Center for AI Policy

AI-Assisted Cheating Appears Widespread Among Students, Educators Struggle to Respond A recent article from Matt Barnum and Deepa Seetharaman for The Wall Street Journal investigated how AI tools are increasingly used for academic dishonesty in schools. “This is a gigantic public experiment that no one has asked for,” Marc Watkins, assistant director of academic innovation at the University of Mississippi, told the reporters. The Journal interviewed a 17-year-old New Jersey student who used ChatGPT and Google’s Gemini for dozens of assignments because “work was boring or difficult,” she “wanted a better grade,” or had “procrastinated and ran out of time.” She only got caught once. AI companies largely deflect responsibility. “OpenAI did not invent cheating,” said Siya Raj Purohit from OpenAI’s education team. “People who want to cheat will find a way.” AI-powered detection efforts face significant challenges. In a Journal experiment, a detection tool from Pangram Labs correctly identified ChatGPT-generated writing as AI-created. But after processing through “humanizing” software, the same piece later passed as “fully human-written.” Pangram Labs’ CEO Max Spero said they’re working to “defeat the humanizers.” AI can complete even university-level assignments. In one experiment, researchers secretly submitted AI-written exam answers at a UK university and found that 94% went completely undetected. Furthermore, the AI submissions received grades that were, on average, several percentage points higher than submissions from real students. Pictured: The front page of Humanize AI.

Roblox Launches Open-Source 3D AI Generator Roblox, a popular online platform where users create and play games, has unveiled “Cube 3D,” an AI system that converts text descriptions into 3D digital objects. The company plans to open source a version of the AI model, making it freely available to developers both on and off their platform. Unlike some AI systems that generate 3D models by constructing them from 2D images, Cube 3D generates objects directly from text prompts. A developer typing “/generate a motorcycle” or “/generate orange safety cone” receives a corresponding digital object within seconds. The system works through what Roblox calls “3D tokenization,” breaking down 3D shapes into individual datapoints similar to how language models process text. This allows the AI to “predict the next shape token to build a complete 3D object.” Beyond individual objects, Roblox hopes to expand the technology to “enable creators to generate entire scenes based on multimodal inputs” including text, images, and other media types. Many people will use these new tools. During the fourth quarter of 2024, over 85 million daily active users spent an average of 2.4 hours per day on the Roblox platform. Virtual worlds are becoming as easy to create as they are to imagine. Pictured: Screenshot from a promotional video Roblox released for Cube 3D.

?? What's in store for U.S. AI policy? We've analyzed key recommendations from over 20 publicly available stakeholder responses to the White House AI Action Plan RFI. From emergency response planning to export control reforms, organizations across the political spectrum are advocating for ideas that could shape the future of AI governance in America. Read our latest article for a concise list of notable proposals from the Machine Intelligence Research Institute, U.S. Chamber of Commerce, Center for Democracy & Technology, OpenAI and many others—including our own recommendation to gather cyber incident data by designating frontier AI as critical infrastructure.

On Tuesday, March 25th at 12pm, CAIP will host a panel discussion on AI and Cybersecurity: Offense, Defense, and Congressional Priorities. Our distinguished panel brings together leading experts from across industry, academia, and policy research: -Fred Heiding, Postdoctoral Researcher, Harvard University -Daniel Kroese, Vice President of Public Policy & Government Affairs, Palo Alto Networks -Krystal Jackson, Non-Resident Research Fellow, Center for Long-Term Cybersecurity -Kyle Crichton, Cyber AI Research Fellow, Center for Security and Emerging Technology (CSET) The session will include a demonstration of an automated spear phishing AI agent, followed by discussion of current cybersecurity challenges, AI's evolving impact, and policy recommendations for Congress. RSVP here: https://lnkd.in/eKeciDqN

Last week, the Center for AI Policy (CAIP) shared suggestions with the U.S. Office of Management and Budget (OMB) regarding the implementation of President Trump's Executive Order 14179, Removing Barriers to American Leadership in Artificial Intelligence. CAIP's letter urges OMB to maintain and improve requirements for testing AI models as part of the government procurement process included in Memorandum M-24-18. It emphasizes the importance of testing for five forms of risk: hallucinations (the frequency of erroneous outputs), control (adherence to human commands), model security (resilience against external attacks), dangerous capabilities (capacity for destructive action), and privacy (protection of classified data). Additionally, CAIP recommends that AISI serve as the central hub for coordinating model testing across government agencies.

AI Policy Weekly No. 67: ??? OSTP received over 8,000 responses to the AI Action Plan RFI, with proposals ranging from developing emergency response protocols for AI crises to coordinating economic partnerships with African nations. ?? Roblox released 'Cube 3D', an open-source AI that generates 3D objects directly from text prompts using '3D tokenization', with plans to expand to full scene generation. ?? AI-assisted student cheating is widespread, with detection tools struggling against 'humanizing' software while AI companies maintain they 'did not invent cheating.' Quote of the Week: Bridgewater Associates CEO Nir Bar Dea offers advice on how to prepare for AI. Full stories at the link below.

+ The progression towards more autonomous and capable systems, underscored by these findings, emphasizes the urgent need for proactive governance and robust safeguards to prevent misuse and mitigate these escalating risks.

On Tuesday, March 25th at 12pm, CAIP will host a panel discussion on AI and Cybersecurity: Offense, Defense, and Congressional Priorities. Our distinguished panel brings together leading experts from across industry, academia, and policy research: -Fred Heiding, Postdoctoral Researcher, Harvard University -Daniel Kroese, Vice President of Public Policy & Government Affairs, Palo Alto Networks -Krystal Jackson, Non-Resident Research Fellow, Center for Long-Term Cybersecurity -Kyle Crichton, Cyber AI Research Fellow, Center for Security and Emerging Technology (CSET) The session will include a demonstration of an automated spear phishing AI agent, followed by discussion of current cybersecurity challenges, AI's evolving impact, and policy recommendations for Congress. RSVP here: https://lnkd.in/eKeciDqN

+ The Center for AI Policy (CAIP) commends NIST for its continued leadership in establishing structured and pragmatic guidance to manage the misuse risks associated with dual-use foundation models.?

CAIP in the News GovTech: Govt., Industry Respond to Federal Info Request on AI Plan + A Request for Information in February on the federal “Development of an Artificial Intelligence Action Plan” has garnered responses from a variety of industry and public-sector stakeholders offering recommendations. + The Center for AI Policy, a nonpartisan research organization, stated in comments that AI is currently “fundamentally insecure and unreliable,” urging the administration to introduce third-party national security audits for advanced AI. https://lnkd.in/ejAdUBuz