Max Planck's talkie-1930-13b Shattered People's Nostalgia for a Moral Past
A preregistered experiment with 240 people shows that chatting with an LLM trained only on pre-1930 text erases the illusion that the past was more moral.
- Researchers trained talkie-1930-13b, a 13B LLM on only pre-1930 text.
- Preregistered N=240 experiment compared it against GPT-5.5 as contemporary control.
- Interacting with the 1930 model reduced the illusion of moral decline; control showed no effect.
- Participants predicted the model's answers on topics like women working and racial equality.
- Authors propose knowledge cutoff as a controllable variable for social science experiments.
- Main critique: model size, tuning, and style also differ, confounding the causal claim.
People often compare the present with a reconstructed past assembled from books, films, family stories, and memories revised over time. A study from the Max Planck Institute for Human Development and its collaborators tests whether direct interaction with a historically bounded language model can change judgments about moral decline.
The Time Machine Experiment uses a model trained exclusively on text published before 1930. The researchers report that participants assigned to this model became less likely to describe morality as declining over time, while participants assigned to a contemporary model showed little change.
Why nostalgia survives the archive
The study targets the illusion of moral decline, a documented tendency to believe that earlier generations were kinder, more honest, and more respectful. Long-running surveys provide little support for a broad decline, yet the belief persists because recollections of the past are selective and difficult to test through conversation.
Historical documents cannot answer new questions, while surviving witnesses have incorporated decades of later experience into their memories. A contemporary chatbot instructed to role-play someone from 1930 also retains knowledge acquired from post-1930 training data, which can leak into its answers.
A temporally bounded model offers a different form of evidence: generated responses constrained by a period-specific corpus. It simulates patterns in surviving text rather than the mind of an average person from 1930. Published material from that period overrepresents institutions, editors, professional writers, and people with access to print, so corpus composition determines whose attitudes the model reflects.
Turning 1930 into a cutoff
The researchers trained a 13-billion-parameter model called
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
