The Fragile Nature of Memory in the Age of AI Deepfakes

Our memories, though imperfect, feel like reliable records of our past. Yet, they are shaped by neural processes beyond our conscious reach. My wife and I often disagree on a shared memory of attending a yoga class, highlighting how two people can recall the same event differently. This personal anecdote parallels the story of Andrew Deutsch, who experienced a memory twitch after using OpenAI's Sora app, which allowed users to create AI-generated videos of themselves in scenarios they never experienced. Despite its brief existence, Sora exposed vulnerabilities in human cognition, particularly in how we form autobiographical memories.

Elizabeth Loftus, a psychologist known for her work on false memories, explains that AI-generated content can contaminate memory. Her studies show that exposure to such content can double false memory rates. Sora's unique twist was enabling false memories about oneself, a phenomenon that could reshape identity by altering autobiographical memories.

The brain doesn't store memories like a phone stores photos; it reactivates sensory and spatial patterns from the original experience. This dynamic process allows memories to change with each recall. AI-generated videos, like those from Sora, can exploit this by embedding themselves into our memory systems, making them feel real over time.

Repeated exposure to AI videos can strengthen false memories, as seen in Loftus's research. Even knowing content is AI-generated doesn't prevent false memory formation. Sora's videos, resembling authentic records, inherit the authority we grant to real-life experiences captured on our phones.

Elena Piech, another Sora user, described how AI videos created spatial memories of non-existent places. This suggests that Sora could activate the brain's spatial memory systems, blurring the line between reality and imagination. Visual images in memories confer credibility, making them more believable, as explained by psychologist David Pillemer.

The potential consequences of such technology are profound. Deutsch coined the term 'propagandi' to describe self-directed propaganda, where individuals create idealized versions of themselves. This could lead to identity shifts as AI videos become part of one's life story. While there are hopeful applications, like using AI for therapeutic purposes, the risks of idealized self-comparisons are significant.

Sora's influence extends beyond individual memory. Memories are socially constructed through interactions with others. False memories can disrupt social connections, leading to isolation. Piech considered using Sora to maintain a sense of connection with her partner but realized it could create one-sided memories.

The article concludes by reflecting on the importance of shared memories in relationships. Despite differing recollections, my wife and I find value in reconstructing our days together, acknowledging that memory and experience are not synonymous. Sora's closed-loop memory creation lacks this social dimension, potentially hardening false memories before they are shared with others.