World Solve

App permission systems exist, but many apps still collect more data than users realize or intend to allow, often through complex, hard-to-navigate settings. True granular control over personal data remains inconsistently implemented across platforms.

Tools designed to detect AI-written assignments frequently produce false positives, wrongly flagging legitimate student work as AI-generated. This unreliability creates real consequences for students while genuine cheating sometimes slips through undetected.

Fake reviews generated at scale by AI can be nearly indistinguishable from genuine customer feedback, undermining trust in online shopping decisions. Detection tools are struggling to keep pace with how convincing these fakes have become.

It's often unclear exactly what data went into training a given AI model, making it difficult to audit for bias, copyright issues, or harmful content. Greater transparency into training data remains technically and legally unresolved.

Advanced spyware can be installed on phones without any visible sign, and even security experts sometimes struggle to detect its presence through standard tools. Better, more accessible detection methods remain an urgent unmet need.

Automated moderation systems still struggle to tell the difference between genuinely harmful content and similar-looking legitimate content, like educational or satirical material. This tradeoff between over- and under-moderation remains unresolved at scale.

Researchers have shown that inaudible or disguised audio signals can sometimes trigger voice assistants to take unintended actions without the user's knowledge. Fully closing this security gap remains unresolved.

Many data centers rely on significant volumes of water to keep servers cool, straining local water supplies especially in drought-prone regions. Alternative cooling methods that don't compete with community water needs remain underdeveloped at scale.

AI health tools can sometimes provide advice that sounds plausible but is subtly wrong or dangerous, and there's no consistent, scalable system for catching these errors before they reach real patients. This gap grows more urgent as AI health tools become more widely used.

Smart home products from different brands frequently fail to communicate properly with each other, forcing consumers to juggle multiple apps and hubs. A universal, trusted communication standard remains incompletely adopted across the industry.

Fake or substandard electronic parts sometimes slip into supply chains for critical equipment, including medical devices and vehicles, and current detection methods are neither fast nor cheap enough to catch every case. This creates hidden safety risks across many industries.

Solar panels contain valuable and sometimes hazardous materials that current recycling processes struggle to recover efficiently and cheaply. As the first generation of panels ages out, this waste problem is becoming more urgent.

Once a photo or video is edited or generated, there's often no reliable way to trace it back to prove whether the original was authentic. Building this verification into cameras and platforms from the start remains an unsolved industry challenge.

The way solid-state drives store data internally makes secure, complete deletion technically harder than with traditional spinning hard drives. This creates lingering privacy and security risks when these drives are discarded or resold.

Today's most capable AI models are trained on vast datasets scraped from the internet, often including copyrighted works and personal information without clear consent. Finding effective training methods that avoid this legal and ethical gray area remains unsolved.

Most software vulnerabilities trace back to common coding mistakes that better language design could prevent from the start. Despite decades of research, mainstream programming still makes these mistakes easy to make.

Zero-day flaws are unknown weaknesses in software that attackers can exploit before a fix exists, and current detection methods still miss many of them until after damage is done. Finding these flaws proactively remains one of cybersecurity's hardest problems.

Algorithms trained on high-income populations may not generalise to under-sampled groups. Inclusive data-collection initiatives and equity-centric model evaluation are necessary.

People currently juggle dozens of separate logins and identity systems, each with different security and privacy standards. A single trusted, privacy-respecting identity system remains unbuilt at global scale.

Modern AI models produce answers without a clear, checkable chain of reasoning, making it hard to know if a correct answer was reached for the right reasons. Verifiable reasoning is essential before AI can be trusted in high-stakes decisions.