#073 - AI in Engineering | Speed vs Substance

Are we prompting ourselves into obscurity? What's the appropriate amount of human oversight? How do we gauge efficacy?

There's a pervasive narrative accompanying AI's forward march into the engineering sector: unprecedented gains in productivity. While the potential for accelerating certain tasks is already clear, a more critical question emerges: are these increases in speed translating directly to an equivalent uplift in the quality of engineering outcomes?

We see different patterns of AI adoption. Some engineers employ these tools for rapid conceptualization, quickly generating preliminary frameworks or designs. Others integrate AI more incrementally, seeking to augment daily tasks like meeting minutes, summarizing documents, code checking, or initial drafting of emails/reports.

Both paths promise efficiency. Yet, there's a subtle, often uncounted, cost to this new velocity, one that requires disciplined engineering scrutiny.

Experience Determines AI Efficacy

It's becoming clear that the true value extracted from AI in its current state is heavily modulated by the experience of the user. A user with more engineering experience tends to approach AI-generated output as raw material – a starting point requiring interrogation, refinement, and validation against accumulated knowledge and experience. For them, AI can indeed accelerate the journey to a robust solution by handling grunt work, but expertise remains the guarantor of quality.

The inverse scenario presents a significant concern. An engineer earlier in their career, lacking that backlog of real-world problems and solutions, might be more inclined to accept an AI's output with insufficient critical distance. This is where the risk of "brittle" or superficially sound engineering arises – designs that may look plausible but haven't been tempered by the hard lessons of experience, potentially overlooking subtle failure modes, edge cases, constructability issues, or crucial code interpretations. This is normal and has been the case forever, as tech emerges.

There's a curious inversion: AI, ostensibly a knowledge tool, often delivers greatest value not to the novice seeking guidance, but to the expert who can critically direct and correct it. An AI, for now, acts much like a highly efficient but inexperienced apprentice; it needs constant, discerning oversight.

I’ve often likened it to a friend that tends to exaggerate their stories, you just need to be aware of the circumstances.

The Last 30%: Where Engineering Value is Generated

It is a common observation that AI tools can get a task to a substantial level of apparent completion, let’s say 70% of the way, with impressive speed. However, that final 30%, the portion that encompasses true engineering robustness, reliability, and the thorough addressing of real-world complexities, frequently proves to be a far steeper climb, demanding disproportionate human effort.

This "last 30%" is not a trivial component; it is the essence of professional engineering: ensuring things work as they are intended to, under all foreseeable conditions.

Furthermore, the risk of AI-driven changes introducing unintended negative consequences elsewhere in a system is very real. A seemingly clever optimization suggested by an AI in one area could, if not holistically reviewed, compromise serviceability, safety, or conflict with other critical design requirements.

A serious concern is that over-reliance on AI for complex problem-solving, especially without deep engagement with the underlying principles, could inadvertently hinder the development of core engineering intuition and diagnostic skills - this is something I am constantly thinking about.

If solutions are generated without genuine comprehension, we risk skill atrophy. To counteract this, certain practices are essential:

AI for Initial Drafts, Human Intellect for Critical Validation

• Employ AI for what it does well: generating initial concepts, performing routine calculations (via code - not text), or creating preliminary models.

• Mandate thorough manual review, rigorous verification, and thoughtful refinement by appropriately qualified and experienced engineers. Well crafted prompts make a massive difference in the quality of model outputs. Spend the time to create reliable prompts. Check out Google’s Prompting Strategies for a good starting point.

• Use modular design approaches. These allow for more straightforward independent verification of constituent parts. This is an obvious inclination in theory but it’s actually pretty difficult to think like this. If you can develop this skill, it’s the hallmark of good design. You can read more about these principles here:

• You can use robust chained prompts to systematically check AI-assisted outputs against applicable codes, standards, established best practices, and a wide range of plausible edge cases. This approach is not bulletproof but it’s pretty good. There is still no substitute for experience and focused human review of engineering deliverables. For my technical writing deliverables, I use a prompt based on our company’s report style guide with a number of additional instructions that has had a big positive impact on the quality of the documents.

• Insist on transparent documentation of all assumptions, AI inputs, and the complete verification process. "Black box" solutions are anathema to responsible engineering.

The Craft of Polishing

There's a trap in accepting AI outputs that are impressively complete on the surface but lack the robustness required for real-world engineering, the "prototype illusion," if you will. This missing "polish", the comprehensive error handling, the diligent consideration of edge cases, the practicalities of construction and operation, is, in our field, the hallmark of professional diligence. It includes:

• Designing for constructability, long-term maintainability, and eventual decommissioning.

• Ensuring unambiguous, complete, and accurate documentation.

• Realistically accounting for manufacturing/installation tolerances, material variabilities, and challenging site conditions.

• Committing to thorough peer reviews designed to identify subtle flaws and overlooked risks.

This experience-driven aspect of engineering, which demands a holistic view of the entire lifecycle, remains firmly in the human domain. For now, we are bound by the legal and ethical frameworks of our profession, where a licensed individual assumes personal responsibility for the life-safety aspects of their designs.

While AI tools will continue to consume discrete tasks, a human will remain responsible until a technology company is prepared to assume the legal liability for an AI-produced design. A more plausible, and unsettling, near-term scenario involves engineering firms taking on that risk themselves: loading up a single individual with an army of AI design agents while maintaining the fiction of ‘human oversight,’ where one poor sucker is expected to check and validate the design outputs of a fleet of algorithms.

I don’t mean to be a doomsday prophet but I’d be lying if I said I was not nervous about these possibilities.

Agentic AI and Engineering Judgment

The development of "agentic" AI – systems with more autonomous planning and execution capabilities – will undoubtedly continue. This might eventually mean AI agents performing proactive code compliance assessments or exploring vast design permutations based on precisely articulated human criteria. The ability to define these criteria with unambiguous clarity will become an increasingly vital skill. I feel like I am already at this point in my own workflow.

Even these advanced agents will operate within frameworks established by human engineers. Our role will evolve, placing greater emphasis on systems architecture, the meticulous definition of requirements and constraints, and, crucially, the ultimate validation and acceptance of responsibility for the outcomes.

Conclusion

The core challenge in engineering has never been a shortage of calculation speed. It has always been, and remains, the intellectual rigor of conceptualization, the complexities of system integration, the navigation of competing trade-offs, and the unwavering ethical responsibility for public safety.

AI can be a profoundly useful tool. It can accelerate tasks, automate processes, and expand our capacity for exploring alternatives. But it is a lever that must be wielded with thoughtful oversight and critical thinking. It is not a substitute for an engineer's analytical capabilities, ethical obligations, or holistic understanding.

Our objective must always be better, safer, more efficient, and genuinely sustainable engineered outcomes – not just faster outputs. The judgment of the professional engineer must continue to direct the application of these powerful new tools, ensuring they serve, rather than subvert, the true aims of our profession.

The future of engineering with AI will be remarkable. It will also, I suspect, generate a remarkable amount of work for the engineers tasked with cleaning up the mistakes made at such high speed.

I am not sure where this road is leading but I am hurtling down it like I’m part of a Mad Max Convoy.

Thank you all for your time and I’ll see you in the next one.

James 🌊