Test Results
Photo: indiatoday

By Bidisha Saha: The tally of the number of jobs threatened by the world-changing generative AI technology depends on whether the GPT with human intelligence makes you more productive – or obsolete.

It is not surprising anymore if you come across a tweet, essay, or news article and wonder if it was written by artificial intelligence software. There are doubts regarding the authorship of a given piece of writing, like in academic context, employment, or the veracity of the content. Similarly, there are questions about the authenticity of the information, i.e., if a misleading idea suddenly appears in posts across the internet, is it spreading organically, or have the posts been generated by AI to create the appearance of real traction?

The core of a generative AI chatbot like ChatGPT, or GPT-3 is something called the Large Language Model (LLM), which is an algorithm that mimics the form of the written language. Although the backdrop of such an algorithm is complex and opaque, the idea is actually quite simple. GPT (Generative Pre-trained Transformers) are fed with heaps of internet information, to repeatedly predict the next few words and then use a grading mechanism to improve their response and write the real thing.

Tools to identify whether a piece of text was written by AI have started to emerge in recent months, including one created by OpenAI, the company behind ChatGPT. This tool uses an AI model trained to spot differences between generated and human-written text. But identifying AI-generated text is becoming increasingly difficult as software like ChatGPT continues to advance and turns out text that is more convincingly human.

To assess the effectiveness of current AI-detection technology, India Today’s OSINT(Open Source Investigations) team analyses some of these detectors using content generated artificially by OpenAI. The results show that AI detection services are advancing rapidly, but still fall short at times.

Three samples of excerpts written by OpenAI’s ChatGPT are vetted by online detectors like Winston, OpenAI’s AI Text Classifier, GPTZero, and CopyLeaks. For the first question, we wanted the GPT to summarise a piece of already published content. In this case, at least two detectors failed to identify the write-up written by ChatGPT, while only one predicted the result correctly and the other predicted a 50-50 chance.

In the second case, we asked ChatGPT to write on the environmental effects of a volcanic eruption which is a recurrent geological hazard, and information related to it is widely available on open-source. In this case, except for AI Text Classifier, all other platforms were able to predict the text as being artificially generated.

In the last try, we prompted ChatGPT to write an emotionally charged story with a shocking climax. The prompt was given to check whether the tool itself can generate content with artistic value. Interestingly, three of the four detectors failed to predict the content as being AI-generated.

Test Result

Detection technology has been heralded as one way to mitigate the harm from A.I. generated content. Many of the companies behind AI detectors acknowledged that their tools were imperfect and warned of a technological arms race. AI-detection companies say that their services are designed to help promote transparency and accountability, helping to flag misinformation, fraud, nonconsensual pornography, artistic dishonesty, and other abuses of technology.


Two reports were released separately by NewsGuard, a company that tracks online misinformation, and ShadowDragon, a company that provides resources and training for digital investigations.

As per the report, dozens of fringe news websites, content farms, and fake reviewers are using artificial intelligence to create fabricated content online. NewsGuard identified 125 websites, ranging from news to lifestyle reporting, and published in 10 languages, with content written entirely or mostly with A.I. tools.

The sites included a health information portal that published more than 50 AI-generated articles offering medical advice. Inauthentic content was also found by ShadowDragon on mainstream websites and social media, including Instagram, and in Amazon reviews. The websites were often littered with ads, suggesting that the inauthentic content was produced to drive clicks and fuel advertising revenue for the website’s owners.

The misleading AI content included fabricated events, medical advice, and celebrity death hoaxes, the reports said, raising fresh concerns that the transformative technology could rapidly reshape the misinformation landscape online. The looming risk of automation also contributes to threatening a number of jobs even sooner than we can imagine.

Although, Google took a stance last year that it does not like AI-generated content and some even speculated that Google may as well penalise those that are using AI to write content. But, Danny Sullivan from Google recently said that any content written only for the search engines primarily for the rankings will not be preferred, irrespective of the fact if it has been written by an AI or a human.


The decoding mechanism of ChatGPT, which is based on the Large Language Model, considers many options for each generated word, taking into account the response it has written so far and the prompt being asked. It then assigns a score to each option on the list, which quantifies how likely the word is to come next, based on the vast amount of human-written text it has analyzed and the patterns and relationship of words with phrases. Then, it chooses a word with a high score and moves on to the next one.

The LLM’s output is often so sophisticated that it can seem like the chatbot understands what it is saying – but it does not. Every choice it makes is determined by complex math and huge amounts of data. So much so that it often produces text that is both coherent and accurate.

Researchers from the University of Maryland have discussed how the watermarking methods in language models work. With watermarking, the maker of the model delivers its model output with unique fingerprints that is unnoticeable for humans but gets flagged with the help of statistics. It works by ’embedding hidden markers or signatures within the generated text to enable traceability and ownership verification.’

Let’s say, we have 50 thousand English words that the language model knows about, referred to as the ‘carrier dataset.’ When the GPT is sampling the most probable word from the predicted probability distribution, it modifies the output by blacklisting at least 20 percent of the carrier dataset and introducing specific changes to the training examples all while keeping the overall meaning intact. These changes include word substitutions, deletions, and reordering that are carefully crafted to encode the watermark.

So, the language model for decoding can choose from the rest of 80 percent of the words. The seed of the random number generator that chooses which words are blacklisted is the last word of the input. This allows the blacklisted subset of words to be reconstructed each time. This procedure is applied at each generation of the next token, where for each next word, the last word is used as a seed to blacklist 20 percent of the carrier dataset.

To detect generated text from this language model, the detector goes through the generated text and counts the blacklisted words. The watermarked language model would not use the blacklisted words, but humans would definitely use the blacklisted terms. A detection tool that knew which words were on the special list would be able to tell the difference between generated text and text written by a person.

If someone tried to remove a watermark by editing the text, they would not know which words to change. And even if they managed to change some of the special words, they would most likely only reduce the total percentage by a couple of points.

Still, the ‘one-stop’ tool that can reliably detect all AI-generated text with total accuracy may be out of reach, unless the companies in the AI field agree on a standard watermark implementation and would require a lot of engineering.

Read The Full Article

Leave a Reply

Your email address will not be published. Required fields are marked *