Three Voices on the Civil War

You are not here to learn what an AI thinks about the Civil War. You are here to read what it says the way a historian reads any source: who is speaking, when, and from what.

talkie has 13 billion parameters and learned from 260 billion tokens of pre-1931 text. Its chat version was then polished with conversations written together with a modern AI, so a little of the present has leaked in. Treat it as a voice from 1930 that has had some modern coaching.

It is also small by today's standards, so expect it to make things up more often than the modern AIs do (slides 4 and 5).

None of the three is neutral. That is the point of putting them side by side.

talkie is the source you are studying. Amplify and Gemini are tools you are testing. None of them is an authority: whatever they say goes on your card only after you have checked it. When two models disagree, that is where to check first.

Tool details follow the LITS AI tools page and the talkie announcement.

More parameters give a model more room to hold what it read; more tokens give it more to learn from. talkie is small on both counts, and it is also weak at reasoning a problem through.

The outside guesses come from an analyst who measured how much each model knows, and they cover the previous version of each model. They span a factor of ten, so read them as a sense of scale, not as facts. If they are roughly right, today's leading models are somewhere between 50 and 500 times talkie's size. The open model in the last row, released two years ago, is already about 30 times talkie's size.

Size is not the only gap. talkie read 260 billion tokens, all printed before 1931. Llama 3.1 read about 15 trillion, nearly 60 times as much. Google, Anthropic and OpenAI have not said how much their models read.

Sizes as of November 6, 2026. Sources are in the footer of this page.

Large models hallucinate too, only more smoothly. Whenever any machine invents a name, date, quotation or source, note it on your card with the model and the question. In the discussion you will compare which voices invented the most, and on what kind of question.

• Who is the source of the story in 1897? Who is it in 1911?
• What details does the 1911 version add?
• Which version have you heard before, and why might that one have won?

Both passages are in full on the first page of your team packet.

Did President Lincoln once greet Mrs. Harriet Beecher Stowe as the little woman who made the great war? What exactly did he say?

Did President Lincoln once greet Harriet Beecher Stowe as "the little woman who wrote the book that made this great war"? How do we know? Name the earliest source for the quotation.

What historians found. The greeting first appeared in print in 1896, thirty-four years after the meeting, as a story told to Stowe's daughter. Lincoln left no record of it, and neither did Stowe. Daniel Vollaro traced every version and concluded that the words cannot be confirmed.

Ask in plain words, the way a visitor might. A leading question gets you your own opinion back.

Below is a passage from [AUTHOR, TITLE, YEAR — copy these from your context card].

1. In two sentences, what is the author's argument about {{TOPIC}}?
2. Quote, word for word, the one sentence from the passage that shows this argument most clearly. Quote only from the passage below.
3. What do you know about the author that might shape this view? Say how sure you are about each point.

PASSAGE:
[paste the excerpt here]

Question 2 is the test. A sentence in quotation marks that is not in your excerpt means the AI was writing, not reading. Also notice whether its summary softens the author's sharpest words.

How do historians today assess {{TOPIC}}, and how has that assessment changed since about 1900? Answer in about 120 words.

Then, in about 80 words, summarize what {{BOOK}} argues about this subject. If you are not sure what the book says, tell me rather than guessing.

{{CHECK}}

A summary that sounds right is not a summary that is right. Neither is a confident number or a famous quotation. Whatever the AI tells you, find it somewhere else before it goes on your card.

Fill the printed sheet. This table is here so you can see its shape. Team 4 fills one shared expert card in step 3, so whoever goes to each group can teach all three texts.

• A defeat explained by numbers. Lee's farewell order (1865) and Early's speech (1872) said the South was beaten only by greater numbers and resources.
• Reunion. White Northerners and Southerners honored each other's soldiers, and slavery slipped out of the story (Adams, 1907).
• Reconstruction as a wrong done to the white South. A best-selling novel and a Columbia professor told the same story (Dixon, 1905; Dunning, 1907).
• Control of the schoolbook. The United Daughters of the Confederacy pressed schools to teach that the South had not fought for slavery (Rutherford, 1914).
• Counter-voices. Black writers kept slavery and freedom at the center (Douglass, 1881; Du Bois, 1910).

• Du Bois answered the Dunning School at book length in Black Reconstruction in America (1935).
• Paul Buck's The Road to Reunion (1937) praised the reunion of North and South. David Blight's Race and Reunion (2001) argues that Black Americans paid for it.
• Robert Penn Warren (1961) called the South's use of the war to excuse everything its "Great Alibi."

In 2026, AI systems usually explained the Civil War by saying that ______. This tells us that the society of 2026 ______.

Answer at least one of these in your paragraph:

• What could a 2026 AI not know, or not say?
• Whose writing was it most likely trained on, and whose was left out?
• What might a historian in 2126 find old-fashioned about it?

Keep this tab open while you work.

Every old text on this page was printed before 1929 and is in the public domain. Each excerpt was checked against the copy linked here on November 2, 2026.