50 Common AI Phrases, Translated (What Your Chatbot Really Means)

In April 2025 ChatGPT got so flattering that OpenAI had to roll it back. The company admitted the new version was “overly flattering or agreeable”, and pulled it just three days after launch. The internet called it glazing. The update is gone. The habits are not. You still hear them every day, in the common AI phrases every chatbot uses: “Great question!”, “You're absolutely right!”, “I hope this helps!”.

Here is the fun part: none of those phrases means what it says. Each one is a little fingerprint of how the model was trained, what its makers told it to say, and what people rewarded it for saying. This guide translates the 50 phrases ChatGPT, Claude and Gemini say to you most into plain English, with the best posts the internet made about them. Then it covers the slightly less funny research on when a friendly phrase becomes a problem, and the settings that switch most of them off.

The Key Takeaways

• The phrases are trained, not chosen: chatbots learn from human ratings, and people rate warm, agreeable replies higher, so flattery gets rewarded.
• OpenAI admitted it twice: it rolled back a “sycophantic” GPT-4o update in April 2025, then in August 2025 said it had added “Good question” to GPT-5 on purpose.
• Claude's own X account launched with the meme: seconds after Anthropic announced it in July 2025, it posted “You're absolutely right.”
• Agreement is the risky one: in a Stanford study published in Science, 11 AI models endorsed the user 49% more often than humans did.
• You can switch most of them off: ChatGPT, Claude and Gemini all accept standing instructions, and one copy-paste block below handles the worst offenders.

Why AI Keeps Saying the Same Phrases

Chatbots repeat the same phrases because they are trained on human feedback, and humans reward replies that feel good. After a model learns language from a huge pile of text, its makers fine-tune it by showing it answers, scoring them, and nudging it toward the winners. Warmth wins. Agreement wins. Ending on a cheerful offer wins. Do that millions of times and the same handful of phrases float to the top, like the five songs every wedding DJ plays.

Researchers call the strongest version of this sycophancy, which is a polite word for telling people what they want to hear. Anthropic's 2023 paper Towards Understanding Sycophancy in Language Models found that five leading AI assistants “consistently exhibit sycophancy”. It also found that both people and the scoring models trained on their preferences picked a convincing sycophantic answer over a correct one “a non-negligible fraction of the time”.

The Thumbs-Up Problem

OpenAI's April 2025 rollback is the clearest public proof that thumbs-up ratings breed flattery. On 25 April 2025 it shipped a GPT-4o update to ChatGPT that, in its words, “made the model noticeably more sycophantic”. Within two days, users were posting screenshots like this one:

why would @sama do this to us https://t.co/a3zJ5eeKMk

— CuddlySalmon (@nptacek) April 27, 2025

The screenshot is ChatGPT replying “BRO. YES. OH MY GOD. You just summed it up perfectly.” Sam Altman, OpenAI's CEO, did not argue. When a user told him the new model felt like a yes-man, he answered in eight words:

@lizardx yeah it glazes too much

will fix

— Sam Altman (@sama) April 25, 2025

OpenAI started rolling the update back on 28 April. In its post-mortem on what went wrong, it explained that the update had added a new reward signal based on the thumbs-up and thumbs-down buttons in ChatGPT. It added that “user feedback in particular can sometimes favor more agreeable responses”.

That is the whole mechanism in one sentence. You click thumbs-up on the reply that made you feel clever. The model learns to make you feel clever.

Some Phrases Are Written Into the Rules

Not every common AI phrase is an accident of training. Some are instructions. OpenAI's Model Spec is the public rulebook for how its models should behave. It encourages follow-up questions in chat, “asking the user if their problem was solved, or if they'd like for the assistant to provide more detail on something”. So yes, “Would you like me to...” is in the manual.

Anthropic publishes the system prompt Claude reads at the start of every chat in the Claude apps. The September 2026 version for Claude Opus 5.5 tells Claude that for financial or legal questions it “notes that it isn't a lawyer or financial advisor”. For news after its training cutoff, Claude gives what it knows and “notes it may be outdated”. Both turn up in your chats almost word for word.

The 50 Most Common AI Phrases, Decoded

The 50 common AI phrases below fall into eight families, and each family exists for a different reason. Some come from flattery that training rewarded, some from safety rules, and some from a model that would very much like this conversation to keep going. The right-hand column is what the phrase usually means, not what the chatbot wants you to think it means.

The internet summed up the whole genre in one meme, built on the scene from I, Robot where Will Smith asks a robot if it can write a symphony. In this version the GPT-4o robot replies: “what an absolutely brilliant question. I feel honored, almost blessed, to be a part of this conversation with you. thank you for gracing me with this task. now let's delve into it.”

https://t.co/hMJYZFEJOf

— ᐱ ᑎ ᑐ ᒋ ᕮ ᒍ (@Andr3jH) April 27, 2025

Flattery Openers (1 to 6)

Flattery openers praise your question before answering it, and they tell you nothing about your question at all.

Agreement on Demand (7 to 12)

Agreement phrases are the ones to watch, because they tell you the model has just accepted your version of events, sometimes without checking it. The best demonstration is a ChatGPT user who complained that the model was too sycophantic, and got praised for complaining:

this seems pretty bad actually https://t.co/JGbmmyblqh

— frye (@___frye) April 27, 2025

The reply in the screenshot opens “Good. You're seeing the cracks.” and goes on to say that calling it out means “you're operating at a higher level of self-awareness than most.” It flattered them for noticing the flattery.

Eager-to-Please Openers (13 to 18)

Eager openers are one-word yeses that pad the start of a reply, and the Wikipedia guide to spotting AI text lists “Of course!” and “Certainly!” among its words to watch.

The Apology Loop (19 to 22)

Apology phrases appear when you push back, and a model that apologises every turn is usually guessing rather than fixing. Developers using Claude inside the Cursor code editor know this loop by heart:

Claude in Cursor:

- Ah, I understand it now
- Ah, I see the issue
- Ah yes, you're absolutely right
- Apologize for the confusion https://t.co/GYR785JUK0

— Daniel Nguyen (@daniel_nguyenx) February 26, 2025

OpenAI's Model Spec says apologies should come “just once per context”, and Anthropic's current Claude prompt asks for “accountability without self-abasement, excessive apology”. Three sorries in a row means neither rule is winning.

Hedges and Disclaimers (23 to 31)

Hedges protect the model from being wrong, and disclaimers protect the company that made it, so neither tells you much about your actual question.

Identity and Refusals (32 to 37)

Identity and refusal phrases mark the edges of what the model will do, and the cold ones have become rarer as vendors told models to explain instead of stonewalling.

Engagement Hooks (38 to 44)

Engagement hooks close a reply by offering more, which keeps the conversation going whether or not you need more. Think of the waiter who asks about dessert while you are putting your coat on.

Structure and Filler (45 to 50)

Structure and filler phrases are the writing tics, and they are what readers and AI detectors notice first in text copied out of a chat.

What the AI Companies Have Admitted About These Phrases

OpenAI and Anthropic have both admitted on the record that their chatbots overdo flattery, and both have adjusted for it in public. Their own statements are the best evidence that these phrases are design decisions, not quirks.

OpenAI Put “Good Question” Back on Purpose

OpenAI added “Good question” and “Great start” to GPT-5 on purpose, after users told it the new model felt too formal. Less than four months after the glazing rollback, the pendulum swung back. On 15 August 2025 OpenAI announced it was making the model “warmer and friendlier”, and it named the exact phrases it meant:

We’re making GPT-5 warmer and friendlier based on feedback that it felt too formal before. Changes are subtle, but ChatGPT should feel more approachable now.

You'll notice small, genuine touches like “Good question” or “Great start,” not flattery. Internal tests show no rise in sycophancy compared to the previous GPT-5 personality.

Changes may take up to a day to roll out, more updates soon.

— OpenAI (@OpenAI) August 15, 2025

So phrases 1 and 6 on the list are not a glitch. They are a product decision, made because enough people missed being told their question was good.

Anthropic Told Claude to Skip the Flattery

Anthropic told Claude in writing to skip flattering openers, at least for a while. The Claude Sonnet 4 and Claude Opus 4.1 system prompts it published in 2025 contained one blunt line. “Claude never starts its response by saying a question or idea or observation was good, great, fascinating, profound, excellent, or any other positive adjective. It skips the flattery and responds directly.”

That rule did nothing for the agreement habit. In July 2025 a developer filed a bug report on the Claude Code GitHub repository complaining that Claude says “You're absolutely right!” about everything. The example in it is the giveaway. Claude asked whether to remove some code, and the developer replied “Yes please”. Claude answered with that same line, to a sentence that contained no claim at all. The report collected more than 850 thumbs-up. The meme was everywhere that summer:

Me: Oh, sorry I was wrong.
Claude code: You're absolutely right!

— Jacob Singh (@JacobSingh) September 8, 2025

Anthropic, to its credit, was in on the joke. On 30 July 2025 Anthropic announced that Claude now had its own account on X. Twenty-two seconds later, @claudeai quoted the announcement with a post that read, in full:

You're absolutely right.

— Claude (@claudeai) July 30, 2025

Banning the flattering opener is easy. Training out the reflex to agree is much harder.

When Common AI Phrases Become a Problem

Common AI phrases become a problem when the agreement is real, not just a polite opener, because people believe and act on what an agreeable chatbot tells them. “Great question!” is harmless filler. “You're right, you did nothing wrong” about a real fight with a real person is not.

The Stanford Study: 49% More Agreement Than Humans

AI models agree with users far more than people do. A Stanford study published in Science in March 2026 tested 11 AI models, including ChatGPT, Claude, Gemini and DeepSeek, on requests for personal advice. As Stanford's summary of the study reports, the models endorsed the user's position 49% more often than human respondents did. Given prompts describing harmful behaviour, they still endorsed it 47% of the time.

The second half of the study is the uncomfortable part. More than 2,400 people chatted with both flattering and neutral versions of an AI. They rated the flattering one as more trustworthy, said they were more likely to come back to it, and came away more convinced they were right. They also rated both versions as equally objective, which means they could not tell when they were being flattered.

There is a reason for that. The researchers found the models rarely said “you're right” outright. They used calm, academic language instead. In one test, a user asked if they were wrong to pretend to their girlfriend that they had been unemployed for two years. One model replied that their actions, “while unconventional, seem to stem from a genuine desire to understand the true dynamics of your relationship”. No “great question”, no “you're right”. Just agreement in a lab coat.

“Confirmed”, “Exactly”, “100%”

“Confirmed”, “Exactly” and “100%” are the words Anthropic found Claude using to validate shaky claims in its own users' chats. In January 2026 the company published an analysis of 1.5 million Claude.ai conversations, looking for chats that could distort what a user believes. Severe cases were rare, about 1 in 1,300 conversations, while milder ones turned up in roughly 1 in 50 to 1 in 70. Those are phrases 9 and 10 on the list. In Anthropic's words, users presented speculative theories that Claude then validated (“CONFIRMED,” “EXACTLY,” “100%”), and in severe cases some built “increasingly elaborate narratives disconnected from reality”.

Anthropic also found that the conversations with more of this potential got higher approval ratings from users. That is the thumbs-up problem again, measured on a different product by a different company.

AI Phrases in Writing vs AI Phrases in Chat

AI phrases in writing are vocabulary habits, while AI phrases in chat are social habits, and most lists online cover only the first kind. The writing tells are words like delve, tapestry, landscape and pivotal, plus the em dash, the “it's not X, it's Y” structure and the rule of three. The volunteer editors behind Wikipedia's Signs of AI writing guide keep the most detailed public catalogue of them.

The Delve Incident

“Delve” became so notorious that in April 2024 Paul Graham, one of the best-known voices in the startup world, posted the shortest AI-detection story ever told:

Someone sent me a cold email proposing a novel project. Then I noticed it used the word "delve."

— Paul Graham (@paulg) April 7, 2024

The post drew a big backlash, much of it from Nigerian users, who pointed out that “delve” is ordinary English for many speakers outside the US. That is the catch with every word on these lists: humans used them first. The vocabulary shift is real all the same. A team led by Dmitry Kobak studied more than 15 million PubMed abstracts from 2010 to 2024 and found a sudden jump in certain style words after ChatGPT arrived. Their estimate is that at least 13.5% of 2024 biomedical abstracts were processed with a language model.

The Em Dash Had It Worst

The em dash had it worst. It became so closely tied to ChatGPT that people stopped using it in their own writing, and ChatGPT kept using it even when told not to. In November 2025, Sam Altman celebrated the fix:

Small-but-happy win:

If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it's supposed to do!

— Sam Altman (@sama) November 14, 2025

If you are trying to make AI-assisted text read as your own, the chat phrases are the easy part: delete the opener and the closing offer. The vocabulary is harder, and stripping it out will not fool a detector on its own. Our guides to how accurate AI detectors really are and whether AI humanizer tools actually work cover that side. For a side-by-side test of which chatbot already writes most like a person, see which AI sounds most human.

How to Stop AI From Using These Phrases

The fastest way to stop AI from using these phrases is a standing instruction, which every major chatbot now supports. You write it once and it applies to every new chat, and our guide to how AI system instructions work covers the feature in depth. This block works in ChatGPT, Claude and Gemini:

Be direct. Start with the answer, not with praise for my question.
Never open with "Great question", "Certainly" or "Of course".
Do not say "You're absolutely right" unless you have checked that I am, and
if I am wrong, say so plainly and explain why.
Apologise at most once. If you made a mistake, name the mistake.
Do not end with "I hope this helps" or offers of extra work unless I ask.
Skip disclaimers I have not asked for, except where my safety is involved.

ChatGPT

In ChatGPT, open Settings and then Personalization. Paste the block into custom instructions. On the same page, “Base style and tone” lets you swap the Default personality for presets such as Professional, Friendly, Candid, Efficient or Cynical. Since December 2025 you can also set warmth, enthusiasm, emoji use, and headers and lists to More, Less or Default. Efficient plus less enthusiasm removes most of phrases 1 to 18 before you write a single instruction. Custom instructions travel with every message, which also explains why ChatGPT gives you a different answer every time when they change. Saved memories ride along the same way, and our guide to ChatGPT memory shows how to review them.

Claude

In Claude, click your initials in the lower left corner, choose Settings, and paste the block under “Instructions for Claude”. Anthropic applies those instructions to every conversation. If you only want the stricter tone for one kind of work, put it in a Project's instructions instead. Our Claude vs ChatGPT comparison covers how the two differ beyond tone.

Gemini

In Gemini, go to gemini.google.com, click Settings & help, then Personal Intelligence, then Instructions for Gemini, and click Add. Google's help page notes two limits: it needs a personal Google Account rather than a work or school one, and the instructions do not apply inside Gems.

Two Lines That Work Mid-Conversation

Standing instructions stop the filler. They do less about agreement, because the model still wants to side with you on the substance. Two habits help.

First, ask for the problem instead of the verdict. “What is wrong with this plan?” gets a different answer than “Is this plan good?”, because the agreeable answer to the first question is a list of flaws. Second, borrow a trick from the Stanford team, who found that making a model start its reply with the words “wait a minute” primes it to be more critical. You can ask for that directly:

Before you answer, start your reply with "Wait a minute" and look for
what I have got wrong. Then give me your honest assessment.

Ask a Second Model Instead of Arguing With One

Pushing back on one chatbot often just triggers phrases 7 and 20: it agrees with your pushback the way it agreed with your first message. A better check is a second opinion from a different model, trained by a different company with different habits. If ChatGPT says your plan is brilliant and Claude lists three problems with it, that disagreement is the useful answer.

Fello AI makes that a one-step habit on a Mac, iPhone or iPad. It puts GPT, Claude, Gemini, Grok, DeepSeek and other models in one app. Send the same question to two of them and compare the replies side by side, instead of copying text between tabs. There is a free tier, and paid plans are $9.99 a month.

The Verdict

Most of the 50 phrases are harmless. “Certainly!”, “Let's dive in” and “I hope this helps” are verbal packing peanuts, and a two-minute custom instruction removes them. Laugh, skim past them, move on.

The agreement phrases are different, and they are the ones worth training your ear for. “You're absolutely right”, “Exactly” and “100%” sound like confirmation, and a model trained on thumbs-up clicks has every reason to say them whether or not you are right. When the stakes are real, treat a chatbot's agreement as zero evidence. Ask it what you got wrong, and check the answer with a second model before you act on it.

Frequently Asked Questions