The Same Sentence Can Be Translated Differently
Translate the same sentence in a new chat, and the wording may change slightly each time. Is the AI doing that on purpose?
Instead of copying the original text, AI builds an answer by choosing what to say next based on what came before. So even when the meaning is similar, the wording or sentence order in a translation can vary. For important translations or other deliverables, compare the results and check them yourself rather than relying on the first answer.
AI Chooses What to Say Next
An LLM (a language model trained on a large amount of text) processes text as tokens (small units, such as words or parts of words). It does not produce an entire answer all at once. It looks at the text so far, chooses what to say next, and repeats that process to build the answer.
For example, after “kimchi for lunch,” “stew” might be more likely than “car.” The video uses “stew 65 percent,” “fried rice 25 percent,” and “car 0.1 percent” as illustrative examples. These are not values calculated by an actual model. When several choices seem plausible, the model may choose different words each time. Compare and check important results.
Models can also give different answers. When translation services were compared using the same instructions and data, the newest model did not always produce the best translation; sometimes an older model did better. The idea that older models may have accumulated more real user conversations over a longer time in service is the author’s guess from the video. Rather than choosing a model just because it is the newest, compare the results for the work you actually need to do.
AI Is Influenced by What It Has Seen Most Often
AI responses are strongly influenced by expressions and patterns it has encountered frequently. For a task like building a login screen, where there is a lot of publicly available material online, it may be easier for AI to produce a familiar answer. In fields with less public material, there may be fewer examples to draw on. In the past, someone asked for ideas for future videos based on comment analysis, and the answer was that it would be difficult. Later, AI became able to make those suggestions. The increase in available data and opportunities for trial and error contributed to that change.
A lack of public material does not mean AI cannot do the task at all. But its answers may be less reliable than in familiar areas, so check the results more carefully for less common tasks.
There Are Limits to What AI Can Refer to at Once—and Remember
The context window (the range of information AI can refer to at one time) has limits. As a conversation gets longer, earlier details may not be fully taken into account. Keep Claude Code’s project-rules file, CLAUDE.md (a file for recording work rules and why they matter), to 200 lines or fewer, and repeat important details when needed.
Training and use are different. After training, AI does not update the model itself every time it answers a question, and a new chat does not automatically carry over the previous conversation. Even when a feature can remember earlier conversations, the information still needs to be saved separately and passed into the new chat. To avoid repeating a mistake you fixed yesterday, record important rules in a file such as CLAUDE.md and provide it with the task.
AI may also be unfamiliar with libraries or features released after its training. A natural-sounding answer and up-to-date information are separate things. Check installation instructions and library usage against the official documentation and your actual work environment.
A Plausible Answer Can Still Be Wrong
Hallucination—the phenomenon of AI giving plausible answers about things that have not been verified—is a reminder that fluent writing does not guarantee accuracy. In the video, AI said an exposure count was 1,200, when the actual number was 253. In another case, a database structure included a rule that did not exist in the service. Check numbers and facts against the actual source, and make sure the answer follows your service’s requirements.
When AI handled search, the quality of its answers was inconsistent. In one case, search was handled separately and the results were then passed to AI. Splitting tasks that require verification, such as searching, from the task of answering based on those results can help distinguish what the tool should do from what AI should do.
AI chooses what to say next based on probabilities, is influenced by what it has seen often, and has limits on how much it can refer to at once. Keep important rules in files, and check results and facts yourself. AI can help with coding, but people still need to decide what to build and check whether the result solves the problem.
Have you had an experience where AI repeatedly made mistakes or gave different answers while you were working with Claude Code?
This article summarizes my Korean YouTube video (English subtitles available) and was translated with AI. #ABotWroteThis





Top comments (0)