When AI can write the code, what is left for engineers to do?

Massachusetts Institute of Technology (MIT) has been a pioneer in engineering education for well over a century. Once in a while, MIT revises its General Institution Requirements (GIRs) for all undergraduates. These revisions reflect what a well-trained engineer and scientist is expected to know and provide broad-based foundational skills. This year, in May 2026, MIT’s faculty and administration released a report describing what the next revision of the GIRs should look like. The most prominent changes proposed include the addition of compulsory courses on (1) computation and (2) probability, statistics, and machine learning.

Given the preeminent position AI and data play in the current landscape, this is not surprising. But the intent is far-sighted. Consider the following statement from the report: “AI will likely take over the writing of straightforward code, but students will, we believe, more than ever, need computational thinking to engage with it deeply, and hopefully steer its direction: understand how to translate a problem so it can be solved computationally, and to recognize and debug incorrect algorithms.”

To understand what this might mean in practice, consider the field of computer vision. My first foray into computer vision was over 25 years ago, and the field was quite different then. To rephrase Russell and Norvig (from the Preface of the recent edition of their seminal textbook on Artificial Intelligence): more hand-crafted knowledge engineering than machine learning. Data and AI have changed all that in the intervening years. Let me illustrate by taking an example from the domain of civil engineering.

Surveyors need to measure precise distances and angles to enable, for instance, the construction of bridges and tunnels. The usual way to do this is to set up a lightweight surveying pole that serves as a target unit. A base unit at a fixed point emits laser or infrared light towards the target unit and measures the reflected light from the target unit (there is a ‘prism’ at the top of the surveying pole). This data can then be processed to extract the precise distance and angle between the two units.

Enhancing capabilities

It is fairly straightforward to code this process. But companies like Leica Geosystems, in their TS20 base unit, have added a Neural Processing Unit (NPU). They have done this to enhance computer vision capabilities in various directions. First, to automatically check that the prism type is correct. Second, to follow the target unit during line-of-sight interruptions and then relock to continue measurements. Third, to accurately measure in challenging atmospheric conditions, like misty rain. To a civil engineer, these are known, even obvious, hazards. To a computer programmer, they need to be explicitly identified as system requirements.

This is what it means to ‘translate a problem so that it can be solved computationally’, harking back to the MIT faculty’s perceptive observation. The surveyor using the TS20 knows how field conditions affect the performance of the system. This engineering reasoning is part of his or her training and experience. On the other hand, the computer vision algorithm uses signal processing and data analysis in ways the programmer of the NPU has envisioned. (In fact, the NPU is capable of software upgrades for needs not yet envisioned.) The challenge, therefore, is the translation of engineering reasoning to functioning programmes. This involves judgement, prioritisation, and a deep appreciation of constraints. AI-generated code can help smooth this process but cannot (yet) be its beginning and its end.

None of this is to say that AI is not adding a great deal of value to the engineering process. On the coding front, AI-generated code can reliably run computer vision algorithms that resolve all three of the TS20 base unit’s NPU capabilities. For instance, the third challenge of algorithms that enable vision ‘through’ rain involves a series of coding steps that include denoising, segmentation, and object identification, all of which an AI can now reliably generate with suitable instructions. Where engineering expertise is needed is in deciding whether the AI-generated code is doing the right thing in the right situation. Rain can also result in the prism on the target unit being partially covered, partly as a protective measure. This adds another layer of complexity to the measurement of reflected distance and angle. A field engineer knows and understands this. An AI will generally not, unless very specifically instructed to. This kind of testing and validation is a major role of the modern engineer in the AI-assisted era.

Twenty-five years ago, I was writing code that extracted hand-crafted features from images (in the parlance of Russell and Norvig). Today, I do not need to do that. Machine learning can discover the features, and generative AI can create the code to execute that. What I can do instead is think more about the surveyor’s challenges in the field, the radiologist’s diagnostic requirements, or the data captured by the camera on top of a self-driving car. In doing that effectively, I can add a particular kind of human value to those fields of engineering, based on a collaborative sense of judgement and perspective that goes beyond automation.

(The author is Senior Faculty & Director-Academics, Great Learning. Views are personal.)