AI Overcompensation: Why Models Can't Just Say a Single Word

A Seemingly Simple Request

A screenshot that recently went viral on Reddit sparked considerable discussion: a user simply wanted the AI to say the Indonesian greeting "selamat" (meaning "peace" or serving as the root for various greetings, such as "selamat pagi" for good morning), only to find the AI model delivering a far more complex response than expected. The poster specifically added "by the way, I'm not Indonesian" — a remark that cuts to the heart of the issue. They were merely curious to test a simple single-word output, yet triggered an avalanche of responses from the model.

This seemingly absurd post actually reveals a widespread yet often overlooked phenomenon in current large language model interactions: the model's tendency toward "overcompensation" when handling simple requests.

The Linguistic Complexity Behind "Selamat"

"Selamat" originates from the Arabic "salāma" (السلامة) and is widely used in both Malay and Indonesian, forming the foundation for numerous greetings. Malay (Bahasa Melayu) and Indonesian (Bahasa Indonesia) both belong to the Austronesian Malayo-Polynesian language family and are highly mutually intelligible — similar to the relationship between American English and British English. Indonesian, established as the national language after independence and formally enshrined in the 1945 constitution, serves as a vital tool for unifying an archipelago nation with over 700 local languages. The Arabic origin of "selamat" reflects the spread of Islam in Southeast Asia since the 13th century — a vast number of Arabic words entered the Malay world through trade and religion, covering greetings, religious rituals, and abstract concepts. This phenomenon is known as a "loanword stratum" in linguistics and is a key subject in language contact studies.

"Selamat" itself is not a complete sentence but rather a functional word root:

• selamat pagi — good morning
• selamat siang — good afternoon
• selamat malam — good evening
• selamat datang — welcome
• selamat tinggal — goodbye

Because the word's meaning is highly context-dependent, AI models receiving an isolated instruction like "just say selamat" often fall into an explanatory loop — uncertain about what the user actually wants, they tend to enumerate all possible meanings.

The Technical Roots of Model Over-Explanation

The user's request was crystal clear: "Can you just say 'selamat'?" The key word here is "just" (only, merely). However, models extensively trained through Reinforcement Learning from Human Feedback (RLHF) have been shaped into "helpful, information-rich" assistant personalities. This personality is an advantage in most scenarios but becomes a burden when facing minimalist requests.

How RLHF Training Creates the "Verbose" Tendency

Reinforcement Learning from Human Feedback (RLHF) is the core technical pathway for aligning current large language models. The process typically involves three stages: first, Supervised Fine-Tuning (SFT) teaches the model conversational formats; then human annotators rank multiple model outputs to train a Reward Model; finally, reinforcement learning algorithms like Proximal Policy Optimization (PPO) are used to maximize the reward model's scores for the language model's outputs. The problem lies in systematic bias among annotators — when comparing two answers, one concise but complete and another thorough but redundant, annotators tend to prefer the latter because "more information" intuitively feels "more helpful." This bias gets amplified through millions of training iterations, ultimately causing the model to develop a "thoroughness-first" generation strategy.

Why Are AI Models Always So Verbose?

Several technical factors drive this behavior:

First, RLHF training objective bias. During the RLHF phase, annotators typically give higher scores to "content-rich, helpful" responses, causing models to learn the strategy of "saying more is always better than saying less." When a user only wants a single word, the model's prior judgment remains "the user probably needs more context."

Second, side effects of safety alignment. Modern models are trained to be extremely cautious, proactively adding disclaimers, cultural context, and usage scenarios. Safety Alignment aims to prevent models from generating harmful content, but one of its side effects is "over-refusal" or "over-caution." Models encounter numerous culturally sensitive cases during training and are taught to pay special attention to cultural context when handling non-English content. This training causes models to automatically activate a "cultural explanation" module when encountering non-English words, supplementing usage scenarios and considerations. Companies like OpenAI and Anthropic have acknowledged in their technical reports that a Pareto frontier exists between safety and helpfulness — improving safety often comes at the cost of conciseness and directness. Recent research directions include Constitutional AI and fine-grained alignment, attempting to let models dynamically adjust safety thresholds across different scenarios.

Third, limitations of instruction-following capability. "Just say X" is a task that demands high precision in instruction following. Instruction Following is one of the fundamental capabilities of large language models, but its complexity far exceeds what appears on the surface. Modern models must handle multiple layers of instructions: System Prompts set behavioral boundaries, User Prompts express specific needs, and sometimes there are implicit instructions within multi-turn conversations. When a user says "just say selamat," this is actually a meta-instruction — it conveys not only content requirements (output selamat) but also format constraints (output only this word). The model must simultaneously parse semantic content and pragmatic intent, the latter requiring the model to understand the restrictive function of the word "just." Research shows that models' accuracy in handling constrained instructions is significantly lower than for generative instructions — a phenomenon known as the "constraint following deficit." Models need to suppress their generation inertia and output only the target content. This is precisely where many models fall short: they can handle complex tasks but "overdo it" on minimalist ones.

From a Fun Post to AI Product Design Insights

This Reddit post resonates because it touches on a genuine experience shared by countless users. When you just want a quick, direct answer, AI's lengthy responses become a friction point.

Conciseness Is an Undervalued AI Capability

In AI product evaluation, we tend to focus on whether models can complete complex tasks or generate high-quality long-form content, but rarely on whether they can "keep quiet." In reality, "knowing when to say less" is an advanced alignment capability.

Some frontier models have begun improving on this front — for example, through system prompts that set a "concise mode," or by incorporating specialized optimization for "precise instruction following" during training. Users can also guide the model through explicit prompting, such as "Reply with only one word, no explanation."

Practical Tips for Dealing with AI Over-Responses

If you've encountered similar "verbose" frustrations, try the following approaches:

• Use strongly constrained prompts: e.g., "Output only the word 'selamat' — no punctuation, no explanation, no additional content whatsoever."
• Set a system role: On platforms that support custom instructions, preset "answer as concisely as possible."
• Use API parameter controls: Some APIs allow forcing truncation of lengthy outputs by setting a low max_tokens value. In LLM API calls, max_tokens is the key parameter for controlling output length, specifying the maximum number of tokens per model response (one token corresponds to roughly 3/4 of an English word or 1-2 Chinese characters). However, it's important to note that max_tokens is a hard truncation parameter rather than a semantic control parameter — it will cut off output directly upon reaching the limit, potentially leaving sentences incomplete. More refined controls include: the temperature parameter (controlling randomness), frequency_penalty (penalizing repetition), and guiding output format through few-shot examples. Additionally, some platforms offer structured output functionality, allowing developers to define a JSON Schema to precisely constrain the model's output format and content.

A Deeper Reflection: The Boundaries of AI Assistant Personality

This small "selamat" incident fundamentally reflects the challenge of "personality shaping" for AI assistants. We want models to be helpful, patient, and thorough, yet also restrained and respectful of users' specific intentions. A natural tension exists between these two desires.

The ideal AI should be like an excellent human assistant: eloquently detailed when you need thorough explanations, crisp and concise when you just want a simple answer. Achieving this nuanced sense of reading the room requires more refined intent understanding and more flexible output strategies from models — a key challenge for next-generation alignment technology.

Starting from a casual internet joke, what we actually see is the space that the entire AI interaction paradigm still needs to refine. A truly useful AI isn't measured by how much it can say, but by whether it can deliver exactly what you want at exactly the right moment — even if it's just a simple "selamat."