Can AI Really Code? MIT Researchers Map the Roadblocks to Autonomous Software Engineering

Picture this: a future where artificial intelligence quietly handles the tedious parts of software development—cleaning up messy code, updating legacy systems, or tracking down elusive race conditions—while human engineers get to focus on what they do best: designing architecture, solving tricky problems, and pushing the boundaries of innovation.

But before we settle into that vision, it’s worth pausing for a reality check. New AI tools have made serious strides, but a recent MIT study throws some cold water on the notion that AI will soon take care of all our software headaches. Instead, it maps out the obstacles we’ll need to overcome if we want AI and developers to truly work hand-in-hand, rather than at cross-purposes.

One thing the study is clear about: writing code is just a small part of what real-world software engineers do. Beyond the fun of inventing clever algorithms, there’s a mountain of “grunt work”—refactoring old code, moving huge systems to new platforms, testing, debugging, keeping legacy projects alive, even the less glamorous business of documenting and reviewing code. AI has started to chip away at some of these chores, but the road ahead is far from smooth.

For all that AI can automate, it doesn’t mean programmers are on the chopping block just yet. Armando Solar-Lezama, MIT professor and one of the study’s authors, points out that while current AI tools are impressively capable, we’re still far from the point where we can turn over full responsibility for most development tasks to machines.

There’s an especially thorny problem when it comes to evaluating how “good” AI actually is at taking on real-world coding. Most benchmarks look at neat, self-contained problems—nowhere near the messy reality of full-scale software projects. As a result, it’s hard to tell whether AI is ready to help with massive performance overhauls, not to mention complex situations like collaborating with humans or patching up tangled, decades-old code bases.

Another major sticking point is the handoff between human and machine. Communication often falters—AI has a habit of churning out giant, opaque snippets of code, making it tough to spot what might cause trouble once the application is live. A recipe for disaster? Maybe not, but it’s definitely a reason for caution. If AI systems could “know what they don’t know” and ask developers for clarification, trust in the human–AI partnership would be much easier to build.

And then there’s the problem of context. Many AIs are trained on code from public repositories, which doesn’t always reflect the unique conventions and standards of specific organizations. The result? Code suggestions that don’t fit the needs—or even compile correctly—in real-world enterprise environments.

The study doesn’t just leave us staring at the hurdles. It suggests a community-wide approach: better, more realistic benchmarks for AI and developer collaboration, and making AI systems more transparent, so humans can intervene when needed. The message is that AI shouldn’t aim to replace developers, but to boost their abilities—freeing up humans to do more of the designing, innovating, and decision-making that still surpasses what machines can offer.

Given how much of our world runs on software—from hospitals to highways—getting this partnership right matters more than ever. As AI’s role in coding continues to grow, keeping a clear-eyed view of what it can, and can’t, do is crucial as we chart the next steps forward in software engineering.

Read the original article on MIT News here: https://news.mit.edu/2025/can-ai-really-code-study-maps-roadblocks-to-autonomous-software-engineering-0716