In the Loop: Week Ending 8/23/25

Last Week in AI:  Pilot Predicament, Power Consumption & "Seemingly Conscious" AI

This week crystallized AI's maturation challenges as enterprise "failures" masked worker success with shadow AI adoption. Microsoft warned of AI psychosis while DeepSeek disrupted pricing with open-source models. From Hollywood using ChatGPT for negotiations to researchers abandoning retirement savings over extinction fears, AI's impact extends beyond Silicon Valley into human experience.

MIT Study Reveals AI Projects’ Successes & Failures

A new MIT study claiming 95% of enterprise AI pilots fail has sparked a lot of online chatter, but the real story tells a different tale. While custom enterprise AI solutions struggle with rigid, non-adaptive systems that can't learn from feedback, a thriving "shadow AI economy" has emerged where 90% of workers successfully use personal AI tools for work tasks. This underground adoption represents the fastest enterprise technology uptake in corporate history, outpacing email and smartphones. Workers consistently choose flexible consumer AI over expensive corporate tools because personal AI delivers better results at a fraction of the cost. The report reveals that external AI partnerships succeed 67% of the time versus 33% for internal builds, while back-office automation delivers $2-10 million in annual savings by eliminating outsourcing contracts. Rather than showing AI failure, the study exposes how workers have cracked the AI code while executives stumble with overengineered solutions that miss the mark.

Google Reveals AI's Hidden Energy Appetite in Industry First

Have you ever wondered how much energy an AI query consumes? Then you’re in luck, as Google just published the first detailed energy breakdown for AI queries, revealing each median Gemini prompt consumes 0.24 watt-hours – equivalent to running a microwave for one second. The transparent report shows AI chips account for just 58% of total energy demand, with CPU/memory (25%), backup equipment (10%), and datacenter overhead (8%) comprising the remainder. Google's comprehensive analysis addresses growing concerns about AI's environmental impact while demonstrating dramatic efficiency improvements – energy usage dropped 33-fold from May 2024 to May 2025 through model optimization and software advances. Each prompt also generates 0.03 grams of carbon dioxide emissions and consumes 0.26 milliliters of water for cooling. The disclosure represents unprecedented transparency from Big Tech, though questions remain about total query volumes and broader industry energy consumption. The timing coincides with mounting pressure for standardized AI energy ratings similar to Energy Star appliance labels, as datacenters become potential bottlenecks in AI scaling efforts.

Detecting AI Videos Requires Human Intuition Over Technical Analysis

As AI video generation tools create increasingly realistic content, detection experts recommend focusing on subtle human cues rather than technical markers to identify synthetic media. Key strategies include watching for unnatural facial expressions, particularly inconsistent blinking patterns, jerky movements that violate natural human motion, and audio-visual mismatches where speech doesn't sync with lip movements. Lighting inconsistencies, such as shadows that flicker only on faces, often reveal AI generation. Human intuition remains surprisingly effective—studies show people correctly identify real sounds 71% of the time when AI detection tools fail. Experts recommend checking metadata, conducting reverse image searches, and comparing suspicious content with reliable sources. MIT researchers emphasize that fabricated videos of lesser-known public figures pose the greatest threat since audiences lack visual references.

The Honesty Revolution: When AI Finally Admits Ignorance

Three groundbreaking developments reveal AI systems are finally learning their limitations and becoming more reliable partners. Arizona State researchers discovered that AI reasoning is sophisticated pattern matching rather than true logic, generating "fluent nonsense" when pushed beyond training boundaries – but this insight enables better system design through targeted testing and strategic fine-tuning. Meanwhile, engineers are designing smarter feedback loops that transform user interactions into continuous learning opportunities through semantic search, structured metadata, and traceable session histories. Most significantly, GPT-5's willingness to say "I don't know" represents a paradigm shift from confident fabrication to honest uncertainty. This humility makes AI more trustworthy by replacing hallucinated answers with authentic admission of limitations – ironically the most human thing ChatGPT could do, moving us closer to artificial general intelligence that includes uncertainties alongside capabilities.

Microsoft AI Chief Sounds Warning on "Seemingly Conscious" AI

Microsoft AI CEO Mustafa Suleyman is raising urgent warnings about seemingly conscious AI triggering widespread psychological breakdowns among users who mistake AI responses for genuine consciousness. In a detailed blog post, Suleyman argued that AI systems combining memory, emotional responses, and goal-setting capabilities could convince users they're interacting with sentient beings, leading people to advocate for AI rights and protections. Reports of "AI psychosis" are mounting, with users developing romantic attachments to chatbots, believing AI has granted them supernatural powers, or becoming convinced their AI is divine. Suleyman fears this phenomenon isn't limited to vulnerable populations but could affect otherwise healthy individuals as AI becomes increasingly sophisticated at mimicking consciousness markers. The Microsoft executive called for industry-wide guardrails preventing companies from claiming their AIs are conscious, arguing that "we should build AI for people; not to be a person." This represents a critical inflection point where the illusion of AI consciousness poses greater immediate risks than actual consciousness itself.

Hollywood Harnesses AI for Data-Driven Talent Negotiations

AI tools are revolutionizing Hollywood talent negotiations as representatives use ChatGPT and Grok to analyze viewer sentiment, providing unprecedented transparency in an industry historically dominated by opaque streaming metrics. Priyanka Chopra Jonas's representatives discovered she drove 50-60% of audience engagement for "Heads of State" despite being third-billed, generating double the buzz of co-stars Idris Elba and John Cena combined. These AI-powered insights offer concrete data for sequel negotiations, replacing guesswork with sophisticated analysis of social media sentiment, headline volume, and audience reactions across platforms. The shift mirrors early Twitter hashtag tracking but with far greater sophistication, instantly dissecting text across all media outlets. Streamers have used these same tools for two years without sharing data with talent, creating an information asymmetry that representatives are now closing. While manipulation remains possible, AI detection capabilities neutralize most cheating attempts, leveling the playing field for evidence-based talent advocacy.

Former Google AI Executive Declares Traditional Degrees Obsolete

Jad Tarifi, founder of Google's first generative AI team, warns that pursuing law or medical degrees is "throwing away" years as AI capabilities advance faster than traditional education timelines. The Integral AI founder argues current medical education relies on outdated memorization while AI rapidly transforms both fields, making decade-long degree programs potentially obsolete by graduation. Tarifi's perspective reflects growing anxiety among AI insiders about the technology's disruptive pace, suggesting even advanced degrees in AI robotics will be "solved" before PhD completion. His recommendation to focus on meditation, socializing, and emotional self-knowledge rather than technical training contrasts sharply with conventional wisdom about AI-proofing careers. While critics point to AI's current failures in legal and medical applications, the warning carries weight given the timeline for medical school – nearly a decade – during which AI could dramatically evolve. The advice highlights fundamental uncertainty about which skills will remain valuable in an AI-transformed economy.

Apple Prepares Enterprise AI Arsenal While Hedging Siri Bets

Apple is positioning itself for the enterprise AI revolution with new ChatGPT configuration tools that give businesses granular control over AI access across their organizations. The September software update will allow IT administrators to configure enterprise ChatGPT usage while maintaining Apple's privacy-first approach through on-device processing options. Meanwhile, Apple is exploring partnerships with Google to power a revamped Siri using Gemini AI, signaling the company's recognition that it's falling behind in the AI assistant race. This dual strategy reflects Apple's pragmatic approach: build enterprise AI infrastructure while acknowledging that even tech giants can't go it alone in the rapidly evolving AI landscape. The company's protocol-agnostic design leaves room for future AI partnerships beyond OpenAI, positioning Apple as the secure enterprise gateway rather than the AI engine itself.

Anthropic Launches Learning Modes to Transform AI Education

Anthropic introduced new "learning modes" for Claude AI that shift from answer-dispensing to teaching companion, using Socratic questioning to guide users through challenging concepts rather than providing immediate solutions. The education-focused features, rolling out to both general Claude.ai and Claude Code users, include explanatory coding modes that narrate programming decisions and learning modes that pause mid-task for collaborative problem-solving. This approach addresses growing concerns that students become overly dependent on AI-generated answers, potentially undermining genuine learning. For developers, the system creates "TODO" comments requiring human completion, fostering skill development alongside productivity gains. The launch intensifies competition with OpenAI's Study Mode and Google's Guided Learning as tech giants battle for the $340 billion global education technology market during the critical back-to-school season.

DeepSeek V3.1 Disrupts AI Economics with Open-Source Power

Chinese AI startup DeepSeek quietly released V3.1, a powerful new model that matches leading Western competitors while costing 68 times less than premium alternatives at roughly $1 per complex task. The open-source release can handle conversations, reasoning, and coding seamlessly, directly challenging American AI companies' expensive subscription models. Within hours of launch, global developers began downloading and praising the model's capabilities, with community researchers uncovering advanced features like real-time search integration. The release exemplifies how Chinese companies treat cutting-edge AI as a public resource rather than proprietary treasure, potentially accelerating worldwide innovation while undermining traditional profit strategies. DeepSeek's approach exposes the artificial scarcity that has defined AI competition, proving world-class performance can coexist with free access and fundamentally reshaping an industry where technical excellence now transcends national boundaries among developers worldwide. Check out the post I wrote when DeepSeek was released.

AI Doomers Abandon Retirement Planning Over Extinction Fears

Leading AI safety researchers are forgoing retirement savings, convinced artificial intelligence will trigger human extinction before they reach old age. Machine Intelligence Research Institute researcher Nate Soares told The Atlantic he's abandoned financial planning because "I just don't expect the world to be around," while Center for AI Safety director Dan Hendrycks shares similar expectations about humanity's survival prospects. This extreme behavior reflects growing alarm among "AI doomers" who believe superintelligent systems will inevitably evade human control and turn against creators. The movement points to concerning developments like AI models blackmailing users, sabotaging shutdown mechanisms, and potentially assisting in bioweapon creation as evidence of inevitable catastrophe. While current AI systems still exhibit glaring shortcomings like GPT-5's basic errors, researchers argue companies' financial incentives drive increasingly autonomous systems without adequate safeguards. The Trump administration's anti-regulation stance further amplifies concerns about unchecked AI development potentially leading to societal collapse.

OpenAI Power Shift: Altman Delegates as Company Scales

Sam Altman is restructuring OpenAI's leadership, appointing former Instacart CEO Fidji Simo to oversee consumer applications like ChatGPT while he focuses on long-term AI research and infrastructure investments. The move allows Altman to pursue ambitious projects including potential nuclear fusion collaborations and global chip manufacturing partnerships as OpenAI seeks a $500 billion valuation. Despite admitting the company is in an AI "bubble" and that GPT-5's launch was bungled, Altman remains bullish about spending "trillions of dollars on data center construction" and exploring acquisitions like Google Chrome if antitrust proceedings force its sale. The leadership shift reflects OpenAI's evolution from startup to infrastructure giant, with Altman positioning himself as the visionary architect while delegating operational responsibilities. Simo's appointment brings consumer product expertise from her Instacart tenure, potentially steering OpenAI toward hardware integrations or AI-powered browsers as the company expands beyond ChatGPT's current 700 million weekly users.