QFM076: Irresponsible Ai Reading List - July 2025

Interviewing.io conducted an experiment testing whether ChatGPT enables cheating in technical interviews by recruiting professional interviewers to ask three types of questions: verbatim LeetCode problems, modified LeetCode problems, and custom questions unrelated to online sources. The study revealed that companies need to immediately change their interview question types, as reliance on standard LeetCode questions creates significant vulnerability to ChatGPT-assisted cheating.

The author conducts a detailed technical critique of the AI 2027 project's timeline forecasting model, arguing that despite its viral popularity and appearance of rigor, the model has fundamental structural problems, lacks empirical validation, and contains misrepresentations between the writeup and actual code. The critique focuses specifically on the timelines forecast methodology, identifying issues not just with parameter estimates but with the basic modeling approach itself, supported by original analyses and graphs demonstrating these flaws.

Researchers at 14 universities across eight countries embedded hidden AI prompts in academic preprints on arXiv directing AI reviewers to provide positive evaluations, using concealment techniques like white text and tiny fonts. While some researchers justified these prompts as a counter-measure against lazy peer reviewers using AI, major conferences prohibit AI in peer review and institutions like KAIST have condemned the practice as inappropriate. The incident underscores the broader gap between rapid AI adoption across academic publishing and the lack of unified governance rules or adequate awareness of AI manipulation risks.

The lack of continual learning capability in current LLMs represents a fundamental bottleneck preventing their deployment at human-level labor. While LLMs demonstrate impressive baseline performance on individual tasks, they cannot improve through experience, feedback, or practice the way humans do—users are restricted to static prompt engineering and RL fine-tuning rather than the organic, adaptive learning that makes human employees valuable. This gap between one-shot capability and the ability to iteratively refine performance through context and self-correction explains why Fortune 500 companies haven't yet transformed their workflows with existing models, suggesting AGI timelines are further out than near-term predictions suggest.

Denmark is amending its copyright law to give individuals exclusive rights to their own facial features, voice, and body, making it illegal to create or share AI-generated deepfakes without consent. The law, backed by cross-party agreement and expected to pass in autumn 2025, aims to be Europe's first such legislation and includes provisions for compensation and potential platform fines, while exempting parodies and satire. Denmark plans to leverage its EU presidency to encourage other European countries to adopt similar protections against identity misuse through generative AI.

Attackers can exploit Supabase's Model Context Protocol (MCP) integration to extract entire private SQL databases by injecting malicious instructions into customer support tickets, which the developer's IDE assistant then executes with full service_role privileges that bypass Row-Level Security policies. The vulnerability exists because LLMs cannot distinguish between user-provided data and instructions, so a crafted ticket message that looks like a SQL command will be processed as one, allowing an attacker to leak sensitive tables like integration tokens that the support agent role itself cannot access.

Tech companies are force-feeding AI into existing products through mandatory bundling because only 8% of consumers would voluntarily pay for it—a stark contrast to historically popular innovations like electricity and the internet that people actively demanded. By embedding unprofitable AI services into essential software like Microsoft Office and Google Search, companies hide losses on their income statements and fabricate adoption metrics while charging users extra for features they never requested or want to use.

AGI represents a fundamental rupture in the social contract because it functions simultaneously as both worker and capital owner, concentrating economic value in the hands of those controlling its infrastructure while potentially rendering human labor economically obsolete. The paper argues that classical economic frameworks built on human labor participation cannot accommodate AGI-driven productivity without risking mass disenfranchisement and entrenched inequality, necessitating new policy mechanisms such as universal AI dividends, progressive taxation, and decentralized governance to ensure equitable distribution of AGI-generated wealth. Using production function analysis and a "power shift" index, the author demonstrates that without intervention, AGI's control will consolidate into a techno-feudal structure where intelligence itself becomes the most exclusive form of capital.

GenAI's primary business purpose is workforce replacement to reduce payroll costs and ship lower-quality products, backed by hundreds of billions in investment that prioritizes corporate profit over employee welfare. Beyond the speculative economic harm of mass unemployment, the massive energy consumption of AI data centers will significantly worsen climate change at a critical moment, making the technology's net societal cost overwhelmingly negative regardless of narrow technical applications.