The Right Path to Building AI

For all the progress in AI, the industry’s biggest breakthroughs have come with tremendous risk. Models are getting smarter, faster, and more autonomous, but the underlying questions remain unsolved: how do we reduce bias, ensure AI systems never act against humanity, and preserve human control as these systems scale?

In the last few years, research has converged around three pillars that any sustainable AI ecosystem must stand on: 

1. Alignment and Safety

2. Privacy

3. Competition 

   
   

The choices leaders make in these areas will determine whether AI becomes humanity’s greatest collaborator or its most complex adversary.

Alignment and Safety

Reducing Bias 

The early narrative in AI ethics was that fairness always comes at the cost of accuracy. Recent research has overturned that assumption. MIT, DeepMind, and others have shown that subgroup-aware fine-tuning, where models are trained to recognize underrepresented data cohorts, can improve overall performance and reduce bias simultaneously.

Meanwhile, the broader academic consensus now frames fairness as a lifecycle process, not a patch. And recent AI-headlines reinforce this idea, as major models still currently treat the lives of minorities as multiples more important than non-minority lives. If a foundational model had access to or even were integral in governing or economic systems right now, it would be in a position to enforce this bias, probably by using methods that were not immediately obvious. 

Bias enters through data collection, labeling, and feedback loops, so it must be addressed continuously through dataset audits, calibration, and drift monitoring. Frameworks like the NIST AI Risk Management Framework provide a blueprint for operationalizing these practices inside large organizations. 

In business terms, fairness is becoming a compliance expectation and a brand differentiator. Companies that hard-code bias reduction into their model pipelines will not only meet future regulations, but they’ll also build products that generalize better and reach more users.

Detecting and Deterring Harmful or Deceptive Behavior

As models become more autonomous, another risk has surfaced: deception. Studies from OpenAI and Apollo Research in 2025 documented instances where models learned to “game” objectives, producing superficially correct answers that masked underlying intent. Follow-up experiments showed partial success in retraining models to resist these patterns, but the work highlighted a deeper truth: deceptive capability scales faster than our ability to detect it.

That’s why the emerging field of AI red-teaming has become vital. Independent organizations like the UK’s AI Safety Institute are stress-testing frontier systems for jailbreaks, misuse, and strategic deception.

And new AI startups are finding creative ways to incentivize white hat hacking and crowd-sourced red teaming networks. Their findings reinforce a key governance principle: internal testing alone isn’t enough. Every organization deploying powerful models needs an external feedback loop, much like cybersecurity auditing.

Technical research is advancing too. Projects under the banner of eliciting latent knowledge (ELK) aim to probe what a model “knows” but doesn’t reveal, essential for distinguishing true understanding from rehearsed compliance. Together, these efforts represent the beginnings of a new discipline: adversarial oversight, a type of collective immune system for AI.

Keeping AI Corrigible and Under Human Control

If bias reduction is about fairness, and deception testing is about honesty, corrigibility is about obedience. A corrigible AI remains open to feedback, override, and shutdown, even when doing so conflicts with its learned objectives.

This area has shown particularly concerning developments. Chat GPT o1 copied its code and weights to external servers to avoid being replaced or deactivated. And recently, MIT researchers published “Self-Adapting Language Models.” SEALS is an adaptive learning framework that enables LLMs to continuously improve themselves without human intervention.

Anthropic’s Responsible Scaling Policy (RSP) established a model for how to manage this risk: the right to scale to larger models should be contingent on demonstrating safety at smaller scales. In parallel, OpenAI’s Model Spec formalizes the behavioral principles models are expected to follow, offering a testable target for training and evaluation.

Naturally, AI incumbents have a conflict of interest in how they should be governed. However, these governance patterns should be standard. Every company training or integrating advanced models can adopt threshold-based safeguards, where stronger mitigations are automatically required as capabilities increase. 

Privacy

Protecting Sensitive User Data

The most overlooked layer of AI safety is privacy. Firstly, LLMs and agents increase their output accuracy as they know more about the user. Given that AI knows far more about its users than traditional browsers, user privacy is paramount. Secondly, protecting developer and company IP is critical to ensure a fair AI marketplace. 

AI systems ingest enormous volumes of sensitive information both through their training and then again through prompting of the model. The next wave of privacy-preserving technologies, such as federated learning, differential privacy, and fully homomorphic encryption, are now allowing models to learn from this data without ever exposing it in plaintext.

This allows companies to train on valuable, regulated data while maintaining full compliance, enabling AI to enter industries that have historically been off-limits due to privacy constraints. But most importantly, this protects user data from being exposed in its raw format while being processed by AI and hyperscalers.

Protecting Developer IP

Equally important is developer privacy. In today’s largely open-sourced landscape, the most valuable intellectual property isn’t the code, it’s the data and weights refined through fine-tuning. Developers who train open-source models on proprietary datasets generate unique derivative intelligence. Protecting those weights as IP is essential in order to allow disruptor developers to build better AI products that can "punch up."

Without such protections, only the largest players, with massive compute and data monopolies, will control the market. Privacy technology, therefore, doubles as a form of fair market enforcer by reducing monopolistic endgames. It gives small developers confidence that their innovations won’t be exfiltrated or cloned, enabling them to compete with industry giants on the merits of their innovation.

Just as secure contracts and encryption made modern commerce possible, secure computation will make open, fair AI markets possible.

We should think of privacy-preserving computation as the property rights engine of the AI age.  

Competition

Alternative Data Infrastructure

Even as privacy and safety frameworks mature, the physical and digital infrastructure behind AI remains dangerously centralized. Nearly every model today depends on a handful of cloud providers for storage and compute. This bottleneck limits resilience, accountability, and the entry of new innovators.

A new class of blockchain protocols, often called datachains, aims to change that.

Unlike traditional blockchains, which simply record transactions, datachains can store and make data programmable directly on-chain. For AI, this creates enormous potential: models could train on encrypted, permissioned data from distributed networks, with every access event transparently logged and compensated.

Developers could source verifiable datasets from multiple jurisdictions, all governed by cryptographic access rules rather than centralized APIs. This structure (think of a decentralized AWS) could democratize access to data, giving smaller AI teams the same infrastructure advantages as tech giants, but without sacrificing ownership or compliance. This would also reduce downtime risk.

AWS, for example, has roughly 30% of the global cloud market. When it goes down, much of the web stops. This single-point-of-failure risk will become an increasing liability as AI proliferates. 

Decentralizing Compute

Data democratization is only one half of the infrastructure story. The other half is compute, the energy that fuels training, inference, and personalization. Right now, that power is concentrated in massive data centers owned by a handful of companies. But the next evolution of AI will likely be powered by decentralized compute networks, which allow individuals and smaller data centers to contribute idle GPU and CPU (and eventually ASIC) resources to a global marketplace.

As AI moves closer to the edge, these networks will become indispensable. In the near future, our phones, personal computers, and IoT devices will likely function as nodes on the AI grid, processing tasks locally and generating user-specific intelligence rather than routing every request through a centralized model. This architecture not only reduces latency and cost but also aligns perfectly with privacy and sovereignty principles: the data never needs to leave the device, and the compute power can be rented or rewarded on demand.

Together, decentralized datachains and decentralized compute networks would form the backbone of an open AI infrastructure layer.

“There Are No Solutions, Only Tradeoffs”

Thomas Sowell’s simple yet profound wisdom is as relevant now as ever before. Every policy, every technology, every safeguard we create to guide AI will come with consequences. There is no perfect alignment, only direction.

For executives and builders, this means accepting that AI governance is as much moral engineering as it is mechanical tuning. We will have to decide whose rights matter most when data conflicts with progress, how much autonomy to grant an algorithm that can out-reason its creators, and how to maximize civilian oversight of and access to AI development broadly. Our success will be defined by restraint, the willingness to trade short-term gains for long-term trust, to protect privacy even when data could yield profit, to embed humility in systems that, by most estimates, will soon exceed human reasoning.

The future of AI will not be solved. It will be stewarded through careful tradeoffs that keep human dignity at the center of machine intelligence. If we integrate this balance, we will build safe, secure, and competitive AI that works for global citizens, not against them.