A new study by Seulki Lee-Geiller at UAlbany shows that political bias shapes how voters view AI in campaigns. Surveying over 3,000 participants, the research found that people trust AI-generated campaign messages more when their own party is involved and question fairness when the opposing party uses similar tools. The findings suggest that perceptions of AI in politics are driven less by awareness of the technology and more by partisan loyalty.
Meanwhile, marketing agencies are grappling with the practical side of AI adoption. Leaders from Monks and M&C Saatchi spoke at Digiday's Media Buying Summit about how agencies have adopted AI agents without updating team structures or job roles. Companies are now rethinking hiring for junior positions and limiting access to certain tools. The challenge is capturing AI's productivity gains without making daily work more tedious for staff.
In higher education, experts continue to debate how AI should be used in classrooms. The Gazette reported on ongoing discussions about best practices, highlighting that institutions are still searching for clear guidance on integrating the technology into teaching and learning in meaningful ways.
Concerns about oversight have grown as AI systems become harder to monitor. A Fast Company article warns that both human-in-the-loop and human-on-the-loop models assume a real person is present, but AI can now imitate voices, faces, and behavior convincingly. The piece calls for stronger identity verification to make sure authorized humans are actually the ones supervising AI decisions.
Asian religious leaders are also weighing in on AI ethics. Bishops and scholars meeting in Macau late last month urged that AI serve human dignity. Bishop Marcelino Antonio Maralit Jr. cautioned against deepfakes, surveillance, and AI-generated homilies presented as original work. Dr. Levi Checketts pointed to the environmental and labor costs of technology, while the assembly discussed supporting over 100 digital missionaries across 17 Asian countries who face online risks.
On the biosecurity front, Stanford researchers used AI models Evo 1 and Evo 2 to design nearly 300 viral genomes. Sixteen of those designs became living viruses that attacked harmless lab bacteria targeting ΦX174. The team found that half the working viruses had mutations unlikely to occur naturally, showing AI is learning design rules from nature. However, researchers emphasized that humans still control safety and biosecurity protocols.
Japan is responding to a surge in cyberattacks against its companies by preparing an emergency cybersecurity package and reviewing its AI Promotion Act. Privacy experts are pushing for tougher data-security rules and better compensation mechanisms. The attacks have exposed businesses to significant damages claims, and the debate now centers on preventing misuse of AI and protecting personal data more effectively.
Cohere finds itself at the center of a policy debate in Canada. The government plans to invest up to $1.6 billion in the company through a planned $7 billion funding round. Critics argue that governments should not act as investors in private companies and that better tax policies would serve the AI sector more effectively. The article warns that state-led funding can create dependency and may actually hurt genuine innovation, especially since Cohere has already raised billions from private investors.
In legal settings, a tool designed to detect AI-generated evidence in courts is proving difficult to build. Canadian researcher Maura Grossman leads a team developing a free open-source system that uses multiple algorithms voting together to identify AI content. After two years of work, the tool shows promise in lab tests but struggles in real-world court use, where the algorithms could not reach enough agreement to be reliably helpful.
Separately, research on federated learning for deepfake detection raises questions about explainability. Testing an EfficientNet-B4 detector on FaceForensics++ across five federated clients, researchers found that explanation quality depends heavily on data heterogeneity. When data setups are nearly identical, federated models match or beat centralized ones. But when data differences are extreme, the AI explanations become misleading, offering important guidance for when federated deepfake detection can be trusted.
Key Takeaways
- A UAlbany study of 3,055 U.S. participants found that political bias, not just awareness of AI use, shapes voter trust in campaign messaging.
- Marketing agencies are adopting AI agents but have not updated team structures, prompting calls for staff retraining and revised junior hiring approaches.
- Asian bishops and scholars meeting in Macau called for AI to serve human dignity, warning against deepfakes, surveillance, and AI-generated homilies presented as one's own.
- Stanford researchers used Evo 1 and Evo 2 AI models to design nearly 300 viral genomes, with 16 becoming living viruses that attacked lab bacteria, showing AI is learning natural design rules.
- Japan is preparing an emergency cybersecurity package and reviewing its AI Promotion Act following a surge in cyberattacks against companies.
- Canada's planned $1.6 billion investment in Cohere through a $7 billion funding round faces criticism, with opponents favoring better tax policies over state-led investment.
- A free open-source tool to detect AI-generated evidence in courts, led by Canadian researcher Maura Grossman, struggles with real-world reliability despite two years of development.
- Federated learning for deepfake detection can reduce explainability when data heterogeneity across clients is extreme, making AI explanations misleading in those scenarios.
- Fast Company warns that human oversight of AI is harder to guarantee than assumed, calling for stronger identity verification to ensure authorized humans are actually monitoring systems.
- Over 100 digital missionaries from 17 Asian countries need spiritual guidance and protection from online risks, according to the Macau assembly.
Study shows party bias shapes voter views on campaign AI
A study by Seulki Lee-Geiller at UAlbany surveyed 3,055 U.S. participants in mid-2024 to see how political affiliation affects views on AI in campaigns. Participants read identical campaign letters but were told different parties used AI. People trusted the results more when their own party was involved and doubted fairness when the opposing party used AI. The study highlights that political bias matters more than just knowing AI was used.
Ad agencies face AI staffing challenges and training shifts
Marketing agencies are adopting AI agents but have not updated their team structures or job roles to match. Gila Wilensky of Monks and Nadja Bellan-White of M&C Saatchi spoke at Digiday's Media Buying Summit about the need for staff training. Agencies are starting to rethink hiring for junior roles and limiting tool access to manage costs. The goal is to keep AI from making work tedious while capturing its productivity gains.
Experts discuss AI use in higher education classrooms
The Gazette featured expert opinions on how artificial intelligence should be used in higher education. Vanessa Miller covers higher education and reported on the discussion. The article points to ongoing conversations about best practices for AI in classrooms. Audio versions of the piece are available through Instaread.
Human oversight of AI is not as safe as it seems
An article in Fast Company warns that human oversight of AI systems is harder to guarantee than people think. In a human-in-the-loop model, a person must approve actions before AI proceeds. In a human-on-the-loop model, AI runs while a person supervises. Both assume a real person is present, but AI can now imitate voices, faces, and behavior. The article calls for stronger identity verification to ensure authorized humans are actually monitoring AI.
Asian Church leaders urge ethical AI and support for digital missionaries
Asian bishops and scholars met in Macau from October 27 to 31 and called for AI to serve human dignity. Bishop Marcelino Antonio Maralit Jr. warned against deepfakes, surveillance, and AI-generated homilies presented as one's own. Dr. Levi Checketts highlighted environmental and labor costs of technology. The assembly also discussed supporting over 100 digital missionaries from 17 Asian countries who need spiritual guidance and protection from online risks.
AI writes virus genome but humans still guide biosecurity
Stanford researchers used AI models Evo 1 and Evo 2 to design nearly 300 viral genomes. Sixteen of those designs became living viruses that attacked harmless lab bacteria. The viruses targeted ΦX174, a simple virus that only infects gut bacteria. Half the working viruses had mutations that natural evolution likely caused. The team says this proves AI is learning design rules from nature but humans still control safety and biosecurity.
Japan cyberattacks spark debate on AI rules and data security
A surge in cyberattacks against Japanese companies is pushing the government to consider stronger safeguards. The government is preparing an emergency cybersecurity package and reviewing Japan's AI Promotion Act. Privacy experts are also calling for tougher data-security rules and better compensation mechanisms. The attacks have exposed companies to compensation costs and damages claims. The debate centers on preventing misuse of AI and protecting personal data.
Critics question Ottawa's $1.6 billion Cohere AI investment
The Canadian government plans to invest up to $1.6 billion in Cohere AI through a planned $7 billion funding round. The article argues that governments should not act as investors in private companies. The author says the focus should be on better tax policies instead of state-led investment. The article warns that government funding can lead to dependency and hurt real innovation. Cohere has already raised billions from private investors.
Tool to detect AI evidence in courts is proving hard to build
Canadian researcher Maura Grossman is leading a team building a free open-source tool to help courts spot AI-generated evidence. The system uses multiple algorithms that vote together on whether content is AI-generated. After two years the team has a working model but it struggles in real-world use. Grossman tested it twice in real situations and the algorithms could not reach enough agreement to be helpful. The tool works better in lab tests than in actual court settings.
Federated learning for deepfake detection may hurt explainability
Researchers studied whether federated learning reduces the trustworthiness of deepfake detection explanations. They tested an EfficientNet-B4 detector on FaceForensics++ across five federated clients. The study found that explanation quality depends on data heterogeneity across clients. At near-identical data setups federated models matched or beat centralized ones. But at extreme data differences the AI explanations became misleading and unreliable. The research offers guidance for when federated deepfake detection can be safely used.
Sources
- Study: Partisan ties shape how voters view campaign AI
- Amid growing tension over workplace AI, agencies have a window to reimagine staffing
- What They’re Thinking: Experts on AI in higher education
- We can't always trust the “human” in human oversight of AI
- Asian Church urged to humanize AI and support digital missionaries
- Why human scientists, not AI, still control the fate of biosecurity
- Japan's cyberattack surge fuels debate over tighter AI regulation, data-security rules | MLex | Specialist news and analysis on legal risk and regulation
- Why Ottawa’s Cohere AI investment is a major red flag
- Building tool to help courts spot AI-generated evidence is 'challenging': researcher
- Does Federated Training Degrade Explainability? A Comparative Analysis of Grad-CAM Faithfulness in Centralized vs. Federated Deepfake Detection
Comments
Please log in to post a comment.