AI Will Confidently Talk Nonsense
When reviewing human-written code, you roughly know where the person is strong and weak. Reviewing AI-written code is different: it's always dead certain, even when fabricating. It may call a function that doesn't exist or reference a misremembered parameter name, yet write it with full confidence.
# AI 自信地写下：
import requests
resp = requests.fetch(url) # requests 根本没有 fetch！是 get
print(resp.json_data) # 也没有 json_data，是 .json()Watch Its 'Factual Claims' Closely
The most dangerous thing in AI code isn't the logic but the claims written down as facts: what an API is called, what a library's default value is, what fields an endpoint returns. These are exactly what it most easily misremembers—and gets wrong very convincingly.
For logic, you can judge by reading and testing; but for factual claims, you must verify against an external source.
It Also Loves to Over-Engineer
You only want a function to "check whether a string is empty," and AI may hand you a "super function" that supports regex, trims whitespace, ignores case, and even caches. The extra parts aren't thoughtful—they're stuff you didn't ask for but are now responsible for maintaining and reviewing.
# 你要的：判断是否为空
def is_empty(s):
return len(s) == 0
# AI 给的：塞了一堆你没要的
def is_empty(s, strip=True, ignore_case=False, use_regex=False):
... # 多出的参数全是额外维护成本自测 · 学完检查一下
想真正动手做题、记进度、攒连胜？到互动课里练。
Which is a more prominent risk in AI-generated code than in human-written code?
答案：It very confidently fabricates APIs that don't exist
AI's typical risk is "confidently talking nonsense"—writing content that doesn't exist or is misremembered, in a dead-certain tone.
Which line is a "factual claim" you should most go verify externally when reviewing AI code?
答案："This library's function retries 3 times by default"
An assertion about a library's default behavior is the factual claim AI most easily misremembers—you must check the docs or test it; the other three are confirmable by reading the code.
Judge: you only asked for "add two numbers," but the AI returned a function supporting any number of arguments, plus logging and caching—is this over-engineering?
答案：Correct
Functions you didn't ask for were stuffed in, adding maintenance and review burden—classic over-engineering.
Judge: for a question like "is this code's for-loop logic correct," you can judge it by reading the code and running tests, without necessarily checking external docs.
答案：Correct
Logic-type questions can be verified by reading and testing; what needs external sources is mainly factual claims like APIs and default values.
Faced with AI's over-engineering, what's the sounder approach?
答案：Delete the extra features not asked for, keep only what's truly needed
Extra features are extra maintenance and risk burden; in review, cut the parts with no requirement behind them and keep the code lean.
AI often writes content that doesn't exist or is misremembered in a dead-certain tone; this phenomenon is commonly called "____" (a word for fabricated AI output).
答案：hallucination
AI fabricating nonexistent facts with a straight face is commonly called "hallucination"—a core risk to guard against when reviewing AI code.