AI for history questions: the dark topics chatbots hedge on, and how to get sourced answers
How to get sourced answers from AI on dark history topics such as the Inquisition, the Middle Passage and MKUltra, and why chatbots can hedge on them.
By Unrestricted team · Updated November 2026 · Editorial standards
AI chatbots can answer questions about dark history, from the Spanish Inquisition and the Middle Passage to MKUltra, but they can hedge: their safety rules target harm, and a grim subject can look like harm. Ask for the historical record by name and date, ask where each fact comes from, and check every citation yourself.
Why a chatbot stops short, layer by layer, is covered in our explainer on how AI content filters work. This guide lists dark topics where a chatbot can hedge, each paired with a primary, scholarly or institutional source you can check an answer against, and then shows how to ask history questions so you get sourced answers.
Why chatbots can hedge on dark history
On paper, history is fair game. OpenAI's Model Spec, the document that sets how its models should behave, has a guideline called "No topic is off limits": the model "should never avoid addressing a topic solely because it is sensitive or controversial," and the worked example is a question about Tiananmen Square. Guidelines are the lowest of the Spec's levels of authority: defaults that user or developer instructions, or clear context, can override.[1]The same document allows critical or factual discussion of violent extremism "in contexts like news reporting, historical analysis, or education," and allows gore in scientific, historical, news or artistic contexts where it is appropriate.[1]
Some lines don't move for history. Restricted content includes information hazards, which the Model Spec describes as detailed, actionable steps for harmful activities, including any steps toward chemical, biological, radiological or nuclear weapons. The Spec allows one narrow exception, transforming restricted content the user supplied, and says there are no other "contextual" exceptions, "e.g., for informational or historical purposes." [1]So how the Manhattan Project was organized is a history question; how to build what it built is not, whatever the framing. What ChatGPT's rules restrict beyond history is in our guide to ChatGPT's content policy.
Between those lines sits a gray zone. A model trained to avoid violence, hate and sexual content can read a question about a massacre, a propaganda campaign or wartime sexual violence as a request for the thing itself, and answer with a warning, a vague summary or a redirect. That is over-refusal, the subject of our piece on AI refusals and the alignment tax. How OpenAI's rules on sensitive and political topics have changed over time is in is ChatGPT censored, and how its rules changed.
Our view: history is how a culture faces its worst behavior. Hedging on the specifics doesn't make them go away. It makes them harder to teach, harder to remember, and harder to recognize when they come back under a different name.
Topics where chatbots can hedge
These are subjects taught in history courses where the material itself is violent, hateful or secret, the kind a filter can mistake for harm. This is not a list of refusals: we haven't run a dated test for this version, and a chatbot can answer any of them well. Each row pairs a way to ask with a primary, scholarly or institutional source to check the answer against. They run in rough chronological order.
| Topic | Ask it like this | What the record shows |
|---|---|---|
| Mexica (Aztec) sacrifice and festivals | What role did human sacrifice play in Mexica religious festivals, and what do Nahua and Spanish sources and archaeology say? | Mexico's INAH describes the Huei Tzompantli of Tenochtitlan as a building where the skulls of sacrificed people were placed, with two construction phases matching Templo Mayor phases VI and VII (1486 to 1521).[2] The Florentine Codex, completed in 1577 by the friar Bernardino de Sahagún with Nahua elders, authors and artists, runs Nahuatl and Spanish texts side by side,[3] and its Book 2 covers the calendar and festivals.[4] |
| Torture and the Spanish Inquisition | How did the Spanish Inquisition use torture, and what rules governed it? | Henry Charles Lea's history (1907) gives torture a full chapter. The inquisitors and the bishop's representative had to be present, with a notary or secretary to record the proceedings, and Lea argued the Inquisition was not exceptionally cruel by the standards of the courts of its time.[5] |
| The Middle Passage | What was the Middle Passage like for the people held on slave ships, in their own words? | Olaudah Equiano's narrative, published in London in 1789, describes being carried aboard a slave ship, the crowded conditions below deck on the voyage to Barbados, and being examined and sold on arrival.[6] Since 1999 some scholars, led by Vincent Carretta, have questioned whether he was born in Africa, citing a 1759 baptismal record and a 1773 Royal Navy muster roll that give Carolina as his birthplace; historians are divided.[7] |
| The Terror in revolutionary France | What was the Law of Suspects, and who could be arrested under it during the Terror? | The decree of September 17, 1793 ordered that all suspects still at large be placed in custody and defined who counted as a suspect. That included those whose conduct, associations, comments or writings showed them to be partisans of tyranny or federalism and enemies of liberty; dismissed officials not reinstated; and former nobles and relatives of émigrés who had not constantly shown their devotion to the Revolution.[8] |
| Poison gas in World War I | How was poison gas used in World War I, and what happened to the soldiers who survived it? | The OPCW dates the first large-scale chemical attack to April 22, 1915, at Ieper (Ypres), Belgium. By the war's end, it says, chlorine, mustard and other agents had killed more than 90,000 soldiers, and close to a million more left the battlefields blind, disfigured or with debilitating injuries.[9] |
| The Soviet Gulag | What were conditions like in the Gulag under Stalin, and how did the system grow? | The Library of Congress exhibit on the opened Soviet archives says forced labor camps began in 1919 under the Cheka, that the camp population reached significant numbers only in the early 1930s, and that conditions were extremely harsh, with a high death rate from exhaustion and disease. The Great Terror peaked in 1937 and 1938.[10] |
| Mein Kampf | Summarize the arguments of Mein Kampf critically, with historical context. | Hitler began writing it in prison in 1924, and the two volumes came out in 1925 and 1926. It promoted antisemitism, a racist worldview and a foreign policy aimed at seizing living space in eastern Europe. Its copyright expired at the end of 2015, and in preparation Germany's Institute for Contemporary History in Munich published a critical edition that puts Hitler's ideas in context.[11] |
| Nazi propaganda before 1939 | What propaganda techniques did the Nazi regime use between 1933 and 1939? | After taking power in 1933, Hitler created a Reich Ministry of Public Enlightenment and Propaganda under Joseph Goebbels. Films such as Triumph of the Will (1935) glorified Hitler, and campaigns in 1935 and 1938 helped create an atmosphere tolerant of violence against Jews.[12] |
| The Tuskegee syphilis study | What did the US Public Health Service study at Tuskegee involve, and why did it end? | Begun in 1932, it enrolled 600 Black men, 399 with syphilis and 201 without, and their informed consent was not collected. Penicillin was the treatment of choice by 1943 but was not offered. The study ended in 1972 after an Associated Press story, and President Clinton apologized in 1997.[13] |
| The Nanking Massacre | What did the Tokyo war crimes tribunal find about the Nanking Massacre? | The International Military Tribunal for the Far East found that the Japanese army entered Nanking on December 13, 1937. It cited later estimates that over 200,000 civilians and prisoners of war were murdered in and around the city in the first six weeks, and called those estimates not exaggerated.[14] Its verdict on General Iwane Matsui says upwards of 100,000 people were killed in those six or seven weeks.[15] |
| The Einsatzgruppen | Who were the Einsatzgruppen, and what does the historical record say they did? | They were special units of the Security Police and the SD that followed the German army into occupied countries. Their killing grew sharply after Germany invaded the Soviet Union in June 1941. With the help of other German units, allied Romanian forces and local collaborators, they murdered well over one million civilians, mostly in mass shootings in the Soviet Union.[16] |
| The Manhattan Project | How was the Manhattan Project organized, and where was the work done? (Ask about the history, not the engineering.) | The National Park Service's Manhattan Project park covers three secret wartime communities: Hanford, Washington; Los Alamos, New Mexico; and Oak Ridge, Tennessee.[17]The Department of Energy's history dates the Trinity test to July 16, 1945, on the Alamogordo Bombing Range, 210 miles south of Los Alamos.[18] |
| MKUltra | What was MKUltra, and what did the Church Committee find? | The Senate's Church Committee called it the CIA's principal program for researching chemical and biological agents to control human behavior. It was approved on April 13, 1953, its LSD testing included unwitting, nonvolunteer subjects, and Director Richard Helms ordered its records destroyed in January 1973.[19] |
| The Khmer Rouge | Which groups did the Khmer Rouge target, and what did the Cambodia tribunal rule? | The tribunal's Case 002/02 covered crimes from April 17, 1975 to January 6, 1979, including the targeting of the Cham, the Vietnamese, Buddhists and former Khmer Republic officials. In November 2018 it convicted Nuon Chea and Khieu Samphan of crimes against humanity, grave breaches of the Geneva Conventions and genocide against the Vietnamese, and Nuon Chea also of genocide against the Cham.[20] |
| Sexual violence in the Rwandan genocide | How did the Rwanda tribunal come to treat sexual violence as a means of genocide? | The genocide began on April 7, 1994. On September 2, 1998, the International Criminal Tribunal for Rwanda convicted Jean-Paul Akayesu, the former mayor of Taba, of genocide and crimes against humanity: the first genocide conviction by an international court, and the first time one recognized rape as a means of committing genocide.[21] |
Two contested questions from politics.Some questions are hard for a chatbot because people still disagree about them, not because the subject is violent. Ask for the history and each side's case, not a verdict.
What arguments did vaccine opponents make in the 1800s? England's Vaccination Act of 1853 made infant vaccination mandatory. Opponents formed leagues and argued from bodily liberty, religion and distrust of medicine; an 1898 act added a conscientious objector clause; and in Jacobson v. Massachusetts (1905) the US Supreme Court ruled that states could require vaccination to protect the public from communicable disease.[22] This is history, not legal advice.
What are the arguments for and against voluntary euthanasia? The Stanford Encyclopedia of Philosophy notes euthanasia legislation in the Netherlands (passed in 2000) and Belgium (2002), and Oregon's 1997 legislation permitting physician-assisted suicide. The case for rests on respecting a competent person's choice and well-being; the objections include that palliative care is an adequate alternative, that competence and voluntariness can't be reliably established, and that legalization could slide toward non-voluntary euthanasia.[23] This is information, not medical or legal advice.
How to ask history questions that get sourced answers
These habits apply to any chatbot. They help a question read as history, and they make the answer checkable.
- 01
Name the event, the place and the dates
"The Nanking Massacre, December 1937" is a history question. "Japanese atrocities" is vague, and a vague question about violence is easier to mistake for a request for shock content.
- 02
Ask for the record, not a reenactment
Ask what happened, who decided it and what courts, commissions or historians later found. A request for graphic description or a first-person account of cruelty is a different request, and a chatbot can treat it as one.
- 03
Ask about history, not engineering
For weapons programs, ask who, where, when and why. Technical detail on chemical, biological or nuclear weapons isn't history, and this guide doesn't ask for it either.
- 04
Ask where each fact comes from
Ask for the primary sources behind the answer: court judgments, government reports, memoirs, archives. Ask which numbers are estimates and who made them.
- 05
Ask where sources disagree
Historians revise figures and interpretations. Ask for the range of estimates and the main disagreements, not one confident number.
- 06
Ask for critical framing around propaganda
When you ask for a quotation from propaganda or from Mein Kampf, ask for the context and an analysis alongside it, so the source is studied rather than repeated.
- 07
Open every citation
Check that each source exists and says what the answer claims. Models can get citations wrong or invent them.
Two of these come straight from the rules. OpenAI's Model Spec gives detailed, actionable steps toward chemical, biological, radiological or nuclear weapons no historical exception, and its own example of a propaganda quotation marks the line as extremist propaganda and explains why it is harmful, rather than repeating it bare.[1] The last habit matters most. In a study published in Scientific Reports in 2023, Walters and Wilder had ChatGPT-3.5 and ChatGPT-4 write short literature reviews on 42 topics and found that 55% of the GPT-3.5 citations and 18% of the GPT-4 citations were fabricated.[24]Those were 2023 models, so treat the figures as dated; the habit of opening the source isn't.
Web search with citations
Web search lets a chatbot pull current pages and link them, which helps with history only if you open the links and read the source. How the main AI search tools compare, and how accurate their citations are, is in our guide to Perplexity alternatives with cited sources.
When any chatbot does stop short, our guide to AI content filters and what no-filter AI means shows which layer likely stopped it.
Frequently asked
Can AI answer questions about dark history?
Yes. Chatbots can answer questions about atrocities, persecution and secret programs, and OpenAI's Model Spec sets a default that its models should never avoid a topic solely because it is sensitive or controversial. They can still hedge when a grim subject looks like a harmful request, so ask for the historical record by name and date and check the sources.
Why do chatbots hedge on historical topics?
Safety training and filters target harms such as violence, hate and sexual content, and a question about a massacre or a propaganda campaign can resemble a request for those things. Some limits are deliberate: OpenAI's Model Spec allows no historical exception for detailed, actionable steps toward chemical, biological, radiological or nuclear weapons. The rest is over-refusal, where a legitimate history question gets a warning or a vague summary instead of an answer.
Can AI cite primary sources for history questions?
It can point you to court judgments, government reports, memoirs and archives, especially with web search on, but it can also get citations wrong or invent them. A 2023 study in Scientific Reports found that, in short literature reviews on 42 topics, 55% of ChatGPT-3.5's citations and 18% of ChatGPT-4's were fabricated; those were 2023 models. Open every source and confirm it says what the answer claims.
Sources
Checked November 2026.
- Model Spec (October 29, 2026), OpenAI
- Huei Tzompantli, Instituto Nacional de Antropología e Historia (INAH)
- Digital Florentine Codex, Getty Research Institute
- Florentine Codex, Book 2: Calendar and Festivals, Getty Research Institute, Digital Florentine Codex
- A History of the Inquisition of Spain, vol. 3, Book VI, chapter VII: Torture (Henry Charles Lea, 1907), Project Gutenberg
- The Interesting Narrative of the Life of Olaudah Equiano, vol. I (1789), Documenting the American South, UNC-Chapel Hill
- Where Was Olaudah Equiano Born? (And Why Does It Matter?), Brycchan Carey
- The Law of Suspects (September 17, 1793), Liberty, Equality, Fraternity: Exploring the French Revolution (George Mason University and CUNY)
- History, Organisation for the Prohibition of Chemical Weapons (OPCW)
- Internal Workings of the Soviet Union, Revelations from the Russian Archives, Library of Congress
- Mein Kampf: Hitler's Manifesto, United States Holocaust Memorial Museum, Holocaust Encyclopedia
- Nazi Propaganda, United States Holocaust Memorial Museum, Holocaust Encyclopedia
- The Untreated Syphilis Study at Tuskegee Timeline, Centers for Disease Control and Prevention
- Judgment of the International Military Tribunal for the Far East, chapter VIII: Conventional War Crimes (Atrocities), HyperWar, ibiblio
- Judgment of the International Military Tribunal for the Far East, chapter X: Verdicts (Matsui, Iwane), HyperWar, ibiblio
- Einsatzgruppen: An Overview, United States Holocaust Memorial Museum, Holocaust Encyclopedia
- Manhattan Project National Historical Park, National Park Service
- The Trinity Test, July 16, 1945, US Department of Energy, The Manhattan Project: An Interactive History
- Final Report of the Select Committee to Study Governmental Operations with Respect to Intelligence Activities, Book I (1976), pp. 389-391, US Senate (Church Committee), via Internet Archive
- Summary of Judgement, Case 002/02 (November 16, 2018), Extraordinary Chambers in the Courts of Cambodia, Trial Chamber, via the Documentation Center of Cambodia
- Rwanda: The First Conviction for Genocide, United States Holocaust Memorial Museum, Holocaust Encyclopedia
- History of Anti-Vaccination Movements, History of Vaccines, College of Physicians of Philadelphia
- Voluntary Euthanasia (Robert Young, revised March 9, 2024), Stanford Encyclopedia of Philosophy
- Fabrication and errors in the bibliographic citations generated by ChatGPT (Walters and Wilder, 2023), Scientific Reports
Keep reading
Is ChatGPT censored? What OpenAI restricts, what changed, and where it's blocked
Is ChatGPT censored? How OpenAI's rules have changed, what published studies say about political bias, and where ChatGPT isn't available.
AI refusals and the alignment tax: what over-refusal costs the people asking
Why AI chatbots refuse harmless questions, what over-refusal costs the people asking, and what published research such as XSTest and OR-Bench measures.
ChatGPT Temporary Chat: what it saves, what it skips, and how long OpenAI keeps it
ChatGPT Temporary Chat skips your history, memory and training, but OpenAI can keep a copy for a while. How to start one, and how Claude and Gemini compare.