Generative AI Detectors Are Discriminating Against Non-Native English Speakers

Artificial Intelligence (AI) has profoundly revolutionised our working world, changing how we learn, communicate, and work – but emerging concerns around bias in AI could reveal huge vulnerability.

However, a new problem has begun to ripple beneath the surface of this technological tidal wave.

It seems AI detectors, designed to identify whether content is AI-generated or human-written, are biased against non-native English speakers.

This assertion comes from a significant study recently conducted by Stanford scholars.

Do AI Detectors Even Work?

AI detectors, developed as a response to the breakthrough launch of OpenAI’s ChatGPT, have gained considerable attention.

Promising to distinguish between human and AI-produced content, they serve as a barrier against AI-induced cheating, plagiarism, and disinformation.

However, the Stanford study punctures this image of infallibility, revealing that these detectors are not just unreliable but exhibit bias against non-native English speakers.

The research paints a sobering picture – while the AI detectors performed almost impeccably when evaluating essays written by US-born eighth-graders, the results were dramatically different for TOEFL essays penned by non-native English speakers.

These detectors falsely classified an alarming 61% as AI-generated – the far-reaching implications of such erroneous judgement could disadvantage non-native speakers in academic and professional environments.

I wrote an academic paper from scratch and running it on Ai detector indicated it was written using AI. That shit a scam bruh

Now I had to re write it until AI gives me a clean pass.

— Cerebral Assassin (@xysist) May 17, 2023

This bias in the AI detectors can be traced back to their use of ‘perplexity’ as a scoring metric, which measures the sophistication of writing.

Naturally, non-native speakers tend to score lower on this metric due to differences in lexical richness, diversity, and grammatical complexity.

Stanford Research Highlights AI Ethics Concerns

Stanford’s James Zou, a professor of biomedical data science, emphasises the ethical dilemmas arising from this issue.

Unfounded accusations or penalties for cheating could be levied against non-native speakers due to the skewed objectivity of AI detectors.

The study also shows that these detectors can be deceived by ‘prompt engineering’, a process of instructing the AI to rewrite text in a more sophisticated language.

Consequently, Zou advises against using detectors in educational settings teeming with non-native English speakers.

He also urges developers to reassess their reliance on perplexity as the predominant metric, while hardening their models against circumvention.

Stanford’s study underscores the inherent challenge of developing impartial AI systems.

I wrote an article with the help of ChatGPT.

Then I pasted the article into an AI content detector.

The only sentences it highlighted were two I’d written to fill in and correct where GPT missed.

Hah.

Another checker gave me the score below. The article is probably 80% AI… pic.twitter.com/Wtsml4rKYV

— Anthony Garone (@atgarone) May 20, 2023

It serves as a timely reminder that while AI promises substantial advancements, it’s not without its shortcomings.

Developers and users share the onus of carefully considering AI’s limitations, ensuring it doesn’t inadvertently reinforce discrimination or bias.

As AI continues to evolve, our comprehension and management of its implications must also mature.

This analysis illustrates that the march of AI isn’t simply a story of unbroken progress – it’s a complex interplay of human decisions, algorithms, and every piece of content AI generates or evaluates.

It’s therefore paramount that we remain alert, scrutinising, and critical of AI’s development and its role in our society.

RELATED:

• What Are the 16 Best AI Writer Tools for 2023?
• Beware of the AI Bubble – Will it Pop Like the Dotcom Bubble?
• What Are the 13 Best AI Essay Writing Tools for 2023?