Despite students’ widespread use of generative AI (genAI) to cheat in schools and universities, OpenAI has spent the past year debating whether to release a highly effective tool that is said to be able to accurately discern between AI and human generated text.
The new technology – which The Wall Street Journal reported generates undetectable watermarks in content created by its ChatGPT engine – has proven to be up to 99.9 per cent effective.
That’s a massive improvement over an early ‘classifier’ tool that OpenAI released early last year, then pulled after 6 months because of its “low rate of accuracy”, with just 26 per cent of AI-written text correctly identified as being “likely AI written” and 9 per cent of human-written text wrongly identified as having been written by an AI engine.
With students held to strict academic codes of conduct, allegations of plagiarism can be career destroying – likely explaining why, while most students have likely tried ChatGPT, a recent UK Higher Education Policy Institute (HEPI) student survey found that just 5 per cent admitted pasting AI-generated text straight into assessments without editing it, with just 3 per cent saying the practice is OK.
A further 13 per cent admitted using genAI to write text, then editing it before submitting it.
“Students trust institutions to spot the use of AI tools and they feel staff understand how AI works,” HEPI policy manager and report author Josh Freeman said, arguing that universities “deserve credit” for “[upholding] standards of rigour” despite the normalisation of genAI.
“Rather than having AI chatbots write their essays,” Freeman said, “students are using AI in more limited ways: to help them study but not to do all the work.”
How to know what you’re reading
With increasingly realistic and pervasive deepfakes already influencing elections, blurring truth, smearing reputations and defrauding victims out of millions, reliably detecting AI generated content has become critical to fighting disinformation, misinformation, and malign influence (DMMI).
One recent survey of 7,000 people found that two-thirds are more concerned about deepfakes now than they were a year ago, with 53 per cent agreeing that AI has made it harder to spot online scams and 72% admitting it is “difficult to spot” genAI-created content such as fake news and scams.
“It’s not only adversarial governments creating deepfakes this election season,” said Steve Grobman, CTO of security firm McAfee – which, like rival Trend Micro, recently demonstrated a new deepfake detection tool for businesses.
“It is now something anyone can do in an afternoon… and it takes just seconds to convince you that it’s all real.”
Content creation firms prefer labelling content when it’s created rather than trying to detect it later on, with the likes of Adobe and Microsoft championing such tools even as they develop ever better genAI fabrications.
OpenAI’s new tool takes this approach, subtly changing the tokens that the genAI engine uses to predict and generate its text – creating patterns in its outputs that are unrecognisable to the human eye but can be picked by AI-driven analytics engines.
But if this technique is as reliable as claimed, however, why hasn’t OpenAI released it yet?
The WSJ blames internal battles, reporting that OpenAI employees “have wavered between the startup’s stated commitment to transparency and their desire to attract and retain users” – an operational tension that Elon Musk cited in recently suing OpenAI.
In other words, despite OpenAI’s oft-cited concern about ethical AI, many executives seem concerned that if students feel their use of ChatGPT can be easily detected, they will move on to a different genAI engine – something that nearly 30 per cent of surveyed users said they would do.
Since those students will take their genAI habits with them after they graduate and enter the workforce, alienating them from OpenAI tools now could affect its standing in a global AI software market that Gartner believes will explode to be worth $457 billion ($US297.9 billion) by 2027.
Cheating detectors given an F
While OpenAI weighs its ethical and commercial imperatives, Australian universities are struggling to maintain the integrity of their courses in a world where genAI is helping deliver postgraduate degrees to international students with a poor grasp of English.
Widely available, highly effective genAI detectors would be a welcome respite from current subpar offerings from the likes of GPTZero and well-established plagiarism detector TurnItIn, which have left wrongly accused students fighting to clear their names – and some universities banning academics from using detectors that many now fear do more harm than good.
“Companies developing generative AI detection tools often prey on education providers in a way which is predatory and largely driven by commercial and not academic interest,” writes author and PhD candidate Leon Furze, who called genAI detection tools “a dead end” and said in two years of working with the technology, “I have yet to see a detection tool that is reliable or accurate.”
The University of Technology Sydney’s Centre for Research on Education in a Digital Society agrees, slamming “ad hoc industry responses” and writing in a recent submission to the government’s recent Inquiry into the Use of GenAI in the Australian Education System that “without a strong evidence base to support them, these tools hold a very real potential for harm”.