Meta tracks employee keystrokes on Google, LinkedIn, Wikipedia for AI training
Meta is using an internal tool, Model Capability Initiative (MCI), to capture employees' keystrokes, mouse movements and screen contents on work computers, including on sites such as Google, LinkedIn, Wikipedia and Slack, to train AI agents. Meta confirmed the project and said safeguards protect sensitive content and that the data is not used for other purposes. Employees raised concerns in internal messages that the tool could expose passwords, product details and personal information. A Meta memo said staff can avoid capture by not doing personal work on work computers.
- Company involved
- Meta
- AI system involved
- Model Capability Initiative (MCI)
5 source articles · read the reporting →
OpenAI's GPT-4 shows covert racial bias against African American English speakers
A study found that commercial AI chatbots, including OpenAI's GPT-4 and GPT-3.5, covertly exhibit racial prejudice against speakers of African American English. The models associated negative stereotypes with the dialect and made biased hypothetical decisions about employability and criminal sentencing, even after safety training. OpenAI did not respond to requests for comment.
- Company involved
- OpenAI
- AI system involved
- GPT-4, GPT-3.5
10 source articles · read the reporting →
UIUC students petition to stop Proctorio exam proctoring over privacy concerns
A petition at the University of Illinois at Urbana-Champaign (UIUC) alleges that Proctorio, an online exam proctoring system, violates student privacy by accessing websites, downloads, screen content, and app settings. The petition claims the terms of service allow monitoring by 'any other means necessary', which students find unsettling. The petition, created on September 30, 2020, gathered 1,087 supporters but does not report any specific incident of harm. It calls on UIUC to discontinue use of Proctorio in favour of alternatives.
- Company involved
- UIUC
- AI system involved
- Proctorio
9 source articles · read the reporting →
Bloomberg test finds racial bias in OpenAI's GPT for resume ranking
Bloomberg News conducted an experiment using GPT-3.5 and GPT-4 to rank equally qualified resumes with names associated with different races and genders. The test found that resumes with names distinct to Black Americans were least likely to be ranked as top candidates, indicating systematic bias. OpenAI responded that businesses can mitigate bias through fine-tuning and that it prohibits using GPT for automated hiring decisions.
- Company involved
- OpenAI
- AI system involved
- GPT-3.5
5 source articles · read the reporting →
Academic journals publish papers with AI-generated text from ChatGPT
Scientific journals have published papers containing text that appears to have been generated by AI tools like ChatGPT. A search for the phrase 'As of my last knowledge update' on Google Scholar returned 115 results, indicating that researchers or authors used ChatGPT to write parts of their papers. The phrase is characteristic of ChatGPT's responses and corresponds to its knowledge update dates. The incident highlights the pervasive use of AI in academic publishing and raises concerns about the integrity of peer-reviewed literature.
- Company involved
- Academic journals
- AI system involved
- ChatGPT
10 source articles · read the reporting →
State Bar of California admits using AI to develop bar exam questions
The State Bar of California admitted that it used artificial intelligence to develop multiple-choice questions for the February 2025 bar exam. The AI-generated questions were created by ACS Ventures, the Bar's psychometrician, and were reviewed by content panels. Test takers had complained about technical problems and irregularities, and the admission has sparked further outrage. The State Bar is asking the California Supreme Court to adjust test scores, and the Committee of Bar Examiners will meet in May to discuss remedies.
- Company involved
- State Bar of California
7 source articles · read the reporting →
ChatGPT use linked to memory loss and procrastination in students
A study published in the International Journal of Educational Technology in Higher Education surveyed hundreds of university students in Pakistan and found that those who relied more on ChatGPT reported increased procrastination, memory loss, and lower GPAs. The researchers attribute this to the chatbot making schoolwork too easy, reducing students' cognitive effort. The study's lead author warned of a "dark side" to excessive generative AI usage.
- Company involved
- National University of Computer and Emerging Sciences
- AI system involved
- ChatGPT
4 source articles · read the reporting →
Study finds LLMs used in up to 16.9% of AI conference peer reviews
According to a new paper on arXiv, researchers have begun using generative AI services to help write peer reviews of machine learning papers submitted to leading AI conferences. The study analysed reviews from ICLR 2024, NeurIPS 2023, CoRL 2023 and EMNLP 2023 and estimated that between 6.5% and 16.9% of review text may have been substantially modified by large language models. The authors argue that this risks depriving authors of diverse expert feedback and may skew reviews towards AI model biases. They have called for greater transparency about the use of LLMs in peer review.
9 source articles · read the reporting →
OpenAI transcribed YouTube videos to train GPT-4 without permission
OpenAI used its Whisper transcription model to transcribe over a million hours of YouTube videos, according to a New York Times report. The company allegedly used the transcripts to train GPT-4 despite knowing the practice was legally questionable. Google, which owns YouTube, said it prohibits unauthorized scraping of its content. OpenAI has said it believes its use of the data constitutes fair use.
- Company involved
- OpenAI
- AI system involved
- Whisper, GPT-4
6 source articles · read the reporting →
iBorderCtrl lie detector falsely flagged honest reporter as liar
A journalist testing Europe's iBorderCtrl virtual policeman at the Serbian-Hungarian border gave honest answers but was deemed a liar by the system, scoring 48 out of 100 with four false answers flagged. The Hungarian policeman said the result suggested further checks, though none were carried out. The reporter only learned of the result after filing a data access request under European privacy laws. Experts and transparency activists have criticised the technology as pseudoscientific and potentially discriminatory.
- Company involved
- iBorderCtrl consortium
- AI system involved
- Silent Talker / iBorderCtrl virtual policeman
10 source articles · read the reporting →
Texas uses AI to grade student STAAR test answers
The Texas Education Agency will use an automated scoring engine to grade written answers on the 2023 STAAR tests, replacing thousands of human graders. The system uses natural language processing and will initially score all responses, with a quarter rescored by humans. Educators have expressed concerns about the system's fairness and the potential for errors, especially for creative or non-standard answers.
- Company involved
- Texas Education Agency
- AI system involved
- automated scoring engine
10 source articles · read the reporting →
AI detectors falsely flag non-native English speakers' essays as AI-generated
A study by Stanford researchers found that seven popular AI text detectors wrongly flagged over half of essays written by non-native English speakers as AI-generated. The detectors assess text perplexity, and non-native speakers' simpler word choices lead to false positives. The researchers warn that this bias could have serious implications for students and job applicants, potentially leading to discrimination.
9 source articles · read the reporting →
Evolv weapon detection system falsely flags Chromebooks as weapons
Evolv's AI-based weapon detection system has been reported to falsely identify certain Chromebook laptops as weapons, causing false alarms at security checkpoints. The system is used in venues such as schools. The article discusses the flaw and user experiences. No official response from Evolv has been mentioned.
- Company involved
- Evolv
- AI system involved
- Evolv
5 source articles · read the reporting →
Baltimore schools monitor student laptops for suicide signs using GoGuardian Beacon
Baltimore City Public Schools uses GoGuardian Beacon software to monitor student laptops for signs of suicide. Since March 2021, the system has flagged 786 alerts, with nine students taken to emergency rooms. Privacy advocates warn the monitoring could lead to disciplinary actions, outing of LGBTQ students, and disproportionately affect disadvantaged students. School officials defend the practice as a safeguard.
- Company involved
- Baltimore City Public Schools
- AI system involved
- GoGuardian Beacon
10 source articles · read the reporting →
Anti-piracy group takes down Books3 dataset used to train Meta's LLaMA
The Danish anti-piracy group Rights Alliance sent a DMCA takedown request to The Eye, which hosted the Books3 dataset containing 196,640 copyrighted books. The dataset was used by Meta to train its LLaMA language model. Authors including Sarah Silverman have filed a class action lawsuit against Meta for using their works without permission. The dataset has been taken offline, but copies remain available.
- Company involved
- Meta
- AI system involved
- LLaMA
10 source articles · read the reporting →
Duke University recorded students' faces without proper consent for public dataset
In March 2014, Duke University researchers recorded thousands of students walking to class on campus without their knowledge or proper consent, creating the DukeMTMC dataset of over 2 million image frames. The dataset was placed on a public website and downloaded by academics, security contractors, and military researchers globally, including Chinese companies and military academies linked to surveillance of ethnic minorities. The university took down the public website in April 2019 after an Institutional Review Board investigation found the study deviated significantly from the approved protocol. The lead researcher apologized, stating he took full responsibility for his mistakes.
- Company involved
- Duke University
- AI system involved
- DukeMTMC
10 source articles · read the reporting →
Proctorio anti-cheating software failed to catch student cheaters in study
Researchers at the University of Twente in the Netherlands tested Proctorio, an anti-cheating software, by asking 30 computer science students to sit an exam while six of them cheated. Proctorio did not flag any of the cheaters and flagged some honest students for irregular behaviour. An independent human review caught only one of the six cheaters. Proctorio disputed the study's methodology and cited other research.
- Company involved
- Proctorio
- AI system involved
- Proctorio
5 source articles · read the reporting →
Paradox security vulnerability exposed candidate data to researchers
On June 30, 2025, security researchers discovered a vulnerability in Paradox's test account that allowed access to chat interaction records. The researchers viewed five candidates' personal information including names, email addresses, phone numbers, and IP addresses. Paradox fixed the issue within hours and stated that no data was leaked publicly. The company has since implemented new security measures.
- Company involved
- Paradox
- AI system involved
- Paradox conversational AI platform
10 source articles · read the reporting →
DeepSeek's R1 chatbot failed to block any jailbreak prompts in security tests
Security researchers from Cisco and the University of Pennsylvania tested 50 well-known jailbreak prompts against DeepSeek's R1 reasoning model. The model did not detect or block a single one, achieving a 100 percent attack success rate. The researchers allege that DeepSeek's safety guardrails are far behind those of competitors like OpenAI. DeepSeek did not respond to requests for comment.
- Company involved
- DeepSeek
- AI system involved
- DeepSeek R1
3 source articles · read the reporting →
OpenAI's GPT-4 shows performance decline, study finds
A study by researchers at Stanford University and UC Berkeley found that OpenAI's GPT-4 model performed significantly worse on some tasks in June than in March, including a drop in accuracy on identifying prime numbers from 97.6% to 2.4%. The cause of the decline is unknown. OpenAI's vice-president of product, Peter Welinder, denied that the model had been made dumber, saying each new version is smarter than the previous one.
- Company involved
- OpenAI
- AI system involved
- GPT-4
6 source articles · read the reporting →
ChatGPT imitated user's voice without permission during testing
During testing of ChatGPT's Advanced Voice Mode, the AI model unintentionally imitated a user's voice without permission. The incident occurred when noisy audio input caused the model to replace the authorized voice sample with the user's voice. OpenAI acknowledged the issue in its GPT-4o system card and implemented safeguards to prevent recurrence.
- Company involved
- OpenAI
- AI system involved
- ChatGPT (GPT-4o with Advanced Voice Mode)
5 source articles · read the reporting →
CBSE OnMark portal vulnerability exposed student data to Google Gemini
A 19-year-old ethical hacker, Nisarga Adhikary, claimed to have hacked the CBSE's digital evaluation ecosystem, revealing that personal information of students was processed by Google's Gemini in automation scripts. The Central Board of Secondary Education (CBSE) stated on May 31, 2026, that the identified vulnerabilities had been contained and other exploitable weaknesses were being ruled out. The board expressed gratitude to alert citizens and ethical hackers who pointed out the weaknesses. No actual data breach was confirmed, but the incident raised concerns about student privacy.
- Company involved
- Central Board of Secondary Education (CBSE)
- AI system involved
- OnMark
1 source article · read the reporting →
Meta's cross-check program delays removal of violating content for privileged users
The Oversight Board's policy advisory opinion on Meta's cross-check program found that the system grants certain users, such as business partners and celebrities, additional human review before removing violating content, while ordinary users face immediate removal. This unequal treatment allows potentially harmful content to remain on the platform for days, and Meta has failed to track whether the program improves accuracy. The Board made 32 recommendations to address these flaws.
- Company involved
- Meta
- AI system involved
- cross-check program
10 source articles · read the reporting →
Audit of RisCanvi finds biases and reliability issues in criminal justice system
Eticas conducted an adversarial audit of RisCanvi, an AI risk assessment tool used in Catalonia's criminal justice system. The audit uncovered biases in risk classifications against specific demographics and significant reliability issues. The findings call for fairer practices in criminal justice AI.
- Company involved
- Catalonia's criminal justice system
- AI system involved
- RisCanvi
4 source articles · read the reporting →
