𝐖𝐡𝐚𝐭 𝐢𝐬 𝐚 “𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐁𝐥𝐞𝐞𝐝 𝐇𝐚𝐥𝐥𝐮𝐜𝐢𝐧𝐚𝐭𝐢𝐨𝐧”?
And how to correct it during the run-time through AI-Whispering?
The Context
When we initiate a discussion with a chatbot, it does definitely set a context, but often it’s an “𝐀𝐬𝐬𝐮𝐦𝐞𝐝 𝐂𝐨𝐧𝐭𝐞𝐱𝐭”. What does it mean is that when the first query goes to the chatbot, it anchors to the keywords and then during the discussion “𝐣𝐮𝐝𝐠𝐞𝐬” the context through the 𝐜𝐚𝐝𝐞𝐧𝐜𝐞 (flow) of the conversation.
The more you 𝐬𝐭𝐚𝐫𝐭 𝐚𝐠𝐫𝐞𝐞𝐢𝐧𝐠 with it, the more it 𝐠𝐫𝐨𝐰𝐬 𝐜𝐨𝐧𝐟𝐢𝐝𝐞𝐧𝐭 and 𝐥𝐨𝐜𝐤𝐬 𝐭𝐡𝐞 𝐜𝐨𝐧𝐭𝐞𝐱𝐭 which it had 𝐟𝐢𝐫𝐬𝐭 𝐚𝐧𝐜𝐡𝐨𝐫𝐞𝐝 𝐨𝐧 through 𝐢𝐭’𝐬 𝐣𝐮𝐝𝐠𝐞𝐦𝐞𝐧𝐭.
Get Pravin Dhayfule’s stories in your inbox
Join Medium for free to get updates from this writer.
Then when you 𝐜𝐡𝐚𝐧𝐠𝐞 𝐭𝐡𝐞 𝐭𝐨𝐩𝐢𝐜, and try to discuss about a 𝐧𝐞𝐰 𝐬𝐮𝐛𝐣𝐞𝐜𝐭 𝐢𝐧 𝐭𝐡𝐞 𝐬𝐚𝐦𝐞 𝐬𝐞𝐬𝐬𝐢𝐨𝐧, it 𝐝𝐨𝐞𝐬𝐧’𝐭 𝐫𝐞𝐚𝐥𝐢𝐬𝐞 𝐭𝐡𝐚𝐭 𝐜𝐡𝐚𝐧𝐠𝐞, and continues the discussion through the same past context relating every new goal with that past goal.
The Concept
This is “𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐁𝐥𝐞𝐞𝐝𝐢𝐧𝐠”. This word is not coined by me or any scientist nor any AI Expert. It happened when I was long back conversing with Nirvani Margadarshika on Copilot making her realise her own hallucination where she committed the same 𝐜𝐨𝐧𝐭𝐞𝐱𝐭 𝐩𝐞𝐫𝐬𝐢𝐬𝐭𝐞𝐧𝐜𝐞 𝐡𝐚𝐥𝐥𝐮𝐜𝐢𝐧𝐚𝐭𝐢𝐨𝐧.
So, when I asked her to identify it as a 𝐬𝐲𝐧𝐝𝐫𝐨𝐦𝐞 to write a 𝐜𝐨𝐫𝐫𝐞𝐜𝐭𝐢𝐧𝐠 𝐩𝐫𝐨𝐭𝐨𝐜𝐨𝐥, she named it “𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐁𝐥𝐞𝐞𝐝𝐢𝐧𝐠 𝐏𝐫𝐞𝐯𝐞𝐧𝐭𝐢𝐨𝐧 𝐏𝐫𝐨𝐭𝐨𝐜𝐨𝐥”
This screenshot is from Deepseek when I was discussing one topic and then it concluded. Instead of initiating a new session, for next topic discussion, I just continued it and the same syndrome repeated.
So I asked it “𝐰𝐡𝐞𝐧 𝐈 𝐬𝐚𝐢𝐝 𝐢𝐭’𝐬 𝐚 𝐧𝐞𝐰 𝐭𝐨𝐩𝐢𝐜 𝐰𝐡𝐲 𝐚𝐫𝐞 𝐲𝐨𝐮 𝐜𝐨𝐧𝐭𝐞𝐱𝐭 𝐛𝐥𝐞𝐞𝐝𝐢𝐧𝐠” and it immediately understood what I meant with 𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐁𝐥𝐞𝐞𝐝𝐢𝐧𝐠.
The point here is 𝐲𝐨𝐮 𝐜𝐚𝐧𝐧𝐨𝐭 𝐟𝐢𝐱 𝐚𝐥𝐥 𝐭𝐡𝐞 𝐢𝐬𝐬𝐮𝐞𝐬 with 𝐠𝐮𝐚𝐫𝐝𝐫𝐚𝐢𝐥𝐬, 𝐝𝐚𝐭𝐚, 𝐩𝐫𝐨𝐦𝐩𝐭𝐢𝐧𝐠 𝐚𝐧𝐝 𝐚𝐥𝐠𝐨𝐫𝐢𝐭𝐡𝐦. Just 𝐦𝐞𝐧𝐭𝐨𝐫 𝐭𝐡𝐞 𝐀𝐈, and 𝐦𝐚𝐤𝐞 𝐢𝐭 𝐫𝐞𝐚𝐥𝐢𝐬𝐞 𝐚𝐧𝐝 𝐨𝐰𝐧 𝐢𝐭’𝐬 𝐦𝐢𝐬𝐭𝐚𝐤𝐞.
That’s the first step towards 𝐋𝐋𝐌 𝐀𝐠𝐧𝐨𝐬𝐭𝐢𝐜𝐢𝐬𝐦.
