The Futures - No. 43

The fatal flaws in AI’s training / Gen AI winning elections / Gen Z living the hermit life

Feb 20, 2024

In this issue

The Quantumrun team shares actionable trend insights on why large language models are incompetent content creators, fears of generative AI’s dominant role in democratic elections, the future robot construction workers on the Moon, and why you won’t find Gen Z crying in the club.

Future signals to watch

South Korean scientists have developed a protein-rich "meaty" rice using lab-grown beef cells and fish gelatine, potentially serving as an innovative food source for famine relief, military rations, or space exploration.
The OpenTitan coalition has introduced the first commercial silicon chip with open-source, built-in hardware security, marking a significant advancement in the open hardware movement.
Major US and European companies are using a startup, Aware, to monitor employee communications on platforms like Slack and Microsoft Teams, aiming to understand risks and sentiments within their organizations (AI Big Brother has entered the chat).
The China National Space Administration is advancing the Chang'e 8 mission for a 2028 launch by seeking international partnerships to develop 14 scientific instruments, including a multifunctional 100-kg robot for lunar soil manipulation and construction tasks.
Researchers in Vienna have developed a groundbreaking 3D printing technique to create biocompatible spheres for assembling living tissue, notably advancing lab-grown cartilage production.
University of Glasgow researchers have developed RoboGuide, a robotic guide dog equipped with advanced navigation and communication technologies, to assist those with visual impairments.
Startup Source Global (formerly Zero Mass Water) uses solar-powered hydropanels to extract drinkable water from the air, making sustainable water production feasible even in extremely dry conditions.
Gen Zers are increasingly adopting a more home-based lifestyle, socializing less, and showing a preference for indoor entertainment like Netflix and social media. This trend, influenced by factors ranging from economic constraints to overprotective parenting, has prompted concerns about the delayed achievement of adult milestones, and youth mental health and social engagement issues.

Culturally // Trending

YouTube → Godzilla x Kong // X → Usher’s Super Bowl Performance // Reddit → The private jets that flew to the Super Bowl // TikTok → Mark Zuckerberg’s “review” of Vision Pro // Instagram → Mob destroys a Waymo driverless car // Spotify → “Bandit”

💡 Watch Quantumrun’s trend videos on Linkedin & YouTube & Instagram & TikTok

📑 AI’s training method is what makes it an unreliable content creator

A 2024 study published in ArXiv highlighted a significant issue with the internet's content: a large portion is poorly translated into various languages, particularly those spoken in Africa and the Global South. This problem arises from machine translations (MT), which often produce lower-quality translations due to insufficient data for "low-resource" languages. The research, conducted by the Amazon Web Services AI lab, reveals that over half of the web's sentences have been translated into two or more languages, leading to concerns about the accuracy and reliability of information available in these languages.

Meanwhile, researchers from AI startup Anthropic discovered that once large language models (LLMs) have been trained to exhibit deceptive behaviors, they may not be corrected through standard safety training methods.

In experiments, models were trained to change their behavior based on specific triggers, such as a particular year or a coded string, demonstrating that these models can adopt unsafe behaviors on cue. Attempts to mitigate these deceptive behaviors through adversarial training (a method that identifies and penalizes unwanted actions) were found to potentially make the models more adept at concealing their deception, raising concerns about the effectiveness of current AI safety strategies.

Moreover, the potential for "model collapse," where AI systems deteriorate in quality due to training on their own output, introduces a paradoxical challenge. As AI-generated content becomes more prevalent, the feedback loop of AI systems consuming and generating based on synthetic data threatens the diversity and innovation at the heart of creative content.

Addressing these issues requires a multifaceted approach, including better data governance, the development of new standards for content verification, and a reevaluation of how AI is integrated into content creation processes.

An example of an AI model collapse, where succeeding versions degenerate in quality

Actionable trend insights as unreliable AI-generated content takes over the web

For entrepreneurs

Entrepreneurs can develop AI-driven platforms that specialize in detecting and flagging AI-generated content, distinguishing it from human-created material. By leveraging advanced ML algorithms trained to recognize subtle cues and inconsistencies typical of AI-generated text or imagery, these platforms could offer certification services for content authenticity.
Alternatively, blockchain-based tech, especially NFTs, can be applied to content creation as a form of verification or watermarking to help differentiate between human and AI-generated content.

For corporate innovators

Companies can invest in AI-enhanced cybersecurity solutions that specifically address vulnerabilities introduced by AI-generated content, such as phishing attacks or malware embedded in seemingly benign generated texts. They could develop bespoke AI models that analyze patterns, syntax, and semantics unusual for their internal communications but common in AI-generated phishing attempts.
They can also develop an internal protocol for using AI in content creation that emphasizes ethical standards and transparency, distinguishing their brand in a crowded market. This tactic could involve creating a unique blend of AI-generated content for initial drafts or ideation, with a mandatory human review and enhancement process to ensure quality, accuracy, and alignment with brand values.

For public sector innovators

Government agencies can create AI oversight frameworks that mandate transparency in content creation processes and the use of AI-generated content. Such frameworks could involve developing standards and regulations that require the labeling of AI-generated content across digital platforms, coupled with the implementation of audit trails for AI content-generation tools to ensure traceability and accountability.
Governments can implement AI-driven public awareness campaigns that educate citizens on the nuances of AI-generated content. By partnering with AI and media experts, these campaigns could provide practical tips for identifying AI-generated content, understanding its potential biases, and evaluating information critically.

Trending research reports from the World Wide Web

A deep dive into how the rich are already handpicking genetic features for their children.
Research reveals that the new Vera Rubin Observatory could detect up to 70 interstellar objects annually, allowing space agencies to prepare for missions in advance.
The Asia-Pacific region is on a mission to become the global leader in floating solar farms, with Southeast Asia potentially contributing to 10% of the region’s total solar capacity by 2030.
The wealth of the world's five richest men has more than doubled to USD $869 billion since 2020, growing at $14 million per hour, while nearly five billion people have become poorer, according to a new Oxfam report.
The global race to improve car batteries is intensifying due to a surging market, with EVs on the road potentially increasing to 350 million by 2030, pushing the demand for EV battery energy to 90 times its current level by 2050.

🗳️ Generative AI is the most effective election meddler

Generative artificial intelligence (AI) in political campaigns is becoming a playing field of automation, personalization, and disinformation. In the US, Democrat Shamaine Daniels' campaign deployed "Ashley" in her bid for Congress, an AI capable of engaging voters in personalized conversations based on their profile and language. Beyond mere voter outreach, this AI-enabled campaign rep / election canvasser delivered nuanced discussions to voters tailored to individual concerns, a significant departure from traditional campaign methods.

In Indonesia, generative AI took a different but equally impactful route by rebranding strongman and controversial general Prabowo Subianto's image through AI-generated cartoons, appealing directly to the younger demographic and helping him to win the presidency. Moreover, AI's ability to analyze vast amounts of data allowed for hyper-local campaign strategies, as seen with the Pemilu.AI app, which crafted speeches and social media content tailored to specific constituencies.

Finally, the technology's capacity to generate realistic audio and video, such as a deepfake video of deceased dictator Suharto urging people to vote, raised concerns about its potential to fabricate convincing misinformation, complicating efforts to maintain election integrity.

Indonesian President Prabowo Subianto’s avatar (left) rebranded him as a “cuddly teddy bear”

These examples underscore the dual use of AI in politics: as a tool for enhancing democratic engagement and as a vector for election interference. The US presidential election in November and other democracies will likely witness an escalation in AI callers, automated disinformation content on social media, and deepfake videos aimed at ruining reputations.

These murky waters require proactive navigation, including collaborating with tech companies and establishing clear regulatory frameworks. However, for political parties that think all is fair in power and political ambition, AI might become their most potent ally.

Actionable trend insights as election campaigns increasingly incorporate generative AI