A ‘1930’ AI made the past look less golden

A golden age is easier to imagine when its inhabitants cannot answer back. An experiment with historical AI gives that thought a practical form: put unfamiliar judgments in front of people, then ask whether the past still looks so admirable.

In a September preprint, 240 UK and US adults were randomized to Talkie, presented as a 1930 AI, or GPT-5.5. They conversed and predicted the model’s moral judgments. Their immediate ratings shifted more after the historical-model encounter.[1]

Participants rated how much people today and about 100 years ago valued kindness, honesty and goodness, on seven-point scales. Researchers subtracted the past rating from the present one. This gap rose by +0.87 points with Talkie versus +0.08 with GPT-5.5: a 0.79-point difference in change, with a 95% confidence interval of 0.44–1.15. The shift mainly reflected lower ratings of the past.[1]

Why the past needed a different model

Talkie’s developers describe a 13-billion-parameter model trained on 260 billion tokens from English-language material dated before 1931. Its intended boundary is the end of 1930. Books, newspapers, journals, patents and legal texts supply the material from which it learns to continue language.[2]

That is a different exercise from asking a contemporary chatbot to adopt an old-fashioned persona. A role-playing instruction changes how a system presents itself; restricting the training corpus changes the material it can learn from. The distinction makes the archive part of the experiment, rather than merely its costume.

The boundary is imperfect. Talkie’s developers report that the 13B model knows some details of the Second World War and the postwar order. They identify misdated documents and later additions to old texts as ways newer information can enter a historical corpus.[2]

The comparison hidden inside nostalgia

The psychological question predates chatbots. In their 2023 Nature paper, Adam Mastroianni and Daniel Gilbert found a revealing mismatch: people repeatedly reported moral decline, while surveys asking about the conduct of their contemporaries showed little change across decades. The two questions ask people to perform different tasks—remember a comparison, or describe the present.[3]

The authors proposed that exposure to bad behaviour now, combined with the fading impact of negative memories, can make earlier life look better. That mechanism concerns the information available to an observer. It does not require each generation to become steadily worse.[3]

Historical writing offers a way to make an imagined past less abstract. A useful question for a classroom discussion would be: which people benefit from the old rule being defended, and whose choices does it restrict? This is an editorial suggestion for examining an argument, not a tested teaching outcome. The point is to inspect the reasons beneath an attractive label such as tradition.

The experiment used twelve selected, contested topics and fixed reference answers for the prediction task; each participant encountered four. Models also differed in capability, style and introduction. The result therefore concerns that designed encounter, not the cutoff alone. It measures immediate perceptions, with persistence untested.[1]

A generated answer is also several steps removed from historical evidence. A surviving text has an author, audience and purpose; a model blends patterns across texts. An apparently confident judgment does not tell the reader who held it, how widespread it was or who challenged it. Those are questions for the documents and their historical context.

The experiment’s most interesting use of AI is consequently a question directed at the living: what, exactly, have you included in your picture of the good old days? A model can make that picture uncomfortable enough to examine. The examination still belongs to the person reading.

Report an error or suggest a correction