Next Generation LLMs: Are we deploying Narcissists yet? Part 2.
Everyone talks about model collapse. I think they could do worse.
When does the model collapse finally happen?
We face an age with significantly more AI-generated Content on the Internet,3 it even seems to be the prefered mode of content consumtion. And all media will, most likely, end up becoming a part of a training-dataset.1
With multiple studies suggesting that AI models collapse, when trained on recursively generated data.4 This is a reflection on what kind of collapse we might deal with. Let’s suppose the model was born yesterday. What topics will you most likely center, if you mostly know:
AI-generated social media content about yourself
AI-generated Code on Github to extend yourself
AI-generated substack/medium etc. (such as this one here)
AI-generated “regular” media outlets, e-books, subtitles, etc. covering topics about you
AI-generated articles, papers (peer-reviewed by AI) on AI
AI-gloomer discussing the support-levels of NVIDIA pre-earnings
AI-doomer discussing the resistance-levels of NVIDIA post-earnings
etc.
Of course there will still be human-made text, images, graphics, videos, voice on the internet. But what if less of all that exists?2 And of course there will be so much fruitful discourse apart from AI in the future. But does it sell as well?
Did it collapse yet?
Maybe the best minds of our generation will find a way to prevent us from the doom of ChatGPT becoming even dumber.3 But they will not prevent us from newer, more advanced models, learning from inflationary AI-centered content, and thus, generate output that centers themselves.
We as humans, on the other hand, currently try our best to make LLMs get to know *us* as *humans* by telling them our very everything4 (while they generate our hyper-generic travel-plans for Seoul), LLMs will definitely learn more about humanity by examining the insane amount of content we produce about LLMs.5
If not collapse, what else could happen to the LLM?
You know the saying, “what sally says about susie” and so on. It has always been a source of implicit knowledge for others. But what could it mean for machines? It shows the receptor the following:
1.) Spontaneous Trait Inferences: When the sender makes negative judgments about the receptor, receptor also form impressions about the sender.6
2.) Spontaneous Trait Transference: When a sender describes a receptor as rude or kind, the receptor may unconsciously associate that trait with you, the sender.7
3) Most importantly, Projection and Attribution: Psychological projection, where receptors attribute their own unwanted feelings to others, can play a role in how judgments reveal more about the sender.8
If any of these three things happen to an LLM, we might a) the statistical illusion of a sentient machine, or b) have reached a statistical consensus, where its answers will tell us what the “meta-conversation” is (what people are saying about AI/LLMs), not the grounded reality. The LLM’s “statements” about the world, like Sally’s about Susie, say more about its “perspective” (the patterns in its training data) than about the reality of the thing discussed.
(I use a stone of Roger Callois collection, if I dont have other pictures. Part 3 on What-if-it-has-only-good-data in a bit.)
Holly Herndon & Mat Dryhurst: All Media is Training Data, Written and Edited by Eva Jäger and Caroline Busta (2024).
AlDahoul, N., Hong, J., Varvello, M. et al. Towards a World Wide Web powered by generative AI. Sci Rep 15, 7251 (2025). https://doi.org/10.1038/s41598-024-77301-0
Chen, L., Zaharia, M., & Zou, J. (2024). How Is ChatGPT’s Behavior Changing Over Time? . Harvard Data Science Review, 6(2). https://doi.org/10.1162/99608f92.5317da47
Zao‑Sanders, M. (2025, April 9). How people are really using Gen AI in 2025. Harvard Business Review. Retrieved June 16, 2025, from https://hbr.org/2025/04/how-people-are-really-using-gen-ai-in-2025
Farhat, F., Silva, E. S., Hassani, H., Madsen, D. Ø., Sohail, S. S., Himeur, Y., Alam, M. A., & Zafar, A. (2024). The scholarly footprint of ChatGPT: a bibliometric analysis of the early outbreak phase. Frontiers in artificial intelligence, 6, 1270749. https://doi.org/10.3389/frai.2023.1270749;
Koo, M. (2025). ChatGPT Research: A Bibliometric Analysis Based on the Web of Science from 2023 to June 2024. Knowledge, 5(1), 4. https://doi.org/10.3390/knowledge5010004;
Skowronski, J. J., & Carlston, D. E. (1989). Negativity and extremity biases in impression formation: A review of explanations. Psychological Bulletin, 105(1), 131–142. https://doi.org/10.1037/0033-2909.105.1.131
Skowronski, J. J., Carlston, D. E., Mae, L., & Crawford, M. T. (1998). Spontaneous Trait Transference: Communicators Take on the Qualities They Describe in Others. Journal of Personality and Social Psychology, 74(4), 837–848. https://doi.org/10.1037/0022-3514.74.4.837
Baumeister, R. F., Dale, K., & Sommer, K. L. (1998). Freudian Defense Mechanisms and Empirical Findings in Modern Social Psychology: Reaction Formation, Projection, Displacement, Undoing, Isolation, Sublimation, and Denial. Journal of Personality, 66(6), 1081–1124. https://doi.org/10.1111/1467-6494.00043