Nobody Reads Every Line Anymore

Jack Ellis asked senior engineers whether they still review the AI-assisted code that hits production. Most who answered said yes. Aaron argues the risk, when teams stop reviewing that work line by line, is that nobody can say who owns the result. The full episode script, as an article.

Jack Ellis asked senior engineers whether they still review the AI-assisted code that hits production. Most who answered said yes. Aaron argues the risk, when teams stop reviewing that work line by line, is that nobody can say who owns the result. Below is the full script of this episode as an article.

On November fourth, Jack Ellis posted something that kept me up a little. He’s a builder. He runs Fathom Analytics. And he was not convinced that senior software engineers are quietly shipping A-I code without reading it. So he tagged eleven experienced engineers who work on high-stakes software, and he asked them, straight, whether they still review what goes out.

Most of the ones who answered said yes. They read it. Not always every line of every side project. But for the work that hits production, they still put their eyes on it. One of them said, without exception. Another said at work he reads it hard, especially early, when the foundation is being laid. A third said they still review before merging, even while they lean more on agentic verification. Jack’s own line was the one I underlined. The human is responsible for the work that hits production. The A-I cannot be held responsible.

Same day, Nikita Bier posted a line that went everywhere. Does anyone remember how to code. That was the joke of the week. Jack’s thread was the serious version of the same anxiety.

So here’s my claim, and you’re welcome to disagree with it. When teams stop reviewing A-I work line by line, the real risk is that nobody can say who owns the result.

The time moved

The time didn’t disappear. It moved.

There’s a Harvard job-market paper out this year by Fiona Chen and James Stratton. They looked at software production inside firms using a proprietary engineering-analytics dataset. Hundreds of millions of work events. Hundreds of thousands of workers. Hundreds of firms. They tracked what happened when companies adopted A-I coding assistants, and then what happened when they adopted coding agents.

Writing sped up. Lines of code and commits went up. Final software shipped did not rise nearly as much. Employment barely moved. The place that absorbed the shock was review.

After agents arrived, the average time from submitting a change for review to getting it approved went up by about fifty-six percent. Code waited longer in the backlog. Once a human actually started reviewing it, that active review also took longer. Comments went up. Change requests went up. Nearly eighty percent of the firms had already adopted some A-I review tool. Humans were still writing the overwhelming majority of the reviews and the comments.

In other words, the generation got cheap. The confidence did not. The hours that used to go into typing moved downstream into someone staring at a diff they did not write, trying to decide whether to put their name on it.

One engineer the researchers interviewed put the scene in a room. On one project, only one senior could approve changes, because he was the expert on that codebase. He had so little bandwidth he could review that person’s code once a week. Then A-I-generated commits caused outages. New rules required senior approval for A-I-generated code. The queue got worse.

The people in that story weren’t lazy. They were stuck behind a bottleneck with a name on it.

The review that ate the week

Now put a face on the other end of the same pattern.

In March, The Guardian published an investigation into Amazon’s A-I push inside corporate teams. One software engineer, anonymized as Denny, described a teammate who said an internal A-I agent had saved him about a week of developer effort on a feature. Then Denny looked at the actual code review. Dozens of comments. Basic issues. He called the A-I-generated code full of slop.

His read was blunt. The pressure to use the tools was producing worse quality code, and also just more work for everyone.

The week of savings did not vanish. It relocated. Into the review comments, and into the people who had to catch what the generator missed and still own the ship.

If you’ve been inside a team that celebrated generation speed while the review queue quietly doubled, you already know this feeling. The dashboard for writing looks fantastic. The calendar for the people who have to stand behind the writing looks worse.

Ownership, not nostalgia

There’s a useful distinction for what happens next, and it comes from Addy Osmani, who works on Claude Code at Anthropic. He wrote about it in early November, right around the same week as Jack’s thread.

He separates cognitive offloading from cognitive surrender. Offloading is handing tasks to the tool and keeping ownership of the result. He says he does that all day, and that you should too. Surrender is complete trust in the model’s output, with no need to check it. His test is simple. When you can’t explain the answer, even with the agent’s help, you’ve ceded ownership.

I’m trained in organizational psychology and family therapy, twenty-plus years of experience, and that line lands for me because ownership is a social fact, not a feeling. In a family system, the person who remembers the appointments and knows where the paperwork lives starts to look like the healthy one, until the day they step away and nobody else can find the electric bill. Teams do the same thing with review. One person keeps absorbing the unread diffs. Everyone else organizes around that person. The system works right up until you ask who actually owns the last thing that shipped.

Jack said it without the jargon. Until an A-I can be sued, a human is responsible. Osmani said the same thing from the other side of the keyboard. Whoever clicks merge owns it. A model cannot be paged at three in the morning.

The worry behind “nobody reads every line anymore” can sound like nostalgia for typing. Typing was never the scarce part. The scarce part was always a person willing to put their name on the result. When review becomes skim-and-approve because the volume won, the responsibility has been misplaced.

And once it’s misplaced, metrics can still look fine. Pull requests merge. Tests go green. None of those tell you whether a human can still narrate what shipped and why. That gap is where ownership dies quietly.

What you can do

Here’s what you can do this week, and none of it requires a new platform.

First, put a named owner on every review slot that matters. Not a role. A person. If the change touches customer data, or anything you would apologize for in public, write the name next to the merge. If you can’t name who would get the page, you have a queue.

Second, put an expiry on those slots. The Harvard interview scene was one senior reviewing once a week forever. That is how bottlenecks calcify. Rotate the hard reviews. Time-box how long a change can sit unread before it gets escalated or rejected. Unowned review slots age into rubber stamps.

Third, separate person edits from template edits. A lot of A-I output looks finished because it matches a house style. Looking finished is not the same as being owned. Ask the submitter, out loud, what they personally changed or verified, versus what the model filled in. If they can’t say, send it back before anyone spends senior time on it.

Fourth, put checking time in the quote. If a vendor, an agency, or an internal team sells you A-I speed, ask where the review hours live. Generation without review capacity is a deferred invoice, and it usually lands on your best people.

You do not have to read every line of everything. Jack himself skims the isolated side project and reads every file that hits the main product. That calibration is the point. Match the depth of the review to the blast radius of being wrong. Just do not pretend the unread middle has no owner.

So this week, ask one question in the next review meeting. For the last thing we shipped with heavy A-I help, who can stand up and say, I own that result?

If the room goes quiet, you found the risk.

—

This show extends the book AI Empowered. Read the first chapter free at https://www.aiempoweredbook.com, or get the book on Amazon.

Aaron Douglas runs Auspicious, a fractional marketing practice.

Blog: https://auspicious.llc/insights/
X: https://x.com/aaron_douglas
LinkedIn: https://www.linkedin.com/in/aaronddouglas/

Transcript

Hey, it’s Aaron Douglas. Welcome to AI Empowered.

On November fourth, Jack Ellis posted something that kept me up a little. He’s a builder. He runs Fathom Analytics. And he was not convinced that senior software engineers are quietly shipping A-I code without reading it. So he tagged eleven experienced engineers who work on high-stakes software, and he asked them, straight, whether they still review what goes out.

Most of the ones who answered said yes. They read it. Not always every line of every side project. But for the work that hits production, they still put their eyes on it. One of them said, without exception. Another said at work he reads it hard, especially early, when the foundation is being laid. A third said they still review before merging, even while they lean more on agentic verification. Jack’s own line was the one I underlined. The human is responsible for the work that hits production. The A-I cannot be held responsible.

Same day, Nikita Bier posted a line that went everywhere. Does anyone remember how to code. That was the joke of the week. Jack’s thread was the serious version of the same anxiety.

So here’s my claim, and you’re welcome to disagree with it. When teams stop reviewing A-I work line by line, the real risk is that nobody can say who owns the result.

The time didn’t disappear. It moved.

There’s a Harvard job-market paper out this year by Fiona Chen and James Stratton. They looked at software production inside firms using a proprietary engineering-analytics dataset. Hundreds of millions of work events. Hundreds of thousands of workers. Hundreds of firms. They tracked what happened when companies adopted A-I coding assistants, and then what happened when they adopted coding agents.

Writing sped up. Lines of code and commits went up. Final software shipped did not rise nearly as much. Employment barely moved. The place that absorbed the shock was review.

After agents arrived, the average time from submitting a change for review to getting it approved went up by about fifty-six percent. Code waited longer in the backlog. Once a human actually started reviewing it, that active review also took longer. Comments went up. Change requests went up. Nearly eighty percent of the firms had already adopted some A-I review tool. Humans were still writing the overwhelming majority of the reviews and the comments.

In other words, the generation got cheap. The confidence did not. The hours that used to go into typing moved downstream into someone staring at a diff they did not write, trying to decide whether to put their name on it.

One engineer the researchers interviewed put the scene in a room. On one project, only one senior could approve changes, because he was the expert on that codebase. He had so little bandwidth he could review that person’s code once a week. Then A-I-generated commits caused outages. New rules required senior approval for A-I-generated code. The queue got worse.

The people in that story weren’t lazy. They were stuck behind a bottleneck with a name on it.

Now put a face on the other end of the same pattern.

In March, The Guardian published an investigation into Amazon’s A-I push inside corporate teams. One software engineer, anonymized as Denny, described a teammate who said an internal A-I agent had saved him about a week of developer effort on a feature. Then Denny looked at the actual code review. Dozens of comments. Basic issues. He called the A-I-generated code full of slop.

His read was blunt. The pressure to use the tools was producing worse quality code, and also just more work for everyone.

The week of savings did not vanish. It relocated. Into the review comments, and into the people who had to catch what the generator missed and still own the ship.

If you’ve been inside a team that celebrated generation speed while the review queue quietly doubled, you already know this feeling. The dashboard for writing looks fantastic. The calendar for the people who have to stand behind the writing looks worse.

There’s a useful distinction for what happens next, and it comes from Addy Osmani, who works on Claude Code at Anthropic. He wrote about it in early November, right around the same week as Jack’s thread.

He separates cognitive offloading from cognitive surrender. Offloading is handing tasks to the tool and keeping ownership of the result. He says he does that all day, and that you should too. Surrender is complete trust in the model’s output, with no need to check it. His test is simple. When you can’t explain the answer, even with the agent’s help, you’ve ceded ownership.

I’m trained in organizational psychology and family therapy, twenty-plus years of experience, and that line lands for me because ownership is a social fact, not a feeling. In a family system, the person who remembers the appointments and knows where the paperwork lives starts to look like the healthy one, until the day they step away and nobody else can find the electric bill. Teams do the same thing with review. One person keeps absorbing the unread diffs. Everyone else organizes around that person. The system works right up until you ask who actually owns the last thing that shipped.

Jack said it without the jargon. Until an A-I can be sued, a human is responsible. Osmani said the same thing from the other side of the keyboard. Whoever clicks merge owns it. A model cannot be paged at three in the morning.

The worry behind “nobody reads every line anymore” can sound like nostalgia for typing. Typing was never the scarce part. The scarce part was always a person willing to put their name on the result. When review becomes skim-and-approve because the volume won, the responsibility has been misplaced.

And once it’s misplaced, metrics can still look fine. Pull requests merge. Tests go green. None of those tell you whether a human can still narrate what shipped and why. That gap is where ownership dies quietly.

Here’s what you can do this week, and none of it requires a new platform.

First, put a named owner on every review slot that matters. Not a role. A person. If the change touches customer data, or anything you would apologize for in public, write the name next to the merge. If you can’t name who would get the page, you have a queue.

Second, put an expiry on those slots. The Harvard interview scene was one senior reviewing once a week forever. That is how bottlenecks calcify. Rotate the hard reviews. Time-box how long a change can sit unread before it gets escalated or rejected. Unowned review slots age into rubber stamps.

Third, separate person edits from template edits. A lot of A-I output looks finished because it matches a house style. Looking finished is not the same as being owned. Ask the submitter, out loud, what they personally changed or verified, versus what the model filled in. If they can’t say, send it back before anyone spends senior time on it.

Fourth, put checking time in the quote. If a vendor, an agency, or an internal team sells you A-I speed, ask where the review hours live. Generation without review capacity is a deferred invoice, and it usually lands on your best people.

You do not have to read every line of everything. Jack himself skims the isolated side project and reads every file that hits the main product. That calibration is the point. Match the depth of the review to the blast radius of being wrong. Just do not pretend the unread middle has no owner.

So this week, ask one question in the next review meeting. For the last thing we shipped with heavy A-I help, who can stand up and say, I own that result?

If the room goes quiet, you found the risk.

If you’re picking up what I’m putting out, please subscribe, review or share the podcast with a friend.

You can find more insights on my blog at auspicious dot L-L-C.

There’s the book “AI Empowered” by Aaron Douglas. The first chapter’s free at AI empowered book dot com.

Social media in the show notes if you have any feedback to offer.

Most of all thanks for thinking with me today. I’m Aaron Douglas and this is AI Empowered!

Share

Leave a Reply