Every breakthrough technology forces leaders to decide how to protect people without suffocating progress. With AI, that decision is overdue. The smartest path isn’t more binders or black-box spelunking—it’s clarity on outcomes and accountability when systems cause harm.
Here’s a simple way to see it. When a neighbor’s golden retriever snaps at a mail carrier, animal control doesn’t ask for a neuron-by-neuron brain map. They ask three questions: Did the dog bite? Was anyone hurt? Was it provoked? We should evaluate AI the same way—by the real-world results that affect people and businesses.
Jon’s take: If you can’t audit the brain, audit the bite. Outcomes are measurable. Opaque inner workings aren’t.
When AI makes a consequential decision—in hiring, lending, healthcare, or criminal justice—regulators and operators should start with a crisp, outcome-based checklist:
We already have laws that protect against discrimination and unfair practices. Enforce them—firmly. Keep humans responsible for outcomes, even when AI is in the loop. That’s more practical and more protective than forcing companies to reverse‑engineer their own black boxes.
Process-first regulation—like Colorado’s SB24-205 as debated—leans heavily on documentation about “how models work.” That’s like demanding a full nervous-system diagram every time a dog barks. You’ll get paperwork, not public safety.
Here’s the hard truth: modern models display emergent behavior—capabilities that arise from billions of interacting parameters in ways even creators can’t fully explain. It’s the AI equivalent of instincts.
Large language models (LLMs) such as GPT, Claude, DeepSeek, and Gemini learn from vast text corpora. No one explicitly programs “write poetry” or “translate between two niche languages.” Yet those skills appear once the system becomes large and well-trained enough.
That’s emergent behavior. It’s real, powerful, and inherently difficult to explain line-by-line.
You can’t predict every capability a frontier model will learn—or how it will generalize across tasks. And it’s not just about the training data.
Data ≠ destiny.
The result also depends on the hidden mechanics—embeddings, weights, and layers—that transform inputs into outputs. Two teams can train on the same dataset and end up with very different behavior: one fair, one biased. The only reliable way to assess that difference? Measure the outcomes.
When lawmakers try to regulate the invisible math inside, they invite regulatory theater: checklists, binders, and bills—without guarantees of safer systems. In Colorado alone, estimates peg administrative costs around $6 million under SB24-205’s approach, with little evidence those forms would stop a single harmful decision.
In other industries, process rules work because steps are visible and auditable. Inspectors can check thermometer readings in a kitchen, review sterilization logs in pharma, or verify torque settings in aviation. Those processes are stable and observable.
AI’s process lives inside advanced math. Lawmakers naturally focus on inputs (data). That matters—but what happens next is harder to see. Embeddings map meaning in high-dimensional space. Weights shift millions or billions of times during training. Deep networks run through dozens or hundreds of layers that resist plain-language explanation.
Think of training as school: the model rewires itself until it learns. Inference is the job: it applies what it learned to new questions. With retrieval-augmented generation (RAG), the weights remain fixed, yet outputs shift based on the external documents retrieved. This is why at Iterate, our Generate platform and Interplay Agentic AI are designed with transparent, document-linked reasoning—so enterprises can trace decisions back to sources rather than black-box guesses.
Engineers prune, distill, quantize, and fine‑tune models after training to make them faster, smaller, or specialized. Useful? Absolutely. Transparent? Not really.
Now add agentic AI, where systems retain memories and take multi-step actions. Industry leaders have suggested memory could be a defining feature of upcoming generations. That means the same model might behave differently tomorrow, not because its wiring changed, but because its experience did. Auditing yesterday’s process won’t guarantee tomorrow’s behavior.
Having served on Colorado’s AI Task Force, I’ve seen both sides up close. Lawmakers want oversight for systems that affect livelihoods. Companies, especially startups, point out that demanding “explainable neurons” is technically unrealistic and strategically hollow.
Here’s the practical risk: mega-cap companies can afford compliance bureaucracy. A three-person startup—often not drawing salaries—cannot. What looks “reasonable” to a trillion‑dollar balance sheet can smother the inventors who create jobs, patents, and local growth.
Research consistently shows young firms drive net job creation, and startup patents tend to be disproportionately influential. In Colorado, that innovative edge spans AI and our world-class quantum ecosystem. Overweight process rules may dull that edge.
Consider a resume-screening system.
Which one better protects applicants? Which one can you enforce this quarter—not next decade?
The best regulations in history focus on outcomes. We don’t micromanage how engines function; we set standards for emissions and crash safety. We don’t script every step of underwriting; we outlaw discriminatory lending.
Apply that logic to AI:
This is aligned with emerging AI governance thinking—including frameworks like Gartner’s AI TRiSM, which emphasize trust, risk, and security management across the AI lifecycle. See: Gartner on AI TRiSM.
Colorado’s debate is a microcosm of a global challenge: how to regulate AI in a way that protects people without stifling innovation. Process-heavy mandates miss the point—they demand explanations no one can give and create paperwork no one reads. Outcome-based accountability, by contrast, is practical, enforceable, and fair.
At Iterate.ai, we believe AI should be built and deployed with this principle at its core: measure what matters, govern what impacts people, and keep humans accountable. Whether through Generate for secure AI assistance, Interplay for agentic low-code innovation, or Extract for precision data handling, our goal is to help enterprises adopt AI that’s trustworthy, transparent, and outcome-first.
If policymakers, innovators, and enterprises align around outcomes—not opaque math—we can unlock AI’s potential responsibly, without dimming the spark of innovation that drives progress.