AI got good at the structural side of writing fast such as outlines, drafts that need reorganizing, research synthesis, editing for clarity, formatting a page so it actually reads clean. That part’s basically solved. I use it for all of it, every day, and it makes the work faster and better. That’s not where I have a problem.
The problem starts the moment AI stops helping me build the container and starts generating the actual words meant for another person. A message, a caption, a post signed with my name, a story. That’s a different kind of output, and right now nothing in the stack distinguishes it from the structural stuff. It all just shows up as text.
The disclosure step is the first thing that gets skipped
You know how this goes if you’ve used any of these tools for more than a week. The output is good enough, often enough, that flagging “this part was AI-assisted” starts to feel like overhead. You’re moving fast, the message needs to go out, and stopping to mark which sentence came from where feels like a tax nobody else is paying. So it stops happening. Not because anyone decided authenticity didn’t matter, it’s because the workflow never asked the question in the first place.
Multiply that by however many people are running the same tools through the same pressure, and you get a feed, an inbox, a comment section quietly filling up with text nobody disclosed and nobody asked about.
Authorship should be a field, not an afterthought
The fix isn’t a disclaimer nobody reads. It’s the same fix as any other data quality problem: information about how something was made needs to live with the thing itself, not in a separate doc no one checks. If a message or a post was meaningfully AI-drafted, that should be visible at the point someone is reading it, not buried in a settings page, not something you’d only find by asking.
This is cheap to build into a workflow from the start and expensive to retrofit later, which is exactly why almost nobody has it. It’s a metadata field, not a moral stance. Build the convention in once and it costs nothing on every post after that.
Trust is the thing actually drifting
Here’s the part that makes this more than a style preference. The WHO put a number on what’s already happening to human connection at scale: their 2025 Commission on Social Connection report tied loneliness to roughly 871,000 deaths a year worldwide, with about 1 in 6 people globally experiencing it directly. The U.S. Surgeon General’s office compared the health cost of chronic loneliness to smoking a pack of cigarettes a day. That’s not a side effect of any one technology, it’s the baseline we’re already operating in. Dating has been at an all time low. Based on a paper written by the Institute of Family Business, only 30% of young adults are actively dating seriously or casually. This is showcasing that loneliness is happening at all levels.
Against that baseline, an inbox full of text nobody can confidently call human or not isn’t neutral. Most LLMs are trained on historical data, not our own personal data. Our brains are so complex that LLMs can only write a perfect text but not OUR text. These texts with our words is what gives these messages its essence. When the source is ambiguous, that signal gets quietly degraded, even if the reader can’t say exactly why. Nobody’s watching that the way you’d watch a drift metric on a dashboard. It just erodes, conversation by conversation, until the thing that was supposed to carry connection stops carrying it.
Where I draw the line
My rule: if I can’t tell the person on the other end which parts of this were me, AI doesn’t touch that layer. Research, structure, and organizing thoughts can be supported by AI to enhance my workflow. The actual words I use to talk to a specific person, or the story I’m telling that’s supposed to be mine, thats the part I write. At the end of the day, the authorship is the message.
This ethical perspective of mine came from a PwC event I attended in Chicago. Anthony Anderson (Senior Manager for Emerging Technology) discussed the balance of AI within the consulting space. Clients prefer dealing with humans because humans are the ones able to create solutions that could bring real change. It allows innovation to occur, which is something we should always strive for. LLMs will enhance our workflow, but it’ll never drive innovation because it’s trained on historical data. Humanize our approach to AI usage will allow us to stand out from others and create things far beyond our capabilities.
The gap is architectural, not moral
I’m not arguing AI-assisted writing is dishonest by default. I’m arguing the tooling doesn’t make disclosure the default output, so it never happens unless someone goes out of their way to add it. Unfortunately, almost nobody does, for the same reason nobody fills out a metadata field that isn’t required. That’s solvable the same way any provenance problem is solvable: decide what needs to travel with a piece of text, build it into the workflow as a default, stop treating “was this AI?” as a question the reader has to ask instead of one the system already answers.
We didn’t lose the ability to tell human writing from generated text. We built tools that stopped telling us which one we were looking at. That’s a design choice, and it’s one we can unmake.