Can AI Make Better Engineers?

Thesis: AI will make engineers faster, and that is precisely the danger. The scarce skill is becoming verification, and our education system is not producing it.

Artificial intelligence now writes code, generates register-transfer-level hardware descriptions, and optimizes chip floorplans better than experienced humans in some benchmarks. My first reaction was anxiety about whether engineering would still need people. After thinking about it further, my concern shifted. The realistic near-term outcome is not that AI replaces engineers but that it produces far more designs than anyone can properly check. Volume of output has never been the limiting factor in engineering quality; understanding has been. If AI multiplies output without multiplying understanding, it creates a new category of risk.

The core technology

Electronic design automation has used optimization algorithms for decades, but recent systems apply reinforcement learning and large language models to design tasks directly. Machine learning agents explore synthesis and place-and-route parameter spaces that a human team could never search exhaustively, discovering configurations with better power, performance, and area. Learned models have produced chip floorplans competitive with expert layouts. Language models generate hardware description code, testbenches, and documentation from natural language specifications. In parallel, surrogate models trained on simulation data approximate expensive electromagnetic or thermal analyses in milliseconds instead of hours. Verification is being automated too, with models proposing assertions and generating stimulus intended to reach states a human-written testbench would rarely visit.

Figure 1. A chip floorplan generated through electronic design automation. Placement and routing tools like this now compete with, and increasingly exceed, expert human layouts.

Figure 2. An AI-agent design loop: a code agent proposes RTL and a testbench, a compiler and simulator check it, and a review agent summarizes bugs before the cycle repeats — illustrating how generation and verification are being chained together.

Recent developments and real problems

The measurable gains are real: shorter design cycles, better power efficiency, and coverage of verification corner cases that humans overlook. But three problems concern me. First, generated designs are often unexplainable, so when something fails in silicon, the team may not understand the structure they shipped. Second, models trained on existing designs inherit existing assumptions and can be confidently wrong in unfamiliar regimes, which is exactly where engineering judgement matters. Third, verification effort does not scale automatically with generation speed, so teams face a growing backlog of unchecked output. There are also unresolved questions about liability and about training data provenance. If a generated circuit closely reproduces a design the model was trained on, neither the engineer nor the tool vendor can currently demonstrate otherwise.

Outlook and solutions

The constructive path is to pair generative tools with formal methods. Formal verification proves properties mathematically rather than sampling them by simulation, and it pairs naturally with AI that produces candidate designs quickly. Engineering education should shift emphasis accordingly: less time producing first drafts, more time on specification writing, assertion design, failure analysis, and knowing what a plausible-looking answer should not look like. Organizations should require that any AI-assisted design carry a human-authored specification of intent and a traceable verification argument. Universities could support this by grading students on how thoroughly they test a design rather than only on whether it works in the demonstration case.

Figure 3. Formal verification methods trade off automation and scalability against precision. Pairing fast, imprecise AI-generated designs with these rigorous methods is the constructive path this essay argues for.

Conclusion

My conclusion is that AI does not make better engineers automatically; it makes leverage, and leverage amplifies whatever judgement already exists. The skill I would most want to develop is the ability to be productively suspicious, to read a working result and ask why it works and under what conditions it stops. That capacity has always separated competent engineers from careful ones, and I think the next decade will make the difference expensive.

Sources

1. Google DeepMind (2024). How AlphaChip transformed computer chip design (on Mirhoseini et al., Nature 594, 2021). https://deepmind.google/blog/how-alphachip-transformed-computer-chip-design/

2. C.-K. Cheng et al. An Updated Assessment of Reinforcement Learning for Macro Placement — reports no published independent reproduction. arXiv:2302.11014. https://arxiv.org/pdf/2302.11014

3. AlphaChip (controversy) — summary of the replication and peer-review dispute. https://en.wikipedia.org/wiki/AlphaChip_(controversy)