How the explosion of AI 'study aids' is arming students to cheat

FE Week has uncovered significant evidence of the harm that generative AI is doing to the integrity of assessment, as educators and regulators are left struggling to catch up with the rapid pace of technological change

Students are now armed with the means to ‘cheat’ their way through almost any non-exam assignment by putting questions to a large language model (LLM) of generative AI such as ChatGPT, Snapchat’s My AI, and a multitude of new AI ‘study aid’ platforms. 

They’re also being exposed to growing numbers of online influencers endorsing these new tools on social media.

But, the AI detection tools that educators are being told by the Joint Council for Qualifications to use to detect AI plagiarism are themselves unreliable, giving rise to false accusations and a breakdown of trust between educators and students. 

Some educators are now questioning whether non-exam assessments are fit for purpose, with this summer’s exams series being seen as the first real test of the system against AI.

Social media AI minefield

There are numerous videos on TikTok of young people endorsing AI tools, sometimes purportedly as a study aid but often more blatantly for cheating. Most of these tools are third-party apps using ChatGPT’s language model. Most of their endorsers are attractive young females.  

But students can also cheat using the social media platforms they’re already socialising on.

Toby Rezio, an American Tiktoker with 92.5 million video likes, admitted to cheating using Snapchat’s My AI (see pic on next page).

Snapchat – used by 90 per cent of 13- to 24-year-olds – unleashed its ‘my AI’ chatbot powered by Open AI’s ChatGPT language model on all its users in April 2023. They customise their avatar with human features and it tells them to “ask [me] questions about anything”.

While increasing numbers of young people are being exposed to new methods of cheating, they’re not all receiving guidance from their colleges about what they can and cannot use AI for. 

A poll of 18 FE representatives at a recent meeting organised by the digital agency Jisc’s National Centre for AI found that a third were part of institutions which had produced AI guidance for learners, a third had not, with a third in the process of doing so.

The centre said its members had “noted that as AI becomes increasingly integrated into everyday tools, learners may encounter challenges in identifying what constitutes as AI”. It was also mentioned that “students may be aware of certain tools that staff are not”.

AI arms race

As chatbots become more sophisticated, new and updated detection tools are being flogged to educators to get the upper hand. Last month, The Joint Council for Qualifications refreshed its AI use in assessments guidance to include an expanded list of detection tools, including Turnitin AI, Copyleaks and Sapling.

These tools are engaged in a war against platforms such as SteathGPT, whose website brashly advertises how it “not only eludes the discerning eyes of Turnitin but also enhances the writer’s voice, ensuring that the work reflects their unique style and intellect”.

But detection tools merely provide a rough probability of whether a learner has relied on AI. FE Week analysis found by that most of the “potential indicators of AI misuse” cited by the JCQ, such as default use of American spelling and a lack of direct quotations, can easily be overcome by using further chatbot prompts to write in specific styles.

Michael Webb, technology director at JISC, says he “managed to beat every AI detector” by telling the chatbot he was using to “use the word ‘and’ less”.

A recent international study of 14 widely used AI detection tools (including Turnitin AI) found them to be “not accurate or reliable enough to use in practice”. Around 20 per cent of AI-generated texts would “likely be misattributed to humans”, rising to half of the AI-generated texts that “undergo some obfuscation”.

It found that some AI detection tools, such as Writer are “clearly aimed to be used to hide AI-written text, providing suggestions to users such as ‘you should edit your text until there’s less detectable AI content’”.

Daisy Christodoulou, director of education for No More Marking believes that AI is being used for cheating far more than most educators realise. Christodoulou warns teachers that “if you spot one AI-generated essay, there’s probably another ten you haven’t spotted. You just haven’t spotted the students using AI well.”

Educators describe feeling like they’re engaged in an arms race when it comes to AI tools. Harald Koch, the author of a book about AI cheating, believes that “the development of AI is progressing far too quickly” for the detection tools to be relied upon. 

“Before an AI checker has been rolled out in a meaningful way, the next level … of AI has already been released”. 

False accusations

The JCQ explicitly states that those accused of submitting AI-generated assignments “may attract severe sanctions”, including disqualification and being barred from exams. And if teachers with “doubts” about authenticity do not “investigate and take appropriate action”, this “can attract sanctions”.

The JCQ advises educators to use more than one detection tool and to consider “all available information”. Yet a growing number of students are now being falsely accused of AI cheating.

Daniel Sokol, lead barrister at Alpha Academic Appeals, says he now “deal[s] with AI-cheating cases regularly”.

In New Zealand last year, AI plagiarism detectors are believed to have falsely accused two high school students of cheating, with one parent describing the use of AI detection tools as playing “Russian roulette”.

An American study published in September 2023, which analysed 49 Reddit posts from college students accused of using ChatGPT, found 38 protesting their innocence. Another found seven AI detectors had wrongly flagged writing by non-native speakers as AI-generated 61 per cent of the time, compared to 20 per cent of human-written essays overall. 

Webb admits the detectors are “particularly prone to false positives when English isn’t your first language, because perhaps you’ve got a more formal style”. 

Caught in the act

As ChatGPT was released in November 2022, last summer was the first set of exams in which it was possible to cheat using generative AI. Cases of exam malpractice resulting in penalties rose from 4,105 in 2022 and 4,895 in 2023, although the proportion relating to tech devices remained the same (44.5 per cent).

The JCQ highlighted examples of students caught misusing AI on their coursework, including two AQA A-level history students, one of whom was disqualified, and two students on the OCR’s Cambridge nationals enterprise and marketing qualification who confessed to cheating and received zero marks. 

One candidate claimed in their defence to believe that using ChatGPT was “no different to asking a teacher for advice”. 

The JCQ last year advised educators to make students do some coursework “in class under direct supervision” to prevent AI misuse. But Christodoulou is calling for them to go further and “pause” all assessed coursework.

“The really high false accusation rate of AI detection tools can be corrosive for classroom relationships,” she warns. “And if a student knows that their friend is using AI and getting away with it, that’s destructive for the values you want to nurture. We must be really careful with these things.”

Scott Hayden, Basingstoke College of Technology’s head of digital, believes the sector needs “a moment to reflect” on AI’s impact. Assessments “need to change”, with essays replaced with “blogs, podcasts, and creative ways we can assess”. 

Claire Whiting, initial teacher education lead at Wiltshire College, posted online that she uses Vivas “with no notice” to assess whether a student has “depth of understanding”. 

“AI is here to stay, and is evolving too fast for staff training … essays are, quite frankly, useless to awarding organisations now.” 

But a new tool being marketed on TikTok, Ecoute, provides real-time AI-generated responses to “hack job interviews” and could potentially be used for any oral assignments. “Invisible” Bluetooth earbuds available to buy on Amazon would make it difficult for teachers to spot such activity.

The regulatory gap

DfE’s deputy director for digital, Bridie Tooher, admits that when it comes to AI, “things are moving so fast that … the tech will always overtake the regulations”. 

AI governance expert Kay Firth-Butterfield points out that in the US – the birthplace of most AI companies – there are “lots of ongoing court cases around LLMs using other people’s data without paying for it” which educators “need to bear in mind”.

“Google, OpenAI and Microsoft are underwriting court cases which may arise from misuse of their tools,” she added. 

Edtech investor Richard Taylor says because “the only real players” when it comes to AI edtech are based in America and China, this makes it more challenging to regulate the industry in the UK. “If you can’t control fundamentally the companies that are doing it, it puts us in a weak position.”

Educators raised concerns to DfE in its AI consultation that their students’ identity, grades or behaviour data may be input into Generative AI tools, with developers being “often opaque” about their use of such data. 

Thea Wiltshire, the Department for Business and Trade’s edtech specialist, says “we have to be very careful” in inputting students’ work into LLMs because “allowing generative AI to learn from it is an abuse of their intellectual property.”

Another key concern is around colleges and young people not adhering to the age restrictions of AI platforms. An Ofcom survey last year found that 74 per cent per cent of 16- to 24-year-olds in the UK have used a GenAI tool. For 13- to 18-year-olds, parental consent is required for ChatGPT. 

Webb says some educators see this as “just a US legal requirement” which is a “very risky strategy”.

Firth-Butterfield warns that in the business world, “the early adopters of AI are having to claw back what they’ve been doing because they didn’t put a good governance structure around AI in at the beginning”. She advises education leaders not to make the same mistake.

AI inequalities

Chatbots have the potential to level the academic playing field by giving more students access to personalised systems of learning, previously only available to those who could afford private tutoring. However, the digital divide could grow because of the rising cost of AI tools and access to online technology.

A recent Internet Matters poll found in households where income is less than £10,000 per year, only 11 per cent of children have used ChatGPT, rising to 45 per cent where income is £80,000.

And the most effective AI tools are the pricier ones: GPT-4, OpenAI’s premium LLM, costs £16 a month.

Webb estimates it costs a student around £80 a month for all the AI tools required to do well academically, giving those students “a significant advantage”. He admits “there’s no easy answer to that.”

Meanwhile, there is “inequality through variations of approach” within the FE sector. 

While some colleges have banned generative AI, others are encouraging students to embrace it and are using it to generate lesson plans, crunch data, and help mark assignments. 

Webb believes that new AI tools can be used by teachers to work “more efficiently and effectively”, allowing them to “spend more time building relationships with the learners, and help them develop soft skills”. 

However, a teacher over-reliant on AI to do all their prep for them could create more “disengaged learners”. “It’s “how we implement these tools, not just the tool itself.”

What are colleges doing?

In a packed session at the Association of Colleges conference last year hosted by Milton Keynes College, around half its attendees said they had either “never used or heard of ChatGPT”, and “two or three said they were experts”, said Alex Warner, principal of its South Central Institute of Technology who led the session. 

The college is choosing to embrace the AI revolution and gave every learner who joined last term an AI induction.

Basingstoke College of Technology is also making strides in engaging students around appropriate use of AI.

After a student survey revealed to Hayden that some of Basingstoke’s students were “hopelessly addicted to their devices and wanted help but were too embarrassed to admit it”, the college created a module on digital wellbeing, now delivered to all new learners.

But having spent the last seven years “encouraging the use of Snapchat and live streaming in the classroom”, Hayden admits he now feels “at a real crossroads” over “doing myself out of work” with AI. “I’m shaking my fist at those Google-Microsoft gods”.

A lack of official guidance around AI use from DfE is leaving many colleges unsure how to proceed. One teacher on Facebook recently described feeling “completely overwhelmed” and “unequipped” by the pace of change. 

Webb suggests that teachers “discuss with students how to use AI to create assessments that are authentic and getting them ready for work”. He admits this is “not easy” and that the problem is “compounded by the complexity around advice” from awarding bodies, which “differs enormously” between them. 

“Collectively, we need to get to a shared understanding.”

Making stuff up

In the meantime, colleges shouldn’t rely on LLMs too much, because they really can’t be trusted. The JCQ warns that chatbots “often produce answers which may seem convincing but contain incorrect or biased information”. 

“Some AI chatbots have been identified as providing dangerous and harmful answers to questions and some can also produce fake references to books or articles by real or fake people.”

Warner has concerns about “ethics, bias and transparency”. 

“The most frightening is where it makes up references by making assumptions. It’s riddled with flaws. But that’s why it’s not going to replace our jobs.”

There are also deeper philosophical considerations about the impact of AI on young people’s faith in democratic systems, and how AI will influence their curiosity for learning.

Firth-Butterfield claims in the US, AI is “helping students to come up from D and C to B [grades]. But it’s not helping to create excellence”.

She believes the tools can really help teach young people who had dropped out of school, been in jail or pregnant and missed large chunks of education to catch up”. 

“But it’s questionable whether these tools will help the people already doing well academically to develop innovation. We have to ask ourselves, do we want to become average? As humans, we need original thought to advance society.”

She’s particularly concerned about the impact of AI that interacts with young people in human-like ways, because “if it can understand how you feel, there’s the potential for it to manipulate”.

Webb says Jisc was in the process of drawing up guidance around AI when ChatGPT was unleashed, and has been in “reactive” mode ever since. But a “common roadmap” for the FE sector is in the pipeline.

“We just need to adapt and centre on what keeps us human. I’m hoping that AI will actually make us more human, because we’ve lost that as a society.”