One point four million real workplace conversations with AI were analysed last year. About 5% of people were using it in ways that improved the quality of their work.

The “Are you using AI?” conversation is over. Nobody won it. We just stopped having it. But the question we skipped is, “Is anyone reading this stuff before it gets sent?” Increasingly the answer is no, and we can all tell. We’re all on the receiving end of each other’s output now, and most of us are silently frustrated by the same thing.

Hi, I’m Rachelle Rathbone. You might remember me from such posts as I’m more productive than ever. I’m also completely fried. A few months on I’ve found a rhythm where AI use isn’t beating me into the ground, but I’ve picked up a new beast instead: AI slop. Less what AI is doing to my brain, more what it’s doing to the work that lands in front of me, when it reads like nobody read it first.

A normal shortcut that causes real damage

There’s a real difference between using AI to help you think and using it to avoid thinking. In the first version, you use the tool to move faster through a problem you understand, explore options you would eventually reach yourself, and handle mechanical work while you focus on the decisions that matter. In the second, you hit “generate,” skim the result for anything obviously wrong, and send it on, hoping nobody looks too closely.

Microsoft Research’s CHI 2025 study surveyed 319 knowledge workers about 936 real uses of AI at work, and found something worth sitting with. The people who put the least effort into checking AI output were the ones with the most confidence in the AI. The people who put in the most effort were the ones with the most confidence in their own judgment on the task. Trust in the tool and scrutiny of the tool move in opposite directions.

Let me be clear: this isn’t a story about people being stupid. Rather, it’s a story about a very normal shortcut that causes very real damage. If the output looks good enough, your brain files it as done and moves on. But “looks good enough” and “is actually good” are not the same thing. That gap is where trust, quality, and shared understanding go to die.

And it isn’t just happening in one corner of the business

You can see the same pattern anywhere people use AI to make work that other humans have to read, review, or trust. Every job has a version of this: the thing you write to tell someone else what you did and what they should look at. Engineers call it a pull request description. Everyone else calls it a handover, a brief, or a status update. Same job, same failure mode.

Pull requests

The PR description is the most useful thing a reviewer gets. It should tell me what changed, why it changed, and what deserves my attention. Instead, I’m seeing long AI summaries that mention every file while explaining nothing. The author saved five minutes. Every reviewer spends fifteen minutes working out what matters.

It’s not just descriptions. Risk classifications marked “low” on changes that touch production infrastructure across several deployment targets. PRs spanning dozens of files and several ownership areas, submitted as one enormous blob because nobody stopped to make the work easier to review.

The numbers are not great. GitClear’s analysis of 623 million lines of code changes between 2023 and 2026 found that refactoring is down 70%, code duplication is up 81%, and error-masking patterns such as empty catch blocks and swallowed errors are up 47%. CodeRabbit’s analysis of 470 real-world pull requests found 1.7 times more issues in AI-generated code overall, and 2.74 times more security vulnerabilities. (CodeRabbit sells code review, so read it with that in mind. The direction still matches everything else here.) We’re producing more code, faster. We’re also producing worse code, faster. The quantity dial went to eleven. The quality dial stayed put.

Confluence pages

You know the pages I mean. They look like a chatbot answer pasted into Confluence and published without anyone changing a word. No voice, no opinion, and no sign that someone who knows the work has edited them. They go on forever, feel like they are going in circles, and give a newcomer no useful context. There are no decisions with reasons, no trade-offs, and no “we tried X and rejected it because Y.” It is confident prose that disappears the moment you need to make a decision from it. A load-bearing wall made of meringue: impressive until you lean on it. Status updates have the same problem. “Progressing well with some challenges being actively managed” tells you nothing about which challenges, since when, or what happens next. If your update survives a find-and-replace of the project name, it isn’t an update.

Slack messages

This one is my favourite because it’s so easy to spot and so awkward to call out. Someone asks, “hey, do you know if we need to update that config?” and gets a five paragraph response in return. The reply has clearly been pasted from an AI chat without being adjusted for the thread. It is oddly formal, full of em dashes, and ends with “I hope this helps clarify the situation.” Nobody talks like that in a DM. You know it, I know it, and the person reading it definitely knows it.

I want to be clear: I’m not talking about using AI to help draft something and then editing it into shape. That’s the good version. I’m talking about the unedited paste, the AI-to-Slack pipeline with no human processing in between. Everyone can tell. It starts feeling like bots talking past each other through human avatars.

Your brain is doing less work

There’s a reason this output lands so hollow, and it isn’t the tooling. It’s what happens to the person on the way through.

An MIT Media Lab study from 2025 used EEG monitoring on 54 people and found that those writing with an LLM showed up to 55% less brain connectivity than people writing without one. The study also found that 83% of the AI-assisted group could not accurately recall or quote from essays they had just written. They produced the words without really taking them in. They were typists, not thinkers, and they didn’t even get the carpal tunnel to show for it. (It’s a preprint with a small sample, so hold it lightly. It’s also not the only signal pointing this way.)

A separate MIT study published at CHI 2026 tracked 67 people using AI to check news over four weeks. In the short term, they were 21% more accurate with help, which sounds great. By week four, their unassisted accuracy had fallen more than 15 points below where they started. The tool helped, then the skill weakened. Critical thinking is a muscle, and AI is very happy to carry the heavy things for you. Generous, until your arms stop working.

The MIT researchers called this cognitive debt, which is a phrase worth sitting with. Not damage. Debt. Something borrowed cheaply now that gets repaid later at a worse rate.

And the part that should worry us most is that the misinformation study found about a quarter of participants believed they were improving while they were measurably getting worse. The skill goes first, the confidence stays. Which means nobody sends an email saying “heads up, I’ve stopped thinking about this.” It just gradually stops being obvious that anyone was.

The real cost

This is what happens when unedited AI output becomes normal in a team. None of it arrives as one dramatic failure. It builds up quietly, like technical debt for your team’s ability to get things done.

Reviewer time gets stolen. Every description that says nothing, every page that is fluent but empty, and every Slack message full of irrelevant context creates work for someone else. They have to work out what is actually being said, ask questions, or make assumptions. The author saved time. The organisation lost more of it. It is the collaboration equivalent of cleaning your room by shoving everything into the hallway.

Trust erodes quietly. When you cannot tell whether someone understands their own work, you review more carefully, ask more questions, and wonder whether the whole thing hangs together. That is a tax on every interaction.

Context disappears. AI output is good at describing what exists and bad at preserving why it exists. The decision, the trade-offs, the “we tried this and it broke in staging” are what someone needs six months later. AI wasn’t in the room when those decisions happened, so it cannot recover them from thin air. Future-you is going to be furious with present-you, and future-you will be right.

The bar drops. This cost is the most dangerous because it reinforces itself. If nobody pushes back on shallow handovers, empty pages, or AI-pasted Slack messages, that becomes normal. Then the people who still put effort into their descriptions, docs, and communication start to look as though they are overdoing it. Doing the work well becomes the odd choice. You are allowed to send poor work back. “Can you give me a two-line summary in your own words?” is a reasonable ask, not a hostile one. So is “which part of this do you actually need me to look at?” If nobody ever asks, the shortcut has no cost, and a shortcut with no cost becomes the default. I refuse to live in that timeline, and if you’re reading this, I suspect you do too.

Argue with it

The thing I’d most like to change about how people work with AI isn’t the editing. It’s the deferring.

Most people use these tools like a vending machine. Request in, output out, move on. The good version looks much more like working with a colleague. You explain what you’re actually trying to do. You push back when the first answer is generic. You ask how it got there. You tell it which part is wrong and why. You have, occasionally, a mild argument.

KPMG and UT Austin’s McCombs School analysed 1.4 million real workplace AI interactions and found the most effective users weren’t the most frequent or the most technical. They were the ones treating the model as a reasoning partner: giving it direction and examples, requiring it to explain how it reached an answer, refining across several exchanges instead of taking the first output, and feeding back what wasn’t working. About 5% of people worked this way consistently. That number is the interesting bit, because it means the behaviour is rare, learnable, and almost never taught. Prompt tips are everywhere. “Disagree with it” is not.

The MIT misinformation researchers landed in the same place from a different direction. Their suggested fix was Socratic dialogue: AI that asks you guiding questions rather than handing over answers, because that’s what preserved people’s skill instead of eroding it. An answer you argued your way to is one you keep.

There’s a version of this you can feel in the moment. If you finish a session having agreed with everything, you weren’t collaborating. You were transcribing.

The test

If you can’t explain your output without re-reading it, you didn’t write it.

That’s it. It works for handovers, pages, Slack messages, design docs, status updates, investigation write-ups, whatever. If someone asks you “what does this say?” and you have to go back and read your own work to answer, you’ve handed over the thinking, not just the typing.

And hand over my heart – I’m guilty of this one myself. Some days you’re flat out, the output looks fine, and reading it properly is the one thing that feels skippable. I’ve sent the thing I didn’t fully read. This isn’t a standard I’ve met and am now recommending from higher ground. It’s a muscle I’m still building, and I put time into it every single day, because the alternative is slowly becoming someone whose name is on work she can’t explain.

How to use AI as a collaborator

  • Use AI to inspect, not to speak for you. AI is great at breaking a complicated piece of work into groups, spotting the blast radius, and finding problems you might have missed. Let it do the checking, then write the summary yourself: “This is a config change affecting all five worker types. I think the risk is medium, and I’d like you to look closely at these two areas.” That takes two minutes and saves every reviewer twenty minutes of guessing.
  • Edit like your name is on it, because it is. Cut the filler, add your actual voice, and remove the headings AI reaches for every time, such as “Key Considerations,” “Important Factors,” and “In Conclusion.” Add the details only you know: the context, the trade-offs, and the things you’re unsure about. “I think we should do X because of Y, but I’m not sure about Z” is more useful than a perfectly tidy page that says nothing. Admitting uncertainty is useful.
  • Structure the work for the humans reviewing it. A 25-file change makes perfect sense to an AI because it can read all 25 files at once. It is much harder for the three humans reviewing it, each with a different area of ownership. The same goes for risk: if your tools flag eleven ungated code regions and you mark the change “low risk,” you have not assessed the risk. You have typed a word. Split the work into sensible pieces, explain what could go wrong, and give reviewers something useful to focus on.
  • Use AI to speed up the parts that should be fast, not skip the parts that should be slow. Boilerplate, scaffolding, and first drafts can be fast. Deciding what to build, explaining the approach, describing your thinking, and asking what could break in production need judgment. If AI makes the judgment-heavy work feel as quick as the mechanical work, that is not a win. It is a sign you may have stopped doing it.
  • Remove the bloody em dashes. Humans rarely use these, AI loves them. The #1 dead giveaway you’ve put zero effort into your ‘work’ are these things —

This is a quality problem dressed up as a productivity win

Each one is a small choice in favour of speed over quality, often made without anyone noticing. One bad handover will not bring down the company. Hundreds of them, thousands of pages, and tens of thousands of Slack messages can slowly wear away the shared understanding that lets an organisation work. It is death by a thousand AI-generated bullet points, and nobody ordered the funeral.

None of this is a discipline problem, and I don’t think it’s a people problem either. It’s a standard problem. Standards aren’t set by policy, they’re set by what gets accepted without comment. Every shallow handover that sails through, every empty page nobody questions, every unread paste that gets a thumbs-up emoji moves the line a little. Do that for a year and the people still writing their own summaries look like they’re gold-plating.

That’s the timeline I’d like us to avoid, and it’s avoidable, because the fix isn’t heroic. It’s a few minutes reading your own work before you hand it over, and a willingness to ask the same of other people without feeling rude about it.

Ultimately, the output has your name on it. Act like it.

📚 References and sources

KPMG LLP and UT Austin McCombs School of Business (2026): Analysis of 1.4 million real workplace AI interactions found that roughly 5% of users consistently engage with AI in ways that materially improve work quality, characterised by problem framing, directing the model’s reasoning, and iterating rather than accepting first outputs. Published in Harvard Business Review, March 2026.

Microsoft Research (CHI 2025): Lee, Sarkar, Tankelevitch, Drosos, Rintel, Banks and Wilson, “The Impact of Generative AI on Critical Thinking.” Survey of 319 knowledge workers and 936 real-world uses. Higher confidence in GenAI predicted less enacted critical thinking; higher self-confidence on the task predicted more.

MIT Media Lab (2025): “Your Brain on ChatGPT: Accumulation of Cognitive Debt.” EEG study of 54 participants; LLM-assisted writers showed up to 55% lower directed connectivity, and 83% could not accurately quote essays they had just produced. Preprint.

MIT Media Lab (CHI 2026): Rani, Danry, Liang, Lippman and Maes, “Dialogues with AI Reduce Beliefs in Misinformation but Build No Lasting Discernment Skills.” 67 participants over four weeks: +21% accuracy with AI assistance, 15.3 point decline in unassisted accuracy by week four, with roughly a quarter of participants believing they had improved.

GitClear (2026): “The Maintainability Gap.” Analysis of 623 million code changes from 2023 to 2026: refactoring line moves down 70%, block duplication up 81%, error-masking patterns up 47%.

CodeRabbit (December 2025): “State of AI vs Human Code Generation.” 470 open-source pull requests (320 AI co-authored, 150 human-only): approximately 1.7x more issues overall (10.83 vs 6.45 per PR), 75% more logic and correctness issues, 2.74x more security vulnerabilities.

Microsoft Work Trend Index (2024): 75% of global knowledge workers reported using generative AI at work.