AI Responsibility Means Deciding Who Absorbs the Risk

When Anthropic disabled access to its Fable 5 and Mythos 5 models this week following a government directive citing national security concerns, the predictable camps formed immediately. Some saw validation for stronger oversight. Others saw government overreach into critical technology. But according to Demetri Giannikopoulos, a healthcare AI leader writing in Forbes, both sides missed the more important question: who absorbs the consequences when AI systems create friction or fail?

Why it matters

As organizations race to deploy AI across healthcare, customer service, and other high-stakes domains, the focus on whether AI can perform tasks obscures a more fundamental issue: where uncertainty should live in these systems, and who bears responsibility when things go wrong. The answer has direct implications for costs, trust, and the viability of AI economics as subsidies fade and inference costs become visible.

The Wrong Layer for Intelligence

Giannikopoulos argues that much of the AI deployment conversation fixates on replacement—will radiologists disappear, will programmers vanish, will customer service representatives become obsolete? He suggests this framing misses how AI might be most effectively used.

Consider clinical trial protocols, which change constantly with new amendments, inclusion criteria, and exclusions. The instinctive approach is to continuously feed every patient through AI to assess eligibility. Giannikopoulos proposes something different: use AI once when the protocol changes to interpret updates and translate complex language into structured rules. Have domain experts review that output. Then let conventional software execute those rules consistently until the next update arrives.

"If the protocol changes once a month, why would I invoke AI millions of times on millions of patients?" he writes. "Use the model ten times a year. Have somebody who actually understands the trial review the output. Then push the rules into software and move on."

One approach concentrates uncertainty where experts can inspect it once. The other spreads it across clinicians, patients, and support teams downstream.

The Economics of Inference

As AI moves from subsidized experimentation to production scale, usage costs become material. Unlike traditional software sold by seat or subscription, AI charges for compute and tokens. Inference—the process of generating responses—isn't free.

But not every interaction carries equal stakes. Password resets differ fundamentally from moments when customers are frustrated, frightened, or considering leaving. The cost of AI failure isn't measured by the first interaction but by the escalation, the second call, or the customer who quietly disappears.

This creates what Giannikopoulos calls "an interesting possibility": instead of replacing three people with nobody, one highly skilled person supported by AI might deliver similar economics while staying closer to the customer. "In some situations, the human may actually be the most cost-effective inference layer," he writes.

Who Owns the Problem?

Giannikopoulos illustrates the ownership challenge with a personal example: meaningful errors in his medical record that multiple people acknowledge but that remain uncorrected months later. "What I still haven't figured out is who actually owns the problem now that it exists," he notes. The issue feels less like a documentation error and more like an ownership vacuum.

In healthcare, clinical trials, and customer service, somebody eventually has to explain what happened. As organizations grow confident in AI technology, Giannikopoulos questions whether equivalent attention has been paid to deciding who absorbs consequences when systems don't work as planned.

The details were first reported by Demetri Giannikopoulos in Forbes.