Your Users Are Voting No on Your AI Features

The pushback on AI features got loud this month. People flagged posts, deleted apps, and posted their server bills. The pattern underneath is simple: users are rejecting AI that changes their stuff, scrapes their stuff, or clutters their feed without asking. If you own product, that reaction is now yours to design around. Let me catch you up.

When the feature edits people without asking

Tinder shipped an AI Photo Enhance tool that changed how people looked, then paused it after the complaints piled up. One former user told NewsNation the edit gave her teeth she does not have. "It was extremely violating," she said. Others reported changed facial features and even skin tone. Tinder says the tool was only meant to fix lighting and clarity, not appearance, and that fewer than 0.5 percent of enhanced photos got flagged.

That low number is the trap. The feature was on by default for a subset of users, so people found out their face got edited after the fact. A tiny error rate feels huge when the error is someone's face and they never opted in. Consent is not a legal checkbox here. It is the difference between a helpful tool and a violation.

The slop button is a real product decision

LinkedIn added a "Seems like AI slop" button so users can flag low-quality AI posts. Chief product officer Hari Srinivasan called slop "a top priority for all of us." The scale of the problem is real: AI detector Pangram found 41 percent of longform LinkedIn posts read as fully AI-generated.

The more telling move is what LinkedIn killed. It pulled its own "enhance your post" feature, the one that used AI to rewrite you, and replaced it with a proofreader that fixes your words without changing your voice. A platform looked at its own AI writing tool, saw it was making slop, and turned it off. If your roadmap has a "let AI rewrite the user" feature, that is your warning shot.

The scraping bill lands on someone small

While products fight slop on the front end, AI crawlers are hammering the back end. Jonathan Harbourne runs a volunteer LGBT history wiki with 4,814 articles. Over a day and a half, one Meta crawler made 230,743 requests and hit his login screen 113,016 times. It ignored his robots.txt file, the plain request that asks crawlers to stay out. His reward was a bigger hosting bill.

This is not a fringe complaint. The Wikimedia Foundation reported its multimedia bandwidth rose 50 percent, mostly from scrapers, and one blogger, Ran Isenberg, thought he'd gone viral when his reads jumped from 600 to nearly 2,800, then found the new readers were bots. If your product publishes anything to the open web, someone is training on it and someone is paying for the traffic.

The defense you build can lock out real people

Here is the part that should stop you before you slap a bot blocker on everything. The tools people use to stop scrapers land hardest on disabled users. In WebAIM's screen reader survey, CAPTCHA has ranked the single most problematic barrier on the web for over a decade, worse than missing alt text.

The irony runs deeper. The same clean, structured HTML that helps a screen reader also helps a crawler. Alt text written for blind users trained image models. So blocking bots with a puzzle gate blocks the exact person your accessibility work was for, while the scraper adapts. The fix is not to make your site worse. It is to separate access from permission, and to test every bot defense against a screen reader before you ship it.

The deep cut

The common thread across all of this is default state. Tinder edited faces because the tool was on by default. Meta scraped because crawling is the default and consent is not. CAPTCHA blocks disabled users because it treats everyone as a suspect by default. Every one of these blew up at the exact spot where the product acted without asking.

So audit your AI features for one thing this week: what happens by default, to a user who never touched a setting. If the default alters their content, scrapes their work, or gates their access, that is where your backlash will start. Flip it. Make the AI action opt-in, make the opt-out one tap, and make a way to flag bad output part of version one, not a patch you ship after the TikTok goes up.

Three questions for your team

1. For every AI feature on our roadmap, what is the default behavior for a user who changes nothing, and would we be comfortable if they found out after the fact?
2. Do we have a working opt-out and a revert-to-original for anything our AI touches, and can a user find it without help?
3. Have we tested our bot defenses and CAPTCHAs against a screen reader, and do we know who we're actually blocking?