because datasets, generators and test conditions changed. Open the study for methodology, false-positive rates and limitations.

Publishers & content teams
Protect content integrity before publication.

Education & research
Support academic integrity with responsible review.

Enterprise & data teams
Turn AI-use policies into repeatable controls.
Decision factor
Originality.ai
Pangram
Mixed documents flagging AI involvement
93.6% · 936/1,000
59.2% · 592/1,000
Mixed classification under study rules
72.7% · 727/1,000
49.4% · 494/1,000
Published AI API usage · 100K words
$10
$50 standard · $40 advertised bulk
Policy control
Five AI Allowance levels
No native allowance selector listed
Our edge: mixed-writing recall, policy controls and workflow breadth.
Inspect the study record →Decision factor
Originality.ai
Turnitin
Primary workflow
Publishers, agencies, enterprise and education
None
Global plagiarism · 15% threshold
99.5%
98.0% · PlagScan (Turnitin)
Patchwork plagiarism · 15% threshold
51.5%
30.5% · PlagScan (Turnitin)
Content-integrity scope
AI + plagiarism + facts + readability + grammar
Similarity, plagiarism and academic AI review
Our edge: strong evidence and a broader integrity workflow.
Read the full Turnitin review →Decision factor
Originality.ai
GPTZero
Testing the Limits · ChatGPT
100%
97.3%
Testing the Limits · Gemini
100%
93.8%
Aggregated-detector study · false negatives
2%
10.2%
Content-integrity scope
AI + plagiarism + facts + readability + grammar
AI detection with education and writing-evidence tools
Our edge: strong evidence and a broader integrity workflow.
Read the full GPTZero review →Data handling
A plain-language view of how personal information moves through Originality.ai.
Collect what is needed
Account, usage and submitted content required to operate the service.
Use for defined purposes
Provide results, secure accounts, support users and improve the service.
Protect and limit access
Encryption in transit, access controls and authorized service providers.
Retain only as necessary
Retention follows legal, contractual and operational requirements.
Give people choices
Access, correction, deletion and applicable opt-out rights.
* Personal or Student Data is not used to train AI or large language models. Users can opt out of scan-text model improvement.

Vision and Invention
Jon Gillham
Founder & CEO

Vision and Invention
Maddie Lambert
Director of Marketing and Sales

Product and Engineering
Conor Watt
Co-founder & CTO

Operations and Scale
Narcis Bejtic
Director of Operations & Integrator

Customer Outcomes
Thomas Bongiorno
Head of Customer Success, Support & Enablement
Responsible use
Strong tools still need human judgment.
Set policy
Define acceptable AI use.
Inspect
Review the signals.
Verify
Add context and provenance.
Clarify
Ask the author.
Decide
Document the outcome.
Our standard
Five checks. No shortcuts.
Traceable
Source
Link to the direct publication or methodology.
Unblended
Label
Separate internal, independent and customer evidence.
Current
Date
Add a last-verified date to time-sensitive claims.
In context
Limits
Keep datasets, thresholds and constraints visible.
Living record
Maintain
Archive retired models and superseded claims.
Sources, security and comparisons in one file.