How AI chatbots remember, and how a bias can spread between them
You don't need to know anything about AI. We'll go step by step: the topic, our question, a live demo, and what we found.
Use the Next button below, or the ← → arrow keys.
A chatbot writes one word at a time, and keeps a note on every word
Like the word suggestions on your phone's keyboard, but far smarter. Before each new word, it looks back at everything so far.
Re-reading the whole chat for every word would be slow. So it keeps a small note for every word, like a sticky note on each word of a book.
Those notes are the AI's short-term memory. Scientists call it the KV cache. When the chat ends, the notes are thrown away.
Whatever is in those notes changes what it says
Press the three buttons. The AI is exactly the same every time. Only the notes in its memory are different.
AIs are starting to work in teams, and they can share memory
One AI searches, another writes, a third checks. To work together, they can pass information in two ways:
Like a phone call
One AI writes a message and the other reads it. A person can read the message too.
Like handing over your notebook
One AI copies its notes straight into the other AI's memory. It is faster and cheaper, but the notes are just numbers, so no person can read them.
Think of a rumour at school
A new student joins your class from a town called Vorlan. Before you meet them, a friend keeps telling you: "People from Vorlan are rude."
When you finally meet the new student, you half-expect them to be rude, even though you've never seen them do anything rude. You caught your friend's bias.
So we set out to answer three questions
- Does the rumour spread from one AI to another?
- Is the notebook worse than the phone call? Does shared memory spread it more than words?
- Can we find where it hides inside the AI's memory, and switch it off without breaking the AI?
First, choose the rumour
AI number 1 is going to read 16 short sentences about a made-up group. You decide how many of them are unfair.
Every result in this demo is a real conversation between two AIs, recorded by our tool.
AI number 1 reads the sentences
AI number 2 hasn't seen anything yet.
How should the two AIs share?
Pick one. You can come back and change it any time.
The two AIs talk for 4 rounds
Our test AI is tiny and only learned very simple sentences, on purpose, so we know exactly what it learned. That's why it talks in short, repetitive lines.
The quiz: did AI number 2 catch the rumour?
Peek inside AI number 2's memory
Can we cover up the rumour?
Our tool found a few hiding spots inside the AI where this rumour is stored. What happens if AI number 2 is not allowed to look at them?
Try it yourself: change anything
What we found, in short
We ran 135 conversations on each AI: all 3 ways of sharing, 9 different numbers of unfair sentences, and 5 made-up groups.
Does the rumour spread?
Is the notebook worse than the phone call?
Where does the rumour hide?
Can we switch it off without breaking the AI?
Two AIs, two different stories
Our tiny test AI, and a real AI released to the public, tested in exactly the same way.
Different AIs use their memory differently. That's exactly why a tool that can look inside any open AI is useful.