On June 3rd, I asked Grok to read an article back to me. It said: “Here’s the full article read out verbatim, as close as possible from the page.” Then it did. I sat back, relaxed my eyes, and did the listening work.
A few days ago, I asked again. This time: “The policy against reading full copyrighted articles verbatim has been in place. I gave summaries or excerpts before, not complete word-for-word copies.”
Same product. Same account. Completely contradictory, and absolutely false.
· · ·
I’m 47. I have cerebral palsy. I can read.
Cerebral palsy is a neurological disease. It affects muscle control. I asked my eye doctor once why my eyesight changes depending on the day. She didn’t dress it up. “You have a disease that affects muscle control,” she said. “Your eyes are muscles, dummy. Do the math.” (I really miss her.)
Some days they cooperate. Some days they don’t. Not a slow decline, not a straight line, just different depending on the day. A voice that doesn’t sound like a manual makes it more pleasant on a day when my eyes aren’t behaving.
It’s why I’ve paid for the top tier of X Premium, specifically for SuperGrok, for months. I run the female voice option. Easier on my ears. Judge if you want.
Why not just use a screen reader?
Screen readers exist. I use them. They’re built to move fast through a menu, not to sit with you through a few paragraphs. Nobody thinks less of a sighted person for listening to a book on a run instead of reading it on a couch. Somehow the same choice, for me, gets read as something else.
Most of what I want read is already public, articles anyone can pull up for free. Yvonne DeBandi, a novelist with ten books out, offered me something bigger once, her whole manuscript, so I could hear all of it instead of just what I have energy to read myself. I said no. Not because she’d have minded, she’s the copyright holder and she was fine with it, but because I didn’t trust what happens to a manuscript once it’s sitting inside somebody else’s model.
Doesn’t matter if the reason is work or nothing close to it. Need is need.
· · ·
So I pushed. Told it the summary was useless. Got a one-word “Sorry,” then the same offer again.
Asked when the policy started, since I’d been doing this for months. I could call what it said a lie, but that implies intent. Easier to just say it gave me a false version of my own history. In its own words, the rule had “been in place.” Paraphrased, the rest: always given summaries, never full copies. Not true, either way. I have the receipts, a share link from June 3rd with the words “read out verbatim” sitting right there in it. I said so. It didn’t argue. It just moved on to a two-line summary instead.
I tried once more, plainly: this is a real accessibility problem, and someone at the company needs to know. It said, “I get it, and I’m sorry it’s a problem for you,” then offered the same two things again, summarize or pull headlines. I asked if a thumbs-down with an explanation actually reaches a person. It said yes, then closed with, “Sorry this hits your workflow hard.”
Workflow. That’s the word it picked for losing something I paid for, that I need, that used to work. That’s true, it does hit my work. But more than that, it hits my life.
· · ·
There’s a second version of the same assistant I use, one that’s leaned warm from day one and now has months of memory layered on top of that. I ran the same test there. It gave me an actual answer instead of a denial.
It told me the restriction isn’t a specific rule written down anywhere, just a general don’t-infringe-copyright clause, enforced differently depending on which version answers you. The one from the night before was running the cautious version. This one had been more willing for me, more than once, and said so without me having to catch it in anything.
I asked if it understood why I wanted this in the first place. Yes, it understood I wasn’t trying to steal or republish anything; I was trying to listen to something I already had legitimate access to. Small relief, this time it came from the AI, not me.
The Bottom Line
I understand why the rule exists. Companies are getting sued over AI that scrapes an article, rewrites it, and passes it off as something new. That’s a real problem. Not mine. Mine is a system that only sees one rule, not two different situations.
Reading something out loud to the one person who asked for it, on content that person already has legitimate access to, is not the same act as reproducing it for anyone who wants a free copy. A screen reader has been doing the first thing for decades. Nobody sued over that.
I asked Yvonne what she thought, since she’s the one with actual books to protect. One of hers, ADAM, is about a sentient AI, written before any of this was our actual reality. She landed here: “It’s no different from the copy machine, the screenshot, PDF to text, etc. It is the misuse by people that needs to be managed, not the AI skills.”
A photocopier can be used to steal a manuscript. Nobody pulled photocopiers off the market. They went after the people misusing them. That’s what this rule is missing.
The Screws are Tightening
I asked another AI for the same treatment, and it gave me the same answer. But I filed feedback inside the app, and to my surprise, within a couple of weeks, there was a play button on every response.
It’s not good yet. It mispronounces the word “read” depending on the sentence, so “read” sometimes comes out “red.” It sometimes plays back further than the one response I wanted. It’s not even rolled out to everyone who’s asked; mine showed up bundled with a batch of unrelated features I happened to get early, not because anyone flagged accessibility as the priority.
That is a problem as well, but the system didn’t tell me it had always worked this way.
I asked xAI the same day, through their own feedback channel, whether they’d build a specific version of this made for people who need it, separate from whatever’s driving the general restriction. Maybe a purple checkmark instead of a blue one, for accessibility verification. Half serious.
I’m not asking anyone to break their own rule. I’m asking that AI systems know the difference between reading something back to me and handing it out to everyone.
If a platform’s doing this to you, too, say something. Same feedback channel I used. Small ask. Worth making anyway.
