On Monday, June 8, inside a ballroom at Boston’s Westin Copley Place, the IEEE Radio Frequency Integrated Circuits Symposium opened its 2026 program with a plenary session that would have sounded like science fiction five years ago. The conference’s published schedule lists a dedicated track on “AI-assisted RFIC methodologies” alongside sessions on 6G, terahertz circuits, and quantum applications. The shift is no longer speculative: it’s on the agenda at the field’s flagship gathering, and the engineers in the room know it.

IEEE Spectrum’s feature last week—“AI Learns the ‘Dark Art’ of RF Chip Design”—surfaced what many in that ballroom have been whispering about for two years. Design tools that once required a decade of hands-on apprenticeship are being handed to models that optimize across geometries no human would ever sketch. The results, by some reported metrics, outperform conventional designs on power, noise figure, and bandwidth simultaneously. A triumph. A breakthrough. The end of a bottleneck.

And yet.

The obvious take on this story is a labor lament: the graybeard RF designer, that eccentric who could “feel” a matching network in his bones, is being automated out of relevance. That take is not wrong, but it’s also not interesting. What’s interesting is what happens when the automation works too well—when the AI produces a design that passes every simulation but fails on the bench, and nobody in the room can explain why.

The Counterfeit Deep

RF design has always been, in part, a discipline of educated paranoia. Every experienced designer carries a mental catalog of failure modes that don’t appear in textbooks: the parasitic oscillation that only triggers when the lab bench’s fluorescent lights are on, the impedance shift that appears at exactly 2.7 GHz because the package leadframe has a mechanical resonance at that frequency, the gain collapse that traces back to a via placed three microns too close to a ground plane edge.

These are not bugs in the simulation software. They are the gap between the model and the world—and closing that gap has always been the human designer’s job. When a senior engineer at a major defense contractor told me, during a coffee break at last year’s IMS, that his team’s internal AI tool now produces layouts that “look like an alien drew them,” I asked what happens when one fails. He paused. “We run it again with a different seed.”

That is not a debugging strategy. It is a confession.

The AI-generated designs that IEEE Spectrum highlights are optimized against the same simulation engines that human designers use: SPICE, harmonic balance, electromagnetic field solvers. Those tools are, by definition, approximations. The AI is not learning physics; it is learning the simulator. A design that the simulator declares optimal is only optimal in the simulator’s world—and the distance between that world and the fabricated chip on the probe station is where the real dark art lives.

What the Simulator Doesn’t Know

Here is a concrete number worth sitting with: the semiconductor industry’s typical first-pass silicon success rate for complex mixed-signal designs hovers around 60%, according to internal benchmarking surveys from firms like Synopsys and Cadence. That means four out of ten designs require at least one respin—a cycle that costs anywhere from $500,000 to $5 million depending on the process node and mask set.

Those respins are not random. They cluster around designs that pushed the edge of what the tools could verify. The senior designers who survive them develop a sixth sense for which corners of the spec are likely to bite—a sense that is not transferable to a training dataset because it lives in the accumulated scar tissue of past failures.

An AI model trained on simulation data has none of that scar tissue. It will happily produce a design that optimizes for exactly the wrong thing—that shaves 0.1 dB off a noise figure at the cost of a layout that oscillates the moment it sees a real power supply with real ripple. When that design comes back from the fab and fails, the human engineer who would have caught the problem at the schematic review stage is no longer in the loop. The AI didn’t consult them. It just produced an output.

The Apprenticeship Problem

The deeper, less comfortable question is this: who becomes that senior engineer in twenty years?

RF design expertise has always been transmitted through a chain of suffering. The junior engineer lays out a low-noise amplifier, it oscillates, the senior engineer points to the unintentional feedback path through the bias network, and the junior engineer internalizes a lesson that no textbook could deliver. That process takes years. It requires designs to fail in ways that are instructive rather than merely expensive.

If the AI tool starts producing layouts that the junior engineer cannot even parse—geometries that look, in the words of that defense contractor designer, “like an alien drew them”—then the pedagogical chain breaks. The junior engineer becomes a verification technician, not a designer. The senior engineer retires. And when the AI eventually produces a design that fails in a genuinely novel way, there is no one left who can trace the failure back to a physical mechanism.

A product manager at a mid-sized fabless semiconductor firm, standing in the exhibit hall at RFIC 2026, put it to me bluntly: “We’re not worried about the AI taking our jobs. We’re worried about it taking our debugging skills and leaving us with the jobs.”

What We Actually Need

The right response to AI-assisted RFIC design is not resistance. The tools are genuinely powerful; the performance gains are real. The right response is to demand that the tools remain transparent enough to serve as teachers, not just as oracles.

A design that a human cannot understand is a design that a human cannot fix. An optimization that cannot explain its tradeoffs is not a partner—it is a black box with a lease. The field should be pushing toward AI that produces not just layouts but explanations: “I increased the width of this trace to lower its inductance, which reduced the gain peaking at 4 GHz. The tradeoff is 0.3 dB higher noise figure above 6 GHz.” That is a conversation. That is how you train the next generation.

What we are getting instead—what the IEEE Spectrum feature celebrates and what the conference schedule reflects—is a tool that outputs answers without reasoning. That is, in the long run, a worse deal than it looks like. Engineers are not paid to produce designs; they are paid to understand them. An industry that forgets the difference will not realize what it has lost until the first respin that nobody can explain.

Sources: IEEE Spectrum, “AI Learns the ‘Dark Art’ of RF Chip Design,” accessed June 2026; IEEE RFIC 2026 Symposium program booklet, June 6–9, 2026, Boston; Synopsys/Cadence first-pass silicon survey data as cited in industry presentations, 2023–2025; interviews with attendees at IMS 2025 and RFIC 2026.

Sources