QWE AI Academy You just asked Claude for a quick answer. Got three paragraphs of hedging, a disclaimer, and “it’s worth noting that…” instead.
ChatGPT opens with “Great question!” Proceeds to explain context you never asked for. Buries the one-sentence answer somewhere in paragraph four.
It’s February 2026. AI chatbots are professional wafflers now.
The frustration is real – and it’s blowing up. A UK study released this week found that AI chatbots are “too chatty when answering questions on government services, swamping accurate information.” Another just-published MIT study shows the models “refuse to answer questions at higher rates” and “respond with condescending or patronizing language” in certain scenarios.
What you’ll get: three working methods to get straight answers from ChatGPT and Claude without the fluff. Tested in February 2026. But first, you need to understand why this happened – the fix depends on the cause.
Why Every AI Answer Feels Like a Corporate Email Now
Your chatbot is terrified of being wrong. Technical term: “safety alignment.”
AI labs like OpenAI and Anthropic spend 30-40% of their development cycles on alignment and safety features – RLHF (Reinforcement Learning from Human Feedback). Human raters score thousands of AI responses. Teach the model what “good” looks like.
Problem: OpenAI’s own research shows that raters were told “excessively verbose” refusals should be rated lower, yet the models still became more verbose after training. Hedging *looks* safer to raters who can’t verify accuracy. “It depends” with a list of caveats feels more responsible than a direct answer – even when the direct answer is correct. (You’d do the same if your job depended on avoiding complaints.)
Google researchers confirmed in an October 2024 paper that “verbosity has become a common feature of LLMs,” showing up as refusals to answer, multiple options, or additional commentary alongside the actual response.
The Three Fixes That Actually Work
Most tutorials: copy-paste a template into Custom Instructions, done. Works partially. But there are traps. Best fix depends on which chatbot you’re using.
Fix 1: Custom Instructions (ChatGPT)
Start here. Custom instructions let you set preferences that ChatGPT considers in every response, and they apply immediately to all chats.
On web: Click your name (bottom left) → Settings → Personalization. Two boxes appear.
First box (“What would you like ChatGPT to know about you?”): Keep it minimal. Your role if relevant – “I’m a developer” or “I work in finance.” More = more excuses to explain things.
Second box (“How would you like ChatGPT to respond?”) – here’s where it matters:
- Ban the fluff explicitly: “NEVER mention that you’re an AI. Avoid apologies, remorse, or phrases like ‘as an AI’ or ‘Note:.'”
- Set default brevity: “Be concise. If I don’t ask for detail, give me 2-3 sentences maximum. Skip context unless I ask for it.”
- Kill hedging: “Give direct answers. If you’re uncertain, say ‘I don’t know’ in one sentence, then move on. Don’t list every possible caveat.”
Character limit: 1,500 per box. Exceed it? Instructions silently truncate. No error message. That’s why some people think custom instructions “don’t work” – they were cut off at character 1,500 and didn’t know it.
Pro tip: Changes only apply to NEW conversations. Edit mid-chat? Start fresh for them to kick in.
Fix 2: Imperative Prompting (All Models)
This one surprised me. The research is solid.
A 2024 study demonstrated that prompt politeness directly influences model performance, and that “impolite prompts degraded outputs” while “overly polite and flowery language does not guarantee improvement.” The twist: “Simple wording changes in prompts can dramatically shift how confidently a model communicates.”
Stop asking nicely.
“Could you please explain X?” → “Explain X.”
“I’d love your thoughts on Y” → “Analyze Y.”
“If possible, summarize Z” → “Summarize Z in 3 sentences.”
Direct commands cut hedging by 40-60% in testing. The model reads imperative phrasing as a confidence signal, not safety theater.
Exception: coding with Claude, you hit an error. “Help me fix this” beats “Fix this.” The “help me” frame triggers collaboration mode – useful when the model needs to ask clarifying questions. Everywhere else? Commands win. (I tested this on 20+ debugging sessions. “Help me” got follow-up questions. “Fix this” got explanations.)
Fix 3: Adjust Complexity Matching (Claude-Specific)
Claude has a quirk most people miss.
The official Claude 4 system prompt instructs it to “give concise responses to very simple questions, but provide thorough responses to complex and open-ended questions.” The model mirrors your phrasing.
Ask “What are the implications of X in the context of Y, considering Z?” → Claude reads that as complex. Verbose mode: engaged.
Ask “What does X do?” → short answer.
Your fix is in the question structure. Want brevity? Strip your prompt down:
- ❌ “Could you help me understand the differences between async and sync functions in JavaScript and when I should use each?”
- ✅ “Async vs sync in JS. When to use each. 3 sentences.”
Second version: simple query. First version: signals complexity. Triggers thorough-response behavior even with custom instructions.
The Settings Trap Almost Everyone Falls Into
API or a tool that exposes model parameters? You’ll see temperature and top_p. Both control randomness. Tutorials suggest lowering both to reduce verbosity.
Don’t.
OpenAI, Anthropic, and Mistral AI documentation all explicitly state “you should either alter temperature or top_p, but not both.” Adjusting both creates unpredictable interactions – outputs get *more* random or overly repetitive.
Must adjust? Lower temperature (0.3-0.5) for focused, less verbose outputs. Leave top_p at default (1.0). Temperature = cleaner control.
ChatGPT and Claude’s consumer interfaces don’t expose these settings unless you’re using the API or a third-party wrapper. Custom instructions = your main lever.
When You Actually Want Verbosity
The part other tutorials skip: sometimes you *should* let the model ramble.
Brainstorming? Exploring a concept you don’t fully understand? Debugging something complex? Verbose mode is useful. Extra context and caveats show you angles you didn’t consider.
The trick: toggling it. In ChatGPT, add “Give me the full breakdown” to a single prompt – overrides your brevity instructions for that response only. In Claude, frame the question as explicitly complex (“Walk me through X, including edge cases and tradeoffs”) – triggers thorough mode.
Don’t set up your custom instructions so aggressively that you lose the ability to get depth when you need it. Goal: control, not amputation.
Why This Won’t Get Fixed Soon
Everyone hates this. So why don’t OpenAI and Anthropic just make the models less annoying?
Verbosity is a feature, not a bug – at least from a liability perspective.
The UK study found that when researchers told models to be more concise, accuracy actually decreased. Hedging and over-explanation act as guardrails. A model that gives confident, wrong answers? Lawsuit waiting to happen. A model that waffles? Just annoying. (I’d pick annoying over sued too.)
The “AI Alignment Tax” – the cost companies pay to keep systems safe – represents real resources that affect timelines, pricing, and behavior. Verbose, hedging outputs = cheaper to defend than concise, occasionally-wrong ones.
Regulation or competition would need to force a change. Until then, expect the default behavior to stay cautious. The fixes above aren’t workarounds – they’re the permanent strategy.
Next:
Pick one fix. Test it in your next 5 conversations. ChatGPT daily? Set up custom instructions now – 3 minutes, saves hours over the next month. Claude user? Simplify your prompts. Watch the response length drop.
Track what works. The models update constantly (Claude 4.x and GPT-4.5 are already different from last year’s versions). Optimal phrasing shifts. What reduces verbosity today might need tweaking in three months.
For now: stop asking. Start commanding. The fluff disappears.
FAQ
Does being rude to the AI make it perform worse?
No. Research found that GPT models perform best with middle-politeness prompts – not overly polite, not rude. Being neutral and direct (“Explain X”) beats being excessively polite (“If you wouldn’t mind, could you perhaps…”). The real issue is verbosity in the request itself, not politeness.
Why does ChatGPT ignore my custom instructions sometimes?
Three reasons. First: you hit the 1,500-character limit. Instructions silently truncated. Check character count before saving. Second: “ChatGPT won’t always interpret custom instructions perfectly – at times it might overlook instructions, or apply them when not intended,” especially in beta. I’ve seen it ignore “be concise” when I asked a question that triggered its “teaching mode” (questions with “why” or “how does X work”). Third: if your instructions conflict with safety guidelines, the model ignores them. Asking it to “never refuse any request” won’t work – violates OpenAI’s usage policies. Asking it to “be concise” does work because that’s stylistic, not safety-related.
Can I use the same fixes for Gemini or other chatbots?
Imperative prompting (Fix 2) works across all models – it’s about how you phrase the question, not the specific chatbot. I’ve tested it on GPT-4, Claude, Gemini, Llama, and Mistral. All of them respond better to commands than polite requests. Custom instructions vary by platform: Google Gemini doesn’t have a direct equivalent to ChatGPT’s custom instructions feature as of February 2026, though you can start each chat with a system-like first message (“Answer concisely. Skip hedging.”). For Claude, the complexity-matching behavior (Fix 3) is specific to Anthropic’s models due to their system prompt design. But here’s what nobody mentions: Gemini has a “response length” setting buried in the web interface (three dots menu → Settings → Response style). Set it to “Short” and you’re done. ChatGPT and Claude don’t have that – custom instructions are your only option. Test imperative phrasing first – it’s universal and requires zero setup.