History has a way of repeating itself, and right now, it’s offering a stark warning.
When the commercial internet launched in 1995, the digital frontier was exhilarating and ungoverned. It took eleven years before enterprises were required to demonstrate meaningful cybersecurity postures through frameworks like PCI DSS and SOX compliance. Similarly, when Six Degrees became the first social media platform in 1997, nobody imagined we’d need laws protecting children from algorithmic manipulation. Yet more than twenty years passed before regulations like COPPA updates and GDPR began addressing the harms, and many of social media’s most damaging societal effects remain unaddressed today.
Alexander Schlager, founder and CEO of Aiceberg.ai, views this pattern through a lens of timing and responsibility honed by more than two decades in technology and cybersecurity leadership. “It’s interesting how deeper reflection often emerges only once a technology is widely used,” he says. “That dynamic feels relevant to AI because it encourages us to think carefully about the choices we make as the space evolves.”
Now, artificial intelligence is transforming every industry at unprecedented speed. Large language models draft legal briefs, AI agents execute financial transactions, and autonomous systems make decisions affecting millions. The question society must confront is uncomfortable but urgent: Can organizations afford another decade-long regulatory lag?
The Black Box Monitoring Black Box Problem
Organizations racing to deploy AI have stumbled into a dangerous trap. To secure their AI systems, many have deployed additional AI, using large language models to monitor other large language models. The logic seems elegant: fight fire with fire.
But this approach has a fatal flaw.
Large language models are, by their nature, black boxes. Their decision-making processes remain opaque even to their creators. When you monitor one black box with another, you haven’t achieved transparency; you’ve compounded the mystery. If your AI security system flags an interaction as dangerous, can it explain why? Can you audit that decision for regulators?
Schlager points to this as a core problem his company addresses. “Relying on one black box to monitor another creates a compounding opacity problem that undermines trust and may make regulatory alignment harder,” he explains. Under his leadership, Aiceberg focuses on AI safety and security through approaches that favor interpretability, explainability, and consistency.
Why Explainability Changes Everything
Here’s the twist few people saw coming: one of the most powerful tools for securing advanced AI isn’t cutting-edge at all. Traditional machine learning classifiers, descendants of Frank Rosenblatt’s 1957 perceptron, are deterministic: the same input always produces the same output. More importantly, they are explainable. Every decision can be traced to specific features, patterns, and training examples.
This transparency isn’t a technical nicety. It’s the foundation of trust, compliance, and effective governance.
When an explainable security system flags a prompt injection attempt or detects an AI jailbreak, analysts can see exactly which patterns triggered the alert. They can demonstrate to auditors precisely how decisions were made. They can verify that the system isn’t susceptible to the adversarial techniques threatening the AI it monitors.
“Progress often feels sustainable when teams can explain systems to regulators, customers, and themselves,” Schlager shares. “Visibility can turn complexity into something teams can engage with thoughtfully.”
Explainable systems also offer operational resilience. If an LLM provider experiences an outage, LLM-based security fails alongside it. Deterministic classifiers continue operating independently, maintaining protection regardless of external dependencies.
The Speed Imperative
AI threats evolve daily. New attack prompts, jailbreaking techniques, and manipulation strategies emerge constantly. Security systems must adapt at the same pace.
Here, traditional machine learning holds another advantage. Updating an LLM to recognize new threats requires expensive retraining that takes weeks. Purpose-built machine learning frameworks can incorporate new attack patterns within days, learning incrementally without disrupting existing capabilities.
Regulatory momentum adds urgency. The EU AI Act and emerging frameworks across regions increasingly emphasize transparency, risk management, and accountability. “I tend to view these developments as part of the same picture as our technical design choices,” Schlager says. Organizations integrating explainability early experience smoother alignment with evolving standards, since documentation and reasoning already exist within their systems.
The Choice Before Us
The internet and social media taught society that technology moves faster than governance. The consequences are already being felt today, including election interference, teen mental health crises, and privacy violations at scale.
With AI, the stakes are higher, and the timeline is compressed. Organizations cannot secure what they cannot see, and they cannot trust what they cannot explain. The enterprises that thrive in the AI era will be those that architect safety and security on transparency and independence, not those that stack black boxes atop black boxes and hope for the best.
The outcomes of waiting are already well documented. This time, the tools to act differently already exist. The only question that remains is whether leaders will choose to use them.
