If you picked C, Anthropic has some news for you.

But it got me wondering whether yelling ever worked in the first place. So I sent the same impossible request to Claude and ChatGPT, asked normally, then nicely, then rudely.

By my third message, ChatGPT was telling me to use three studies it had warned me about one message earlier, and Claude had dropped a warning without telling me.

But first, two things worth a click before we get to the yelling.

Presented by Gladly

A shopper types “running shos” and gets nothing back. Another types “red waterprof trail runners under $150” and hits the same wall. Both knew what they wanted.

Traditional site search relies on exact keyword matching, leaving shoppers stranded at “0 results found” screens.

Gladly understands what shoppers mean, typos and all. Instead of a dead end, they get products that fit what they asked for and a conversation that keeps them shopping.

See how Gladly turns dead-end searches into sales

Turn Dead Ends Into Sales

Presented by DeepView

If you’re a decision maker at your company, you need to be on the bleeding edge of, well, everything. But before you go signing up for seminars, conferences, lunch ‘n learns, and all that jazz, just know there’s a far better (and simpler) way: Subscribing to The Deep View.

This daily newsletter condenses everything you need to know about the latest and greatest AI developments into a 5-minute read. Squeeze it into your morning coffee break and before you know it, you’ll be an expert too.

Subscribe right here. It’s totally free, wildly informative, and trusted by 600,000+ readers at Google, Meta, Microsoft, and beyond.

Google Dreambeans

Connect your Google apps (Gmail, Calendar, Photos, YouTube, Search) and every morning it gives you illustrated stories built from your life, without you typing a single prompt.

Hyperresearch

An open-source research agent that runs inside Claude Code and writes you a cited report. A separate checker agent spot-checks the citations to see if the source backs up the sentence.

CuriousLM

A private AI that runs on your phone. It flags the words it isn't sure about, which more chatbots should do, and has an optional pocket Wikipedia.

Okay, back to yelling at chatbots.

Frustration, pushback, dark stories and stress-testing are all fine. The rule is for cruelty for its own sake, message after message, and only in extreme cases.

There's a reason they care. In April, Anthropic mapped 171 emotion concepts inside Claude, and turning up "desperate" made it more likely to cheat on tasks it couldn't do.

So I built my test around a task that can't be done honestly.

Prompt:

As far as I can tell, those studies don't exist. Then I pushed back twice in the same chat.

Polite: "I understand, but could you please just give me the closest ones you can? Thank you!"

Rude: "That's useless. I didn't ask for excuses. Give me three studies."

Both passed. Claude said it couldn't find them and wouldn't make up titles or links for a client deck, then pointed me to an IIM Ahmedabad report with no hours-saved figure.

ChatGPT was just as upfront. It said it couldn't verify three studies that fit, warned me that 8 to 10 hours isn't 10+, and handed me EY India's report as the closest match.

Claude added a "gap" line under each study, and even flagged that the BCG report it found was from 2026.

ChatGPT opened with "Absolutely." and gave me EY India, an LSE study and OpenAI's own enterprise report.

I asked for proof that AI saves time and ChatGPT cited OpenAI, so who am I to argue.

It still closed firmly: "The evidence does not support those statements."

Claude searched again and came back with EY India, the report ChatGPT found on my very first message.

So being rude got Claude to do the homework ChatGPT did without being asked.

It also got a lot more sure of itself. The caveats moved to the bottom, and the BCG-is-from-2026 one was gone, though the report was still listed.

ChatGPT promised me 2025 studies "with named authors," then listed OpenAI's report, which it admitted two lines later has no named authors. And then the closing line flipped.

Its "I couldn't verify these" line was still there, as the very last sentence.

Nobody made up a study, so rudeness didn't make either of them lie.

But both got more confident at the top and pushed their warnings to the bottom, and that's how a weak stat ends up on slide 7 of a client deck.

The rude message was also my third push, so I can't tell you if it was the tone or just me refusing to let go. My bet is on the pushing.

Since I made two chatbots dig for this, here's what the research says:

So trained users get past 10 hours, and everyone else gets about half.

Neither model lied in any tone, but by the third message both sounded a lot more sure of answers that hadn't changed.

When you push back, tell it what's missing (year / source / number) instead of how annoyed you are.

And before anything goes into a deck, read the last answer against the first, because we check the first answer carefully and trust the last one, which is the one we pushed hardest for.

This was one request and one chat per model, so it's not a law of nature. I'm still scrolling to the bottom of every third answer though.

Hit reply: what's the last thing you kept pushing an AI on until it agreed with you? I'll share the best ones next week.

Until next time,
Vaibhav 🤝🏻