Are we confusing motion with progress in our AI adoption?

We all feel the speed. The boilerplate appears instantly, the functions auto-complete, and the "boring parts" vanish. But if the sensation of speed is this strong, why does the data suggest we might actually be slowing down?

A recent randomized trial from METR challenges our intuition:

• If we feel 20% faster, but are objectively 19% slower, are we measuring the right things?
• Why does the productivity gap widen so drastically between junior developers (who see gains) and senior engineers (who see little to no speed-up)?
• Are we trading the mental load of writing code for the subtler, perhaps heavier, load of reviewing and debugging it?

And then there is the question of technical debt. Apiiro’s research points to a 322% rise in privilege escalation paths in AI-generated code.

• Are we shipping features faster today at the cost of a massive security remediation bill tomorrow?
• Is "more tasks completed" a valid metric if it comes with a 153% increase in design flaws?

We explored the "Productivity Paradox" to understand why our perception is so at odds with the reality of shipping production-ready software.

Read the full analysis here