AI is making us write more code. That's the problem.
I analyzed research papers on AI-generated code quality. The findings:
→ 1.7x more issues than human-written code
→ 30-41% increase in technical debt
→ 39% increase in cognitive complexity
→ Initial speed gains disappear within a few months
We're building the wrong thing faster and calling it productivity.
The bottleneck was never writing code. It's understanding what to build.
If you're using AI coding tools, focus on:
• Smaller features (if it's 1000 lines, it's too big to review)
• Clear acceptance criteria before you prompt
• Tests first, AI-generated code second
• Security audits (AI can't do this)
More code isn't the goal. Solving real problems is.
@mlevison I had a discussion with some developers about this last week. If we expect developers to take responsibility for the code they generate, the amount of work is capped by their ability to review code. And if we expect more (which is absolutely necessary, if we want the promised productivity gains), we can't blame them, if it goes wrong.
If we celebrate huge productivity gains, we give up the right to complain, if AI fails. And it will fail.
@imcdowall @jose Early days. We're still learning.
I mostly use: https://github.com/PaulDuvall/claude-code --- I use /xspec to create the core acceptance criteria. I read and edit. Then use /xtdd to attempt to write the code. I keep all changes small and only commit what I can read and understand.
7.26K Posts, 904 Following, 530 Followers · Certified Scrum Trainer | I help organizations and teams become more effective. I use #Scrum, #Kanban #Agile #BehaviouralPsychology. | I’ve helped over 8000 people build better teams I dabble with #PKM aka Personal Knowledgement Management so the ideas that fall out of my head get written down somewhere #Obsidian. #fedi22
That's my experience too.
Dave Farley's. MSE channe, which I usually respect,l recently claimed the opposite though, based on a study they took part in.
As I've said to a few other people, the best teams with exceptional discipline may get better results.
In which case, I want data and I want measured by the CodeRabbit people so we have a basis for comparison.
Yeah I get the impression that the Dave Farley shop is quite professional and disciplined. Although they use Java and write mainly business / finance apps. Which are perhaps structurally relatively simple? They had a sample of 150 devs.
This is the video:
https://www.youtube.com/watch?v=b9EbCb5A408
Focus is on claude code and the maintainability by humans of code written by AI.
And this is the study:
https://arxiv.org/abs/2507.00788
@mlevison not sure what you are saying..
That they have a good reputation for serious research?
Right. Their methodology seemed sane to me.
The results surprising. But as you say perhaps helped by some very disciplined Devs.
@oschonrock I highly recommend the book: "Your Code as a Crime Scene" - Adam Tornhill.
https://codescene.com/hubfs/web_docs/Business-impact-of-code-quality.pdf this is the paper that I recommend to Scrum Masters and Developers.
Couldn't human made deterministic tools (or changes to programming languages) help with boilerplate work instead of indeterministic intransparent generative AI?
IIRC for some languages there also have been deterministic refactoring tools too that take over the tedious parts of refactorings (like "rename method" which exactly identified callers to adapt them).
What does that mean?
Oh ok. Thanks but no thanks for being judgmental about my thinking process.
@project1enigma these might help
The bigger issue is that GenAI has no judgment. No understanding of correctness; readability etc.
Better refactoring is great, but not enough.
Curious what languages are you referring to?
I personally work with C++ and am old fashioned and code with a text editor.
But the first time I read about refactoring tools, it was about the so called "refactoring browser" for Smalltalk.
IIRC nowadays there are tools for at least some of the usual refactoring steps in many IDEs for common programming languages. I'd be surprised if there were none for Java for example.
I'm personally also somewhat in favor of code generation for example for data marshalling/unmarshalling, parsing etc.
But that's deterministic specific generators then, either existing (lex, yacc and successors, for example), or in house/ad hoc.
Or using higher level languages where this can be done as libraries supporting domain specific languages instead.
(There is stuff like Boost.Spirit for C++, though that still feels less natural than say parser combinators for Haskell)
@project1enigma I do refactoring all the time. I still have a copy 1st edition of Refactoring - Martin Fowler.
They tools exist, but they don't necessarily help the genai tools. Maybe a Claude Plugin for refactoring?
"LLMs are just another layer on top of that, and a bad one"
That's my point. After you told about using those for tedious tasks.
@mlevison I hear you. I am seeing it first hand. And I am being hammered for pointing this out and encouraging caution, while others are being rewarded for shipping 💩 “because AI”.
Just call me ‘Cassandra’ 🤷♂️
Defining “productivity” as building the wrong thing faster has been the usual definition for at least my lifetime. Not just for code.
@mlevison can you post some links to those research papers?
In case you did not know: #acm's digital library is not open and free to use for everybody so most peer reviewed articles published in the us should be accessible. https://acm.org/dl
@mlevison The silver lining here (for developers who aren't using AI, despite demands from employers) is that we are easily able to keep up with developers who are.
My boss saw speed gains from some developers, and asked us all to. I kept going as normal, and after a bit of time for things to settle, I'm still faster.
I like to think that more and more people are realizing that the emperor has no clothes.
@mlevison My anecdotal guess is that people are operating above their own abilities.
If I understand a code base and its architecture, dependencies, business logic needs, etc. Then I can ask an AI to make a targeted change and verify it. It’s basically avoiding some typing.
But if I don’t understand those things I don’t actually know how to ask or how to verify it.
Even worse if I wouldn’t be able to write it in the first place, how can I ask for it and verify it?
Basically, how can I successful operate above my own knowledge and ability? It’s like asking for a piece of text written in a language I don’t speak. How can I have a clue what it says?
I think there are so uses for assisted development. I’m less sure an LLM is the way to achieve that.
Worst part, people aren’t learning anything from all this. There’s no skill and knowledge being built:(
Imho, ymmv, etc.
Also those are some numbers, yikes!
Julian Wecke
Mark Levison
Sam
Konstantin Weddige
jose
Ian McDowall
*_jߍyrope
Charles ☭ 
is a Green
Tormod
LunaticDancer
Oliver Schönrock
Piper
Gothmog
Chris Pitts
always tired
H4Heights 🇪🇺🇵🇸🇺🇦🇨🇦
Paolo Montalto
Ԏєηυкι, 手抜き🚀🐧♏ 🔭 ⚫⚪
Deborah Preuss, pcc 🇨🇦
Viktor Linsel
Minnesota 123456
r-hold
Flesh & bone by the telephone
四