DEV Community

Cover image for Return on Attention: Why AI Code Reviews Are Wearing Us Out

Return on Attention: Why AI Code Reviews Are Wearing Us Out

christine on July 09, 2026

Our team grew this year, and the PR volume has grown with it. Certainly faster than the ticket generation. New people means more code moving throug...
Collapse
 
kenwalger profile image
Ken W Alger

I'm seeing organizations doing random "PR reviews" during weekly standups. The team lead will randomly select a PR from the last week and walk through the code changes with the team with the person whose name is on the PR being the one answering the questions around coding style, edge cases, etc. If the code and PR comments are all AI generated, that's fine, you just need to be able to speak to it and defend it.

Different teams handle the "well, that's just what Claude Code wrote" answers and reasoning differently.

Nice post.

Collapse
 
cseeman profile image
christine

Hmmmm interesting, that probably is a good check on PR Reviews...if a bit nerve wrecking for some devs that might feel called out. I would hope they frame it as a learning opportunity not a pile on anyone ๐Ÿ˜…

Collapse
 
kenwalger profile image
Ken W Alger

I suppose that it depends on the team's dynamic, but the intent as I most often see it is a learning opportunity, yes.

Collapse
 
manolito99 profile image
Lolo

"Return-on-Attention" is a great way to think about it. One thing I've noticed is that AI doesn't just generate more code, it generates more things to verify. If every suggestion requires careful validation, you've only moved the work around. The best AI tools reduce the number of decisions you have to make, not increase them.

Collapse
 
cseeman profile image
christine

Oh boy does it generate more things to verify!

Collapse
 
bucabay profile image
Bucabay

Return on Attention - That's a good way to see it, better than describing the actual activity, but describing the value of attention as a scarcity.

You guys remember when XMLHttpRequest came out? There we thousands of libraries abstracting it, many were bad, some lived on - axios comes to mind. It's such a vital part of software and it took so long to get it down. Using it was tricky, there were many nuances you needed to know.

Now the abstraction is just await fetch(). We had to get await, and a standard to get a simple abstraction that simplifies understanding it.

With AI we seem to be going the other direction. Everyone is in a frenzy to build stuff, so we're building "long tail" software that aren't good under the hood. We are actually turning attention away from the more important nuances.

I think we need to formalize AI development - like we do programming standards. It would be more fuzzy, but it needs better structure. MCP is a great example, but we need something similar for programming structures.

Before we had design patterns to guide the fuzzy stuff, formal standards that guided abstractiosn and implementations. We need these built in, real AI frameworks that choose these and keep the AI in line.

Like how you'd use package.json to describe your dependencies, we need a schema to describe AI development that is predictable.

I believe that would allow us to automate most of the review - the implementation. We just need to focus on the - did it work (because the implementation is standardized it can be tested).

I'm just rambling but right now it seems we're driving AI blindly most the time - or stuck in review. AI will produce more and more faster - we need to change how we build the structure that produces what's reviewed.

Collapse
 
cseeman profile image
christine

I appreciate your thoughts here! It is always nice to see another dev who has been through multiple generations of software. I still remember J2EE and such and oof full hand generated WSDLs, man those sucked.

Collapse
 
alexshev profile image
Alex Shev

Return on attention is the right metric for AI review. More comments are not automatically more review. If the bot creates five weak notes for every one real issue, it turns human attention into the bottleneck. The best AI reviewer may be the one that stays quiet until it has evidence worth interrupting for.

Collapse
 
michael_salinas_472fbf6c1 profile image
Michael Salinas

Thank you for sharing such an excellent post. I really enjoyed reading it.

Iโ€™m a Python Full-Stack Engineer with over 10 years of experience designing and building scalable software solutions for clients across a variety of industries. Along the way, Iโ€™ve learned that successful projects depend not only on strong technical execution but also on creating real business value.

With my recent contract completed, Iโ€™m exploring new opportunities to collaborate with professionals who value innovation, practical problem-solving, and long-term partnerships. I enjoy discussing ideas that combine technical excellence with sound business strategy, creating outcomes that benefit everyone involved.

I believe every connection has the potential to become something meaningful. If you're interested in exchanging ideas, exploring opportunities, or simply connecting with someone who enjoys building impactful technology, I'd be happy to hear from you.

Wishing you success in your future endeavors, and I look forward to connecting.

Collapse
 
cseeman profile image
christine

Michael, thanks for reading the post. And best of luck out there with your next gig! If I hear anything Python-wise, I will send the details your way.

Collapse
 
michael_salinas_472fbf6c1 profile image
Michael Salinas

Thanks. Christine
If you allow, I wanna discuss about Opportunities and collaboration with you as soon as possible. Of course, this is an excellent opportunity for you.
How about you?

Collapse
 
michael_salinas_472fbf6c1 profile image
Michael Salinas

If you have free time, I wanna discuss each other to get income.
Best

Collapse
 
motedb profile image
mote

The ROI collapse in code review is real, but I'd push back slightly on where the problem originates.

You describe bots reviewing bots รขย€ย” but the underlying issue isn't the AI per se, it's that we're running AI review as a parallel human-reviewer simulation. We give it the same artifact (the PR diff), the same output format (comments), and the same authority level (official review status), and then act surprised when it generates the same noise we were already drowning in.

What if AI review worked as a hypothesis generator rather than a verdict machine? Not "this is wrong" but "consider: what if this error propagates to the billing layer?" รขย€ย• a question only a reviewer with full system context would think to ask.

The trickier failure mode is when the PR is fine but the original ticket was flawed. Should the review bot surface that the problem being solved is the wrong problem? That's where human attention genuinely earns its keep.

Have you seen teams successfully reframe AI review as an async analysis layer rather than a parallel reviewer?

Collapse
 
cseeman profile image
christine

Yea, our teams do consider a good amount of the AI review items as nice to haves or nitpicks, so they really aren't considered blocking. Some of the time, then the dev will just make a follow up pr if they need.

Collapse
 
mnemehq profile image
Theo Valmis

The fatigue makes sense once you count what a reviewer actually has to check per PR now: not just "is this correct" but "is this still the pattern we agreed on," and nothing in the diff tells you which of those two questions you're supposed to be answering. That ambiguity is exhausting even before you get to the line-by-line read.

Collapse
 
mads_hansen_27b33ebfee4c9 profile image
Mads Hansen

โ€œReturn on attentionโ€ is a useful way to name the problem.

A lot of AI review tooling optimizes for finding more things to say, not for preserving reviewer attention. That creates the strange failure mode where the team gets more comments, more apparent rigor, and less shared understanding.

The best pattern Iโ€™ve seen is to force AI review output into a smaller contract: severity, evidence, reproduction path, and why a human should care. If it cannot provide those, it should stay silent or ask a question. Otherwise the bot is not reducing review load; it is spending the teamโ€™s attention budget on its behalf.

Collapse
 
glenallen profile image
Glen Allen

The idea of "Return on Attention" really stands out. As AI-generated reviews become more common, success shouldn't be measured by the number of comments produced but by how quickly a reviewer can make a confident decision. High-quality feedback reduces cognitive load rather than adding to it.

Collapse
 
alexshev profile image
Alex Shev

Return on attention is the right metric. AI review can add value, but only if it reduces the human's uncertainty instead of creating another inbox. A good review bot should surface evidence, risk, and suggested next action, not just more fluent commentary to process.

Collapse
 
voltagegpu profile image
VoltageGPU

As someone working on GPU infrastructure for secure ML workloads, Iโ€™ve seen how LLMs are increasingly used to automate code reviews, but the real challenge remains contextโ€”especially when models run on different hardware stacks like VoltageGPU. The fatigue mentioned in the article hits close to home; without proper alignment between tooling and actual deployment environments, even the best AI suggestions can miss the mark.

Collapse
 
muhammad_muntazir_92189c7 profile image
Muhammad Muntazir

The 'slop grenade' concept is so real. Just because an LLM can generate 500 words in two seconds, we forget that someone has to spend their limited cognitive energy reading it. Love the 'Return on Attention' frameworkโ€”definitely a great boundry to set!

Collapse
 
cseeman profile image
christine

Thanks for reading!

Collapse
 
akcumeh profile image
Angel Umeh

This is a really good one. I often catch Claude adding lengthy explanations to my comments that I race to remove because a human would neither understand them, nor believe me ๐Ÿ˜‚๐Ÿ˜…

Collapse
 
cseeman profile image
christine

๐Ÿ˜ I have definitely done this too.

Collapse
 
julianneagu profile image
Julian Neagu

I started treating AI review comments like draft commits. If I wouldn't say it myself, I delete it. Most PRs get better after cutting half the words.

Collapse
 
cseeman profile image
christine

Great way to frame it.

Collapse
 
khx2012 profile image
Khalid Salameh

well soooooooooo true but some AI generate oversimplified code (ChatGPT) some produce duplicated code (Cluade) lol

Collapse
 
threerouter profile image
threerouter

The "1430 chars for 10 chars of change" hit home. I've started treating AI comments like spam filters โ€” if it's longer than my attention span, it's getting collapsed.

Collapse
 
cseeman profile image
christine

Yea....that was definitely not a popular PR ๐Ÿ˜

Collapse
 
krupali_gadhiya profile image
Krupali Gadhiya

The "Return on Attention" idea really stood out. Definitely something worth keeping in mind while using AI.

Collapse
 
cseeman profile image
christine

Thanks for reading Muhammad! I appreciate the comment ๐Ÿ’–

Collapse
 
itskondrat profile image
Mykola Kondratiuk

every AI comment added is another triage call - trust it or flag for juniors. started treating AI review feedback like junior contributions: signal, not authority. that second-order cost never shows in the faster-reviews pitch

Collapse
 
kartik-nvjk profile image
Kartik N V J K

The bot-replying-to-bot loop is the failure I find hardest to contain, because confident wrongness from one model becomes input the next one treats as settled fact. The line about having to leave the PR to get context points at the actual metric worth watching: a review comment that sends you elsewhere has negative value, no matter how thorough it reads. I have started scoring review comments on whether they are self-contained and whether the flagged bug reproduces, and most of the noise fails both checks.

Collapse
 
cseeman profile image
christine

Then do you still respond on those comments? I have a co-worker who refuses to review a PR if it has too long of commit messages and just way too much in the PR description. Which is fair. Her time, and energy can be better focused with shorter, more concise artifacts....and it actually helps everyone on the team.

Collapse
 
voltagegpu profile image
VoltageGPU

As someone working with GPU infrastructure for confidential computing, Iโ€™ve seen how LLMs are increasingly used to manage code review noiseโ€”but the real bottleneck isnโ€™t the code, itโ€™s the signal-to-noise ratio in team communication. Tools like VoltageGPU help isolate sensitive data during training, but they donโ€™t fix burnout from endless bot-reviewed PRs.

Collapse
 
cornellfilip profile image
Cornel Cristian Filip

Hey Christine, your "Return on Attention" piece was great (and clearly did numbers on dev.to). I noticed it lives on christine-seeman.com first: how do you handle getting it onto dev.to after, by hand or a script? I'm digging into that exact problem before building anything, so I'd love to hear your actual workflow.

Collapse
 
cseeman profile image
christine

I do this pretty much by hand. So I store my draft posts locally as .md, then have a python script to post on my site (WordPress). Then to x-post to dev.to I have it pretty much cat the markdown to my clipboard, and then paste into a new post on here. There is a small .py script I use for that cat-ing of the markdown and such ...it's here: github.com/cseeman/blog-scripts#to... . I think there is an API here...but this was what has worked for me without too much effort to post both on WordPress and here. I had RSS setup for a while, but I changed themes and that stopped working. I think the RSS feed into dev.to is a good approach, if that works for your site.

Collapse
 
mudassirworks profile image
Mudassir Khan

the 'bot commenting on a bot's suggestions' loop is the one that finally made us gate AI review behind a delay. we had the bot run 5 minutes after the first human review, not before. completely changed the dynamic: humans anchored the thread, the AI added instead of defining it.

the attention cost is the part your post names correctly and most tools don't acknowledge. you can't review 40 AI comments at the same depth as 5 human ones, and the signal to noise ratio isn't there yet.

are you seeing different patterns across PR sizes โ€” does the bot noise get proportionally worse on larger PRs, or is it a flat overhead regardless?

Collapse
 
polterguy profile image
Thomas Hansen

If you're tired of code reviews, you might want to change your programming language. 20% of the tokens compared to Python. 10% relative to C# and Java. And 10 times as high success ratio as Fable on Python, C#, JS, etc ...

Collapse
 
jugeni profile image
Mike Czerwinski

Return on attention is the right unit, and the version that bites is that a review comment has to be cheap to verify or cheap to dismiss, and slop is neither. Confident wrongness is expensive for one reason: at a glance it is identical to confident rightness, so the reader has to spend attention to tell them apart, and that cost lands on every comment whether or not it earned reading.

A colleague's terse note is cheap to act on because it carries the why in the same breath. A bot comment that makes you leave the thread to reconstruct context has already cost more attention than the bug it found was worth. The move that would actually shift your metric is not better-sounding bots, it is making each automated comment carry the evidence that lets you resolve it in one look, cite the line it checked and what it expected, or do not post. Volume was never the problem. Unverifiable-per-glance was.

Collapse
 
alexshev profile image
Alex Shev

Return on attention is the metric most AI review tooling ignores. A bot can produce technically relevant comments and still make the team worse if it increases the number of decisions humans must triage. The better product question is not โ€œdid it find issues,โ€ but โ€œdid it spend reviewer attention well?โ€

Collapse
 
eduzsh profile image
Edu Peralta

The 10 character diff getting a 1,430 character description is the exact failure mode I hit running a couple of coding agents on the same repo. The AI reviewer treats verbosity as thoroughness, so I end up scanning three paragraphs to find the one line that actually changed behavior. Your accountability question lands hard because I have caught myself pasting a bot's explanation into a PR description without reading whether it was even true. The real fix might not be a better AI reviewer at all. It might be a rule that nothing gets attached to a PR unless a human read it first and would defend it under questioning.