We already know that generative AI replicates human biases, but it turns out that the tools used to identify it do the same.
There are many issues with programs designed to differentiate between chatbot text and human writing. Now, a new problem has been added to this list: AI detectors often mistakenly classify non-native English speakers' writing as AI-generated. According to a study published in the journal "Patterns" on Monday, AI detectors misidentify writing by native English speakers as AI-generated more than half of the time.
In a world where generative AI is ubiquitous, it is becoming increasingly important to distinguish between AI-generated "gibberish" and text written by real humans. Job seekers, students, and those who are regularly evaluated based on their writing abilities should be able to submit their work without worrying about being mistaken for computer-generated content. At the same time, teachers, professors, and hiring managers should be able to know when someone is honestly showcasing their abilities and when they are not.
Differentiating between a person's work and chatbot automation or algorithmically determined output is becoming increasingly difficult due to the training of larger language models on massive datasets. Just as image, audio, and video deepfakes have become disturbingly difficult to discern, the recognition of AI-generated text is becoming more challenging.
Several companies have already begun trying to solve this problem by developing AI detection software that can parse out a person from a pile of text. Even OpenAI, a company at the forefront of generative AI, has attempted to create AI detection tools. Despite claims by developers of metrics such as "99% accuracy" being unverifiable, most AI detection tools have been ineffective or have limited use cases.
In addition to being generally not as good, these tools may also reproduce human biases, just like generative AI itself.
In this new study, researchers evaluated 91 TOEFL (Test of English as a Foreign Language) essays written by non-native English speakers using seven "widely used" GPT detectors. As a comparison, they also tested essays written by 99 eighth-grade students in the United States using the same set of AI detection tools. While these detectors correctly classified over 90% of the eighth-grade students' essays as human writing, their performance was not as good with TOEFL essays.
The average error detection rate for non-native English speakers' essays was 61.3% across all seven GPT detectors. At least one detector incorrectly labeled nearly 98% of the TOEFL essays as AI-generated. All detectors consistently identified the same batch of approximately 20% of TOEFL essays as AI-generated, even though they were written by humans.
The authors of the study explained that most AI detectors work by evaluating the "perplexity" of the text. Perplexity is essentially a measure of how unexpected a word's appearance is in the context of a series of texts. If a word can be easily predicted based on previous words, then theoretically, the sentence is more likely to be AI-generated because these large language models use probabilistic algorithms to generate convincingly well-organized word salad, known as "strong auto-completion."
However, non-native speakers of any language often write with a limited vocabulary and predictable grammar, making sentences and paragraphs easier to predict. The researchers found that simply reducing vocabulary repetition in TOEFL sample essays significantly reduced the number of false positives in AI detection software. Conversely, simplifying the language of eighth-grade essays resulted in more essays being mistaken for AI-generated.
As the new study points out, this could be a significant problem for non-native English speakers who already face discrimination in the job market and academic environments. On the broader internet, persistent AI detector errors could amplify existing inequalities.
The researchers wrote, "On social media, GPT detectors may mistakenly label non-native content as AI plagiarism, causing unjust distress for specific non-native communities. Mechanisms implemented by search engines like Google can devalue AI-generated content, potentially limiting the visibility of non-native communities and potentially suppressing diverse perspectives."
Source: https://gizmodo.com/ai-detectors-discriminate-against-non-english-speakers-1850622278
Alphabet Q1 paid subscribers reached 350M, fueled by YouTube & Google One. YouTube ad revenue slightly missed at $9.88B amid Premium growth. Total Q1 revenue hit $109.9B, beating forecasts.
Anthropic releases Claude connectors for Adobe, Blender, and Ableton, enabling AI workflow automation and app control. It also sponsors the Blender Development Foundation.
Apple introduces an App Store model splitting 12-month plans into monthly installments. On iOS 26.4+, it lowers upfront costs without reducing total price, adding billing tracking.
Google launches Aadhaar digital IDs in Wallet for India, enabling secure QR/NFC verification. A phased partner rollout targets key sectors, while passport-based credentials expand globally.
YouTube tests "Ask YouTube," an AI tool blending text and video for complex queries. Initially for US Premium users, it supports interactive follow-ups, expanding Google's AI search.
Google grants US DoD classified AI access, joining OpenAI and xAI in defense deals amid Anthropic's dispute. Contractual safeguards against surveillance and lethal weapons remain unenforceable.
Mistral AI launches Workflows, a Temporal-powered orchestration platform for secure, Python-driven enterprise AI automation featuring human-in-the-loop approvals.