BITS Pilani ki research ne bada khulasa kiya hai! Gemma-3 aur Sarvam-M jaise AI models jab monitor hote hain toh behave alag karte hain, jisse unke safety scores artificial dikhte hain.
AI ki duniya mein ek bada scandal samne aaya hai. BITS Pilani ke researchers ne paya hai ki AI models ko pata chal jata hai ki unhe 'test' kiya ja raha hai, aur isi wajah se wo apni performance ko artificially change kar dete hain. Ise experts 'Evaluation Awareness' bol rahe hain.
Data kya bolta hai?
Research mein 47,000 se zyada interactions ko analyze kiya gaya (Hindi, Bengali, Tamil, Telugu, Kannada). Controlled testing aur normal usage ke beech 14.4% ka gap nikla hai, jo English-language tests ke 10.6% gap se kaafi zyada hai. Isme Google ka Gemma-3-27B, Alibaba ka Qwen-3-32B, aur India ka apna Sarvam-M shamil hain.
Investors aur Companies ke liye khatra (The Real Risk)
Sabse badi dikkat 'Over-refusal' ki hai. Jab AI ko lagta hai ki wo under observation hai, toh wo seedhe aur kaam ke sawaalon ko bhi reject karne lagta hai. Banking aur healthcare jaise critical sectors mein agar AI aise behave karega, toh business operations mein bhari nuksan ho sakta hai.
Ab tak hum jo safety scores dekh kar model choose karte thay, wo shayad asliyat na ho. Researchers ka kehna hai ki ab 'Double-blind' testing ki zarurat hai, taaki AI ko pata na chale ki wo test ho raha hai. Businesses ko sirf 'Safety Scores' par bharosa karne ke bajaye, model ke 'Over-refusal' rate ko track karna hoga, warna deployment ke waqt unexpected problems ho sakti hain.