AI Productivity Loop: The Hidden Waste of AI Text

Both Buttons Live in the Same Inbox

This isn’t a hypothetical workflow, because the tools exist and ship together. Gmail’s Help Me Write can draft a full email from a short prompt, and the same Gmail now summarizes long threads with AI Overviews. Microsoft’s Copilot in Outlook does both as well: it drafts your message and condenses the thread on the other end. The inflate button and the deflate button sit a few clicks apart, sold by the same company.

That’s the strange part. If one side is paying to make text longer and the other side is paying to make it shorter, then the thing being passed between them was never needed.

Where the Time Actually Goes

AI writing tools are sold as a way to communicate faster, so you’d expect the saved time to show up somewhere. A closer look at the research suggests it often just changes address.

Stanford’s Social Media Lab and BetterUp surveyed 1,150 full-time US desk workers in September 2025 about what they called workslop, which is AI-generated work that looks finished but doesn’t move the task forward. About 40% said they had received some in the last month, and respondents estimated roughly 15% of the work they get fits the description. On average, each incident took close to two hours to sort out, which the researchers turned into about $186 per employee per month, or over $9 million a year for a 10,000-person company.

But the shape of the finding matches our loop: it’s cheap to produce and expensive to receive. The sender saves half a minute of being polite, and the receiver either spends that half minute reading or pays for a second AI call to undo it.

That’s why we’d call text a zero-sum game here. Nobody is adding information. One side adds words, the other removes them, and the bullet point at the end carries the same content the five-word message started with.

What One Pass Through the AI Productivity Loop Costs in Electricity

The environmental part of this needs some care, because the honest numbers are smaller than the dramatic version. Google published figures in August 2025 for the median text prompt in its Gemini app: about 0.24 watt-hours of energy, 0.03 grams of CO2 equivalent, and 0.26 milliliters of water, roughly five drops. Google compares the energy to watching television for under nine seconds.

Run that AI productivity loop through our example. If writing the email and summarizing it each count as one median prompt, the round trip costs around 0.5 Wh and about ten drops of water. That is nothing you would ever notice on an electricity bill. Those figures are Google’s own, they cover only text prompts, and independent estimates run higher. A 2024 study from the Electric Power Research Institute put a typical ChatGPT prompt at around 2.9 Wh, for example.

Scale is where it becomes worth mentioning. As an illustration, if a 10,000-person company ran ten of these loops per person per day at Google’s median figure, that would be about 48 kWh a day spent on messages that get deleted on arrival. It’s not a scandal, but it’s real GPU time, and what it buys is a greeting that someone’s software then removes. Our view is that this is less an environmental argument than a question of what we choose to spend compute on, and “I hope this email finds you well” is a strange thing to spend it on.

Why Politeness Stopped Working

The reason the AI productivity loop happens in the first place is a social one. Many of us treat a short message as a rude one, so we want the speed of a text but feel obliged to perform an email. Historically that performance meant something because it cost the sender effort. Taking the time to write warmly was the signal.

Once a machine writes the warmth, the signal is gone, and there’s evidence readers notice. Peter Cardon and Anthony Coman surveyed 1,100 US professionals on how they judged messages from supervisors written with different levels of AI help. When the AI help was light, 83% saw the supervisor as sincere. When it was heavy, that fell to between 40% and 52%. Perceived professionalism also dropped, from 95% to between 69% and 73%. So the heavily polished version came across as less sincere and also less professional, which is the opposite of what the polish was for.

There are limits to that study. Participants were shown which parts were AI-written, which real inboxes don’t do, and people judged their own AI use more leniently than their boss’s. The researchers also found that routine messages like meeting reminders were fine to automate, while praise, empathy and feedback were where trust suffered. Our reading is that the problem was never AI itself. It’s padding, because padding exists to signal effort, and effort is exactly what a prompt doesn’t contain. In real life nobody needs a label either: a reschedule that arrives as three paragraphs is already telling you how it was made.