A new Cornell study reveals that AI-generated fiction has a strange habit of telling the same story. Researchers found that 70 to 90 percent of stories produced by models like ChatGPT, Claude, and Gemini feature a lighthouse keeper named Elias. The team analyzed roughly 100,000 AI-generated stories and identified this recurring theme as a case of mode collapse, where models gravitate toward familiar patterns despite those themes being rare in their training data. The findings will be presented at a conference in Budapest.
On the software development front, AI code review tools are creating a notable trust gap among programmers. A recent Stack Overflow survey shows that 51 percent of developers now use AI tools daily, and four major editors — Zed, Cursor, VS Code, and Claude Code — have introduced line-by-line review features for AI-written code. The goal is to catch errors before unreviewed AI output makes it into production, but developers clearly differ on how much they rely on these systems.
Cybersecurity concerns are mounting as personal AI agents enter the mainstream. Expert Hed Kovetz warns against trusting tools like Meta Muse and OpenAI Dots with sensitive data, arguing that companies released them without adequate security safeguards. Meanwhile, Walmart's AI agent Sparky reportedly increased shopping basket sizes by about 40 percent, highlighting how these tools can influence consumer behavior in ways users may not fully understand.
In healthcare, AI scribes are being deployed in hospitals to listen during patient visits and generate clinical notes. The practice raises serious data concerns, especially since many scribes are trained primarily on English and American accents, which can lead to errors for patients with different speech patterns. Hallucinations from the AI can end up in official medical records, prompting experts to call for better measurement, regular updates, and fairer outcomes for both patients and doctors.
Efforts to address AI bias in medicine continue as well. The Collaborative to End Race-Based Medicine and Wayne State University recently hosted a virtual summit focused on how AI can deepen existing racial bias in medical decision-making. The group, formed after high COVID-19 death rates in Black communities, aims to eliminate race-based medicine from teaching, research, and policy while incorporating community voices into the conversation.
On the education side, Jisc and Google announced a two-year partnership to help UK universities and colleges adopt AI more effectively. The program will integrate Google tools like Gemini and NotebookLM with Jisc's digital transformation expertise, offering workshops, webinars, and a new reseller framework for Google Workspace for Education. The initiative focuses on improving staff skills and governance around AI use in academic settings.
In hardware performance news, doubleAI launched its Code Execution Models, starting with Maxwell. The technology inspects and optimizes binary code at runtime on CPUs and GPUs without requiring source code changes. According to the company, Maxwell made data pipelines over 38 times faster and AI inference kernels over 4.5 times faster on real-world workloads, with every optimization mathematically proven to preserve program behavior.
Key Takeaways
- <ul><li>Cornell researchers found 70-90% of AI-generated fiction from ChatGPT, Claude, and Gemini repeats a story about a lighthouse keeper named Elias, a phenomenon linked to mode collapse.</li><li>A Stack Overflow survey shows 51% of developers use AI tools daily, while Zed, Cursor, VS Code, and Claude Code now offer line-by-line AI code review features.</li><li>Cybersecurity experts warn against trusting personal AI agents like Meta Muse and OpenAI Dots with sensitive data, citing rushed releases and weak security.</li><li>Walmart's AI agent Sparky reportedly increased shopping basket sizes by about 40 percent, raising concerns about AI's influence on consumer decisions.</li><li>AI scribes in hospitals risk introducing errors into medical records, particularly for patients with non-English or non-American accents, prompting calls for better oversight.</li><li>The Collaborative to End Race-Based Medicine held a summit with Wayne State University on how AI can worsen racial bias in medical decision-making.</li><li>Jisc and Google launched a two-year partnership to help UK universities adopt AI tools like Gemini and NotebookLM through workshops and a new reseller framework.</li><li>doubleAI's Maxwell Code Execution Model achieved over 38x speedups on data pipelines and 4.5x on AI inference kernels by optimizing binary code at runtime.</li><li>Experts recommend multiple verification methods, including reverse image searches and source checks, to detect AI-generated deepfakes in viral social media videos.</li><li>A Philadelphia woman was arrested for blocking a delivery robot in protest over job displacement, highlighting growing public friction with autonomous delivery systems.</li></ul>
AI fiction keeps repeating the lighthouse keeper Elias story
Cornell researchers found that AI-generated fiction often tells the same story about a lighthouse keeper named Elias. About 70 to 90 percent of AI-generated stories include words like lighthouse, keeper, Elias, Mara or Elara. The study analyzed 100,000 stories from models like ChatGPT and Claude. Researchers believe this repetition comes from shared training data and a problem called mode collapse.
Why AI stories keep featuring Elias the lighthouse keeper
A Cornell study found that large language models repeat the same fictional themes, especially stories about a lighthouse keeper named Elias. Researchers looked at 20,000 AI-generated stories across ChatGPT, Claude, Gemini and OLMo. The stories often use familiar names, jobs and settings even though these themes rarely appear in training data. The team suspects mode collapse and will present the work at a conference in Budapest.
AI code review tools create a 46% developer trust gap
Four popular code editors, Zed, Cursor, VS Code and Claude Code, now let developers review AI-written code line by line. A Stack Overflow survey found that 51 percent of developers use AI tools daily, making review more important. Each editor takes a different approach to showing and rejecting AI changes. The new tools aim to reduce errors from unreviewed AI-generated code.
Experts warn about trusting personal AI agents with data
Cybersecurity expert Hed Kovetz says he does not trust personal AI agents like Meta Muse or OpenAI Dots with sensitive information. He believes companies released these tools too quickly without strong security. Consumer protection expert Lindsay Owens pointed out that Walmart agent Sparky increased shopping basket sizes by about 40 percent. Experts advise users to limit what personal data they share with AI agents.
Woman arrested for blocking a delivery robot in Philadelphia
A woman in Philadelphia was arrested for standing in front of a delivery robot and blocking its path. Police charged her with disorderly conduct after she refused to move. She said she was protesting because delivery robots take jobs from human workers. Philadelphia is also the city where the robot HitchBot was vandalized and destroyed during a previous incident.
AI tests the credibility of viral videos
Social media videos are no longer reliable proof because AI can create realistic deepfakes that alter faces, voices, and events. Detecting fake videos is now harder without technical tools. Users should check the original source, date, and location of a video before sharing it. Reverse image searches and looking for mismatches in lip movements or lighting can help. Experts recommend combining multiple verification methods and thinking carefully before sharing any video online.
AI scribes in health care raise data concerns
AI tools called scribes are being used in hospitals to listen during patient visits and write clinical notes. Experts say this changes how medical data is collected and may introduce errors. The scribes are often trained on English and American accents, so they may work poorly for patients with different accents. Mistakes and hallucinations from the AI can end up in official medical records. Experts call for more measurement, regular updates, and fair benefits for patients as well as doctors.
Virtual summit tackles AI and racial bias in health care
The Collaborative to End Race-Based Medicine and Wayne State University hosted a virtual summit on AI and racial bias in health care. Experts say AI can worsen existing bias in medical decision-making and lead to poor health outcomes. The group wants to stop race-based medicine in teaching, research, and policy. They also aim to include community voices and create new education strategies. The collaborative was formed after high COVID-19 death rates in Black communities.
Experts warn AI makes cyberattacks harder to spot
Texas is among the top states for cybercrime losses, and experts say AI is making attacks more sophisticated and harder to detect. Common attacks include identity theft, voice cloning, and phishing emails. Experts recommend watching for urgent messages, grammar errors, and suspicious links. Businesses using AI systems from large companies may face higher risks if outdated coding patterns are not updated. AI can also help in medicine and cybersecurity, but the future impact remains uncertain.
Jisc and Google launch two-year AI partnership
Jisc and Google announced a two-year partnership to help UK universities and colleges adopt AI more widely. The program will combine Google AI tools like Gemini and NotebookLM with Jisc's digital transformation expertise. It aims to improve staff skills, governance, and understanding of AI's impact in education. A new Jisc reseller framework for Google Workspace for Education will give schools a tailored way to buy tools. The partnership will include workshops, webinars, and knowledge-sharing activities.
doubleAI launches Code Execution Models to boost hardware performance
doubleAI introduced Code Execution Models (CEMs), a new technology that optimizes running programs directly. Their first CEM, Maxwell, inspects and improves binary code while it runs on CPUs or GPUs without changing the source code. Every optimization Maxwell makes is mathematically proven to keep the program's behavior unchanged. On real-world workloads, Maxwell made data pipelines over 38x faster and AI inference kernels over 4.5x faster. The technology aims to get more work out of every piece of hardware in a compute-constrained world.
Sources
- Who is Elias the lighthouse keeper? AI fiction tells the same old story
- Who is Elias the lighthouse keeper? AI fiction tells the same old story
- AI Code Review Tools: 4 Editors, 46% Trust Gap [2026]
- Should you trust personal AI agents with your money and data? Here's what experts say.
- Philadelphia Is Holding the Line on Robot Resistance as Woman Gets Arrested for Blocking a Delivery Bot
- How can fake videos be detected? Artificial intelligence puts the credibility of circulating videos to the test
- The AI is listening: Implications of machine generated clinical notes on health care
- Virtual summit addresses artificial intelligence and racial bias in health care
- Experts warn AI making cyberattacks harder to detect
- Jisc and Google launch two-year partnership to accelerate AI adoption
- doubleAI Unveils ‘Code Execution Models’ to Improve Hardware Utilization
Comments
Please log in to post a comment.