The development of A.I. leads to asking the Turing question over and over again: Is this information of human origin? Am I dealing with a human or a robot?
To better understand the logic behind detection tools, I invite you to read this interesting article by Olivier Duffez with two AI and SEO specialists: Vincent Terrasi and Sylvain Peyronnet - The state of the art of AI-generated content detection - presents different methods to identify the origin of texts and their limits. A real race is on.
- An artificial intelligence does not make mistakes, neither in spelling nor in grammar.
- An artificial intelligence proceeds by probabilistic judgment; it takes the most likely and most often encountered explanations. It rarely uses rare words because those rare words are rarely used. Its language is generally composed of common words.
- The algorithms used each have their signature, their fingerprint, which can eventually be recognized.
- An artificial intelligence ensures the consistency of its texts by taking into account the context. One does not encounter mixtures of terms without context.
- Algorithms are characterized by their results such as the frequency of certain words, the density of keywords, and rules of language that they follow "to the letter."
- An A. I.A. can write implausibilities without any restraint.
From these kinds of observations, here are services that calculate the degree of probability that a text is or is not written by an artificial intelligence, a new kind of Turing Test.
- AI Detector Pro - by Westbourne International... so you don't have to pay a lot for machine-generated content
- Compilatio AI Detector by Compilatio - multilingual
- Content at scale AI detector - by Slone &McGill with marketing ambition
- Copyleaks AI Content Detector - by Copyleaks, 99.12% efficiency... claims to be verified.
- Crossplag AI Content Detector - Artificial intelligence learns to detect artificial intelligence.... as long as we are not asked for a DNA test to verify that we are not a robot.
- Draft&Goal - Specifically calibrated for GPT chat but its approach will allow it to detect other A.I. signatures - In French and English.
- GPT-2 Output Detector Demo based on RoBERTa (A Robustly Optimized BERT Pretraining Approach)
- Giant Language model Test Room by MIT-IBM Watson AI lab and Harvard NLP)
- GPT Radar by Neuraltext
- GPT Zero - by Edward Tian of Princeton University
- Grover - by the Allen Institute with a broader ambition of detecting fake news
- Open AI Text Classifier - by Open AI, the creator of GPT Chat. With reliable data on the detector's true effectiveness.
- Originality - Starting as a plagiarism checker. now augmented with an AI checker.
- Sapling - expands its AI offerings with this service. Knowing what is human and what is not is really a curious ethical question.
- Writer AI Content Detector - by Writer - Free version limited to 1500 characters
Many of us submitted this text to them and in every case it was certified 99% not composed by an A.I.
We can assume that other such tools will soon follow and that we will come to create A.I.s that will each have their own identifiable writing style.
Will they go so far as to imitate the writing style of middle schoolers or
children? There is no reason to believe that they would not be
capable of doing so, as they are already doing for art, as with Deep Art Effect.
Illustration: Deposit Photos - paulfleet
Artificial Intelligence in a frenzy - Try more of the dozens of public
Knowledge as a public good to be seized
See more technologies from this institution