နားလည်ထားရမယ့် အချက်
AI-assisted research က ordinary pipeline အတိုင်းပဲ လိုက်နာပြီး AI ကို step နှစ်ခုတည်းမှာသာ ထည့်သုံးပါတယ် - angle generation နဲ့ comparison။ search, evaluation, synthesis ကို ဘယ်တော့မှ AI က အစားထိုးမပေးပါဘူး။
1. Question
Real research question တစ်ခု သတ်မှတ်ပါ။
2. AI: Search Angle
AI tool ကို angle plausible list တောင်းပါ - findings မဟုတ်ဘဲ search plan ပါ။
3. Real Source များ Search
Angle တစ်ခုစီအတွက် real source တွေကို တကယ် search ပါ။
4. ဖတ်ပြီး Note ရေး
Source တစ်ခုစီကို ကိုယ်တိုင် ဖတ်ပြီး structured note ရေးပါ။
5. AI: Compare
Note များကို AI နဲ့ ယှဉ်ပြီး pattern (agree/contradict) ရှာပါ။
6. Human Verification
Citation တိုင်းနှင့် claim တိုင်းကို လက်ဖြင့် ပြန်စစ်ပါ - ဒီ step ကို ကျော်လို့ မရပါဘူး။
7. Output
Verify ပြီးမှသာ output ကို ထုတ်ပါ။
AI-ASSISTED RESEARCH FLOW
-------------------------
AI-ASSISTED RESEARCH FLOW
--------------------------
[question] -> [AI: search angles] -> [search real sources]
|
v
[read + take notes]
|
v
[AI: compare notes]
|
v
==[ HUMAN VERIFICATION ]==
|
v
[output]လက်တွေ့ scenario နဲ့ ချိတ်ကြည့်မယ်
Research session ကို AI ဆီက angle ၅-၆ ခု တောင်းတာနဲ့ စတင်ပါ - အဖြေ မတောင်းပါနဲ့။ Angle တစ်ခုစီကို real search ထဲ ထည့်ပါ။
Source တိုင်းကို note ထဲ မဝင်ခင် ကိုယ်တိုင် ဖတ်ပါ။ AI summary တစ်ခုတည်းကို trust မလုပ်ပါနဲ့ - conclusion ပြောင်းစေမယ့် caveat ကို ချန်ထားနိုင်ပါတယ်။
- Structured note ရလာတာနဲ့ AI ကို claim compare ခိုင်းပါ
- Output ကို draft comparison အဖြစ်သာ သတ်မှတ်ပါ - final judgment မဟုတ်ပါ
- Citation တိုင်းကို လက်ဖြင့် ဖွင့်ပြီး verify ပါ
Verify မလုပ်ရင်
Real-looking DOI/URL ရှိတဲ့ citation တောင် ဖွင့်မကြည့်ဘဲ ယုံရင် fabricated source တစ်ခု output ထဲ ရောက်နိုင်ပါတယ်။
AI-generated citation ကို Blindly မယုံပါနဲ့
AI tool တစ်ခုက format မှန်၊ ကြည့်ကောင်းတဲ့ citation ထုတ်ပေးနိုင်ပေမယ့် လုံးဝ fabricated ဖြစ်နိုင်ပါတယ်။ Output ထဲ မထည့်ခင် citation တိုင်းကို ဖွင့်ပြီး source ရှိမရှိ၊ claim နဲ့ ကိုက်ညီလား လက်ဖြင့် စစ်ပါ။
အတူတူ စမ်းရေးကြည့်မယ်
function flagCitations(citations) {
const doiPattern = /^10\.\d{4,9}\/\S+$/;
const urlPattern = /^https?:\/\/[^\s]+\.[a-z]{2,}(\/[^\s]*)?$/i;
const placeholderWords = ["example", "citation needed", "source name", "journal of somewhere", "n.d."];
return citations.map((c) => {
const text = c.trim();
const lower = text.toLowerCase();
const looksLikeDoi = doiPattern.test(text);
const looksLikeUrl = urlPattern.test(text);
const looksFabricated = placeholderWords.some((w) => lower.includes(w));
let status;
if (looksFabricated) status = "needs_verification"; // placeholder-shaped, likely invented
else if (looksLikeDoi || looksLikeUrl) status = "check_it_resolves"; // real shape, still confirm it loads and says this
else status = "needs_verification"; // no verifiable identifier at all
return { citation: text, status };
});
}
const aiCitations = [
"10.1145/3442188.3445922",
"https://arxiv.org/abs/2005.11401",
"Example et al. (2024), Journal of Somewhere, p. 42",
"Smith, J. (n.d.). Citation needed for this claim.",
"https://openai.com/research/gpt-4",
];
for (const r of flagCitations(aiCitations)) {
console.log(r.status.padEnd(20), "-", r.citation);
}check_it_resolves - 10.1145/3442188.3445922
check_it_resolves - https://arxiv.org/abs/2005.11401
needs_verification - Example et al. (2024), Journal of Somewhere, p. 42
needs_verification - Smith, J. (n.d.). Citation needed for this claim.
check_it_resolves - https://openai.com/research/gpt-4
"check_it_resolves" ဆိုတာတောင် fully verified လို့ အဓိပ္ပာယ် မဟုတ်ပါဘူး - shape မှန်ရုံသာ ဖြစ်ပြီး ဖွင့်ပြီး ဖတ်ရဦးမယ်။၅ မိနစ် စမ်းကြည့်
AI tool တစ်ခုကို research question တစ်ခုအတွက် citation ၅ ခု ထုတ်ခိုင်းပါ။ flagCitations() ထဲ ထည့်ပြီး "check_it_resolves" ရတဲ့ citation တစ်ခုစီကို တကယ် ဖွင့်ပြီး claim နဲ့ ကိုက်ညီလား စစ်ပါ။
သတိလေးတစ်ချက်
AI ရဲ့ angle list ကို finding အဖြစ် ထင်မှတ်ပြီး search မလုပ်ဘဲ ရေးတာ
Citation format မှန်ရုံနဲ့ verified ပြီးသားလို့ ယူဆတာ
Wikipedia: Hallucination (artificial intelligence) — Productivity Systems