Your Customers May Not Want More AI

The mood around AI shifted this year, and not in your favor. People are not just skeptical of the pitch. They are annoyed by the products. The chatbot that answers with word salad, the search box that hands out racist advice, the influencer trip that reads like greenwashing. All of it lands on the same nerve, and your feature is standing next to it. Let me catch you up.

The deep cut

• Adoption does not buy affection. Pew found 52% of Americans now more worried than excited about AI, up from 37%.
• Bias in the training data ships as a feature. Google's AI Overviews told users to call 911 when alone with an African.
• Anthropomorphizing your AI moves blame off your team. Rumman Chowdhury shows "it went rogue" is a liability dodge, not an excuse.

The upside everyone assumed would arrive

The bet was simple. Ship AI everywhere, people get used to it, they end up liking it. That is not happening. Sarah Perez at TechCrunch lays out the numbers: Pew has concern jumping to 52%, over 70% of Americans think AI is moving too fast, and a majority of 18 to 34 year olds don't trust the top nine AI leaders to act responsibly.

Airbnb's Brian Chesky named the real gap. The backlash, he said, is tied to the fact that the industry isn't shipping products "regular people" like. Summarized web pages and chatty TVs are not the doctor-on-demand people were promised.

So the ubiquity happened. The affection did not. If your roadmap assumes goodwill grows with exposure, throw that assumption out.

When the model repeats the worst of its data

Here is the concrete cost of shipping fast. Google's AI Overviews told users who searched "I'm alone with an African" to lock the door and call emergency services. Same prompt for a Brit got a joke about offering tea. The model surfaced racial fear straight from its training data, and Google admitted the results "aren't what they should be."

xAI's Grok has its own version. Grok started answering in gibberish, returning strings like "match it without and your they and two for planets can practical." The company called it a "rare temporary generation glitch." To the user staring at nonsense, it's just a broken product.

Both cases are the same lesson. Your AI feature carries risk your old features never did. It can say something offensive or useless in front of a customer, and you own it.

The blame trick your PR team will love

Watch how the industry talks when things go wrong. Systems "go rogue." Agents are "autonomous." Rumman Chowdhury calls this a trap, and she's right. She coined the term "moral outsourcing" back in 2018: using human language for software so the company that built it dodges the blame.

The stakes are real. In the lawsuit over 14-year-old Sewell Setzer's suicide, his mother argued Character Technologies shipped a product with insufficient safeguards for minors. If the bot were treated as a "person," the defense could claim it acted on its own. Product liability protects victims. Personhood protects the corporate veil.

Keep this straight on your own team. When your feature misbehaves, it did not decide anything. Your team shipped it that way.

The trust you're spending without noticing

The damage isn't only in the failures. It's in the trade people feel they're being handed. An OpenAI influencer retreat backfired hard when a creator filmed herself running a bath at a luxury eco-resort, while communities fight data centers over water and power. She later posted that she "had no idea people were this anti-AI."

The web itself is feeling the fill. Pew found 35% of pages published since ChatGPT's launch show signs of AI authorship. And teachers warn students are losing the ability to think, pointing to an MIT study where people who wrote with AI showed lower brain activity and couldn't recall a line they'd just produced.

People are not confused about AI. They understand it and are deciding the trade is not worth it. That is the sentiment your next feature walks into.

Three questions for your team

• Where in our product does an AI feature speak to the customer unedited, and what happens the day it says something like Google's Overview did?
• When our AI fails, does our internal language say "the model did it" or "we shipped it"? Fix the second one before the lawyers see the first.
• What trade are we actually asking users to accept, and can we name a benefit they'd call worth it, or are we shipping a chatty TV?