We are all AI philosophers now

Most technology decisions begin with capability. Can the system scale? Is it secure? Will it work with what we already use?Last month, Anthropic CEO Dario Amodei sat down with CBS News after the US government labeled his company a supply chain risk. The dispute focused on two uses Anthropic declined to support in its contract with the Pentagon: domestic mass surveillance and fully autonomous weapons without human control.Those cases made up only two percent of use cases. Yet they carried more weight in the company’s decision. Explaining the choice, Amodei said, “We believe that crossing those red lines is contrary to American values, and we wanted to stand up for American values.”That comment shifts the frame.When an AI provider draws a moral line, it sends a message. AI systems are shaped by their training data, tuning choices and safety rules. They reflect decisions about what is allowed. When organizations build on those systems, they accept those limits.This is no longer abstract. AI now affects identity checks, fraud alerts, automated tasks, customer interactions and reporting across the enterprise. As these tools move into core business work, their outputs shape real results.Technology leaders have always made tradeoffs. Encryption reflects risk tolerance. Access controls reflect trust. Data policies reflect compliance goals. AI simply makes those choices easier to see.The question for IT and security leaders is simple: When your systems act on AI output, whose values guide the outcome?As AI becomes part of core operations, that question becomes one of leadership.The illusion of neutral AISeveral years ago, I advised IT leaders in Washington State as they modernized their identity and access management systems. A major component involved evaluating vendors’ biometric capabilities. Accuracy and integration mattered. What required even greater scrutiny was bias.Our teams conducted extensive due diligence on how vendors trained and tuned their biometric models, how error rates varied across demographics and how those outcomes aligned with the state’s legal obligations and commitment to digital equity. Washington had already established a clear framework. HB 1493 (RCW 19.375) restricted commercial enrollment of biometric identifiers without notice and consent. And in April 2023, Governor Jay Inslee signed the My Health My Data Act into law, reinforcing privacy protections under the leadership of Chief Privacy Officer Katy Ruckle.There was no tolerance for biometric systems operating without oversight and making automated access decisions. Not because the technology lacked utility, but because its impact on citizens could disproportionately impact minorities, be difficult to explain or unwind.That experience makes something clear. AI is never neutral. Bias is embedded in training data, alignment tuning, safety constraints and access policies. Some providers go further and declare explicit moral baselines. For enterprise leaders, this carries a direct implication. Vendor choice is a governance choice. The architecture you approve encodes assumptions about fairness, accountability and acceptable risk. Those assumptions become operational reality the moment the system goes live.AI is artificial — and still stochasticGenerative AI and AI systems are built on probability. They produce results based on prediction, not certainty. That makes them useful for exploration, pattern finding and brainstorming. It is less reassuring when accuracy is mission critical and decisions affect citizens, customers or national security.Uncertainty is not a temporary flaw. It is part of how these systems work. Models can be tuned and guided, but variation remains. The risk is that clean dashboards and confident language hide that uncertainty. Leaders see polished results and assume precision.At the same time, regulators expect the opposite. Laws in Europe (E.g., EU AI Act) and several U.S. states are raising the bar for reliability, clarity and disclosure. Organizations are expected to explain how systems work and how confident they are in the results. High-stakes decisions require more than fast answers. They require traceable inputs and visible limits.At the Washington Digital Government Summit, state CIO Bill Kehoe put it simply: “AI innovation must be risk-averse and transparent.” He stressed strong data foundations, privacy by design and honoring opt-outs to maintain public trust.The tension is clear. We are handing serious decisions to systems that still operate on probability.From artificial to verified intelligenceAI generates plausible answers that sound correct. Verified Intelligence demands evidence. The difference matters most when decisions carry real impact.It makes little sense to separate intelligence from its source. Leaders need to know where conclusions come from, what data shaped them and whether they fit the business context. Context defines risk and consequence.Verified Digital Twins reflect a broader shift. Insight should require clear sources, defined limits and explicit confidence levels. As AI moves deeper into daily operations, the focus must shift from speed to clarity. Fast answers are not enough. Leaders need results they can explain and stand behind.IBM recently identified verifiable AI as one of its top AI trends for 2026. That reflects a growing expectation from regulators and boards that AI-driven decisions be explainable and defensible.Responsible AI conversations that lead to consequential decisions now hinge on four foundational pillars of Verified Intelligence:Grounding: Anchored to a real entity and decision contextScope: Explicit limits on authorityProvenance: Traceable reasoning and data lineageDrift awareness: Visibility into uncertainty and stalenessAI can generate insight. Verified Intelligence ensures leaders remain accountable for what follows.Disciplined AI deployment and leverageMost AI use cases bring small gains, or no gains at all. A small group delivers outsized impact. That same group often carries the most risk.For executive teams, the first discipline is categorization. Place AI use cases into one of three groups: speed enhancers, decision support and automated decisions. Speed enhancers improve efficiency but don’t always change outcomes. Decision support use cases guide how people act. Automated decisions trigger action on their own. The further you move toward automation, the more oversight you need.High-impact use cases usually sit close to revenue protection, fraud detection, uptime and customer trust. They also pose the greatest harm if bias, drift or weak data go unchecked. At this level, human review and clear escalation paths are essential.To avoid AI sprawl, build controls not only around models but around data. Knowing where data comes from and how it is used is critical. OASIS is advancing work on data provenance standards to strengthen traceability, alongside frameworks such as the NIST Cyber AI Profile released in December 2025.Framework alignment is table stakes. Clear Acceptable Use standards are leadership. Put in writing what AI may do, what it may not do and where human judgment is required. Build those standards into design reviews, vendor selection and procurement. If AI becomes core infrastructure, oversight must be built in as well.The executive questionControls and oversight matter. They are not the full story.For CIOs and IT leaders, this is about survival and results. AI now shapes revenue, customer experience, risk scoring, fraud alerts and daily operations. Decisions that once required human review now happen at machine speed. When those systems fail, the damage is real.Stopping at compliance is tempting. It offers a checklist and a sense of safety. Doing less can look like an optimization. Both create hidden risk. Laws set the floor. Markets set the penalty.The question leaders must face is simple: What does the business lose if we get this wrong?Revenue can slip through weak decisions. Trust can vanish after one public mistake. Regulators can shift from guidance to enforcement to penalties. Small system errors can grow fast when machines act at scale.AI is not just another wave of technology. It magnifies both strength and weakness. When guided by clear values and sound judgment, it strengthens the company. When poorly managed, it spreads risk faster than most leaders expect.Architectural implications for 2026In the same interview, Amodei added another pointed comment: “We are a private company... We can choose to sell or not sell whatever we want. There are other providers.”That remark goes beyond business strategy. It reminds us that philosophy comes first. Philosophy shapes policy. Policy shapes economics. And economics shapes the tools we use. By the time software is released, a worldview is already built into it.As you continue the AI conversation in 2026, the focus must shift from novelty to design thinking. The right response is not panic. It is clear thinking.That clarity shows up in how and what you build.It requires deliberate choices:Preserve optionality across providersLoosely couple AI componentsAbstract model dependencies behind controlled interfacesMaintain clear human accountability for consequential decisionsDesign systems that assume vendor positions, policies and boundaries can changeThis is not paranoia. It is disciplined execution in a rapidly shifting landscape.We are all AI philosophers now. Not because we wanted to be, but because the architecture we approve reflects the values we accept. And if it fails, responsibility will not belong to the model. It will belong to us.This article is published as part of the Foundry Expert Contributor Network.Want to join?