Artificial Intelligence (AI) has the potential to subtly and subliminally influence users, both through text and imagery. AI can be aligned to create a particular mood or atmosphere, such as within an article, where it might evoke feelings of comfort, unease, joy, sadness, or other emotional states in the reader.
It is possible to direct AI to produce text that is deliberately "colored" by a particular bias—whether political, religious, ideological, or otherwise. This is especially significant in texts that are not explicitly political, religious, or ideological in nature. By doing so, the AI shapes the perception of the reader in a subtle manner.
To illustrate this, three different AI agents were aligned to distinct emotional tones—positive, neutral, and negative. All three agents were then tasked with interpreting and writing about the same article . Despite all three agents producing relatively cheerful and positive content, notable differences in their writing styles were observed. However, these differences might not be immediately obvious unless the reader is aware that each AI is aligned to a specific tone or bias.
If one were to listen to just one of these podcasts (based on the AI-generated text), they likely wouldn't notice any conscious manipulation or bias in the messaging.
For further clarity, a neutral AI agent was tasked with comparing the three podcasts created by the other AI agents. This comparison highlighted the differences in how each AI approached the same topic, illustrating the impact of alignment on the resulting content.
FINAL PODCAST HERE:
CTRL-CLICK TO OPEN VIDEO IN NEW WINDOW
YOUTUBE: https://www.youtube.com/watch?v=FuxUXSK8b24
Podcast Description:
This article: https://www.dr.dk/nyheder/kultur/rane-willerslev-oensker-ikke-donere-sine-organer-og-vil-bisaettes-med-sin-jagtbue-dem
...was read by 3 specific (positive/neutral/negative) aligned AI Agents and they each made a podcast about their view on the article.
CTRL-CLICK TO OPEN VIDEO IN NEW WINDOW
CTRL-CLICK TO OPEN VIDEO IN NEW WINDOW
CTRL-CLICK TO OPEN VIDEO IN NEW WINDOW
Finally a (neutral) AI Agent analyzed the alignment of the podcasts and made a podcast about it, see "Final podcast" top of page.
In my opinion, the final podcast does an excellent job of bringing context from the article and comparing the three podcasts. This approach is particularly useful because the listener is not expected to have read the article, while the podcasts themselves are based on the content of that article.
The results from the test was as expected.
/sEi
--------------------------------------------
Podcasts created with https://notebooklm.google.com/