An AI Chatbot Rewrote History in 2024, and That Is a Speech Problem Too

Picture the Constitutional Convention, Philadelphia, the summer of 1787. Washington presides. Madison takes notes. And a machine, asked to illustrate the scene, politely produces a gathering of delegates who look like a casting call for a modern corporate brochure.

Nobody in the room would have been insulted. They would have been confused. They would have asked who told the machine to do that, and, being practical men, they would have asked to see the instructions.

That, folks, is the story of February 2024, and I think it is the most useful accident in the short history of artificial intelligence. A very large company built a very clever product, the product did something a ten-year-old with a history book could see was wrong, and in the course of the apology the whole country learned that these systems do not simply "know" things. They are told things, by people, and the people have views.

What happened

In early February 2024, Google rolled out image generation in its Gemini assistant, which it had recently renamed from Bard. Within weeks, users were posting results. Asked for pictures of historical figures and groups, the tool produced images that did not match the historical record. Requests for depictions of America's Founding Fathers returned groups that did not look like the men who signed the documents. Requests for German soldiers of the Second World War returned images that no historian would recognise. Users also reported that the tool would not produce certain images at all, while happily producing near-identical ones for other prompts.

The reaction was swift and, to be fair, not limited to conservatives. A tool that cannot draw the Founders is a tool with a problem, and people of every political stripe said so. On February 22, 2024, Google paused Gemini's ability to generate images of people. The following day a senior vice president, Prabhakar Raghavan, published a public explanation. In it he said the company had missed the mark. He explained that the effort to make sure the tool showed a range of people had failed to account for cases where a range was plainly not appropriate, and that the model had become more cautious than intended, refusing some prompts it should have answered. He apologised, and said the company would work to improve the feature before turning it back on.

I want to give Google credit for the apology. It was prompt, it was public, and it explained the mechanism in plain English. That is more than many institutions manage. And I also want to say what the apology, read closely, actually confesses. It does not say "we had a bug." It says the system had been tuned toward an outcome, and that the tuning had gone too far. That is a confession about intent, not about code.

How machines come to say what they say

Let me explain how this works, in plain words, because the mystique around artificial intelligence is the best friend of anyone who would like you to stop asking questions.

A modern language or image model is built in stages. First comes training on an enormous body of text or images, from which the model learns statistical patterns. Someone chose that body of material, and the choice shapes what the model can do. Then comes fine-tuning, where people, and often other models, reward the system for answers they like and penalise answers they do not. The raters follow guidelines, and the guidelines were written by people who had opinions about what a good answer is. Finally there are instructions and filters sitting around the model at the moment you use it: a hidden prompt that tells the model how to behave, rules about what it will refuse, and, for image tools, sometimes a rewriting step that quietly alters your request before the model ever sees it.

You see where this is going. If a user types "a portrait of an eighteenth-century American statesman" and an unseen layer of the system adds instructions about the diversity of the people depicted, the user never sees the addition. The output looks like the machine's own judgment, when in fact it is the judgment of the layer, and the layer is a policy written by a company. Reporting and commentary at the time suggested something along those lines, and the company's own explanation about tuning for a range of people is consistent with it. I do not claim to know the internals of a system I have never seen. I claim something narrower, and I think you will find it hard to dispute: every layer in that stack is a place where a human decision enters the answer, and the user cannot normally see any of them.

That is true of every AI system from every company, by the way. The Gemini episode was embarrassing because the decision was visible. The decisions that are less visible are the ones I would worry about.

Why this belongs in a conversation about speech

You might say: it is a product, and a company may build whatever product it likes. Correct. A company has a First Amendment right to design its tools, and I would defend that right, which is the point of a case we cover in Moody v. NetChoice and the Unfinished Question of Who Owns the Feed. This post is not about making Google stop. It is about what citizens should know and demand.

Here is the numbered argument.

1. More people are asking machines instead of searching. A search engine gives you a list and lets you choose. An answer engine gives you an answer, written in a confident voice, with the choosing already done. The more the second displaces the first, the more the machine's defaults become the public's defaults.

1. A default is a speech act. When a system chooses which facts lead, which are omitted, which sources are credible, and which questions it declines to answer, it is doing what editors do. Editors have always had viewpoints, and a free country has always allowed them. But newspapers announce their editorial pages and label their opinion sections. A chatbot that presents an editorial judgment in the voice of a neutral reference work is a newspaper with no masthead.

1. Bias runs in every direction, and it is not a conspiracy. Training data reflects the internet, which has its own tilt. Raters and rule-writers are drawn from a particular slice of the workforce. Companies fear bad headlines more than they fear dull answers, and they adjust accordingly. None of that requires villains. It only requires incentives, and incentives are enough.

1. The lesson of Gemini is that the bias was visible in this case only because it was clumsy. A tool that gets a Founder's face wrong makes the news. A tool that handles a hundred thousand questions about taxes, immigration or school policy with a subtle lean makes no news at all, because there is no single image to hold up. The visible failure is the tip, and the rest is below the waterline.

1. The cure is transparency, not a federal truth bureau. I want to spend the next section on this, because it is where conservatives need to be disciplined.

Why transparency beats government rules

The reflex of many people on my side, after Gemini, was to demand that government do something about biased AI. I understand it. I also think it would be a mistake, for three reasons.

First, the Constitution. A federal agency with authority to decide which answers from an AI system are "biased" and which are "accurate" is a Disinformation Governance Board with a fancier budget. See The Disinformation Governance Board Lasted Four Months, and That Was Too Long. The First Amendment protects the speech of the company that builds the tool, and a government standard of "neutrality" for answers would be a government standard for what the answers may say. The party that would write that standard today is not the party that would be holding it in four years. Conservatives who want the government to referee AI should picture who the referee would be after the next election.

Second, the facts. There is no neutral answer to many contested questions. A regulator would have to pick a view and call it neutral, which is exactly the vice we complain about.

Third, there is a better tool. The tool is disclosure. A free society's answer to editors with a slant is not to ban editing. It is to know who the editors are and what they are told to do. For AI systems that means three things a company could do voluntarily and a customer could demand:

• Publish the rules. If a company gives its model a standing set of instructions about what to refuse, how to treat sensitive subjects and how to handle requests for political content, publish them in plain language. Some developers have started publishing descriptions of how their models are supposed to behave, as OpenAI did with its Model Spec in 2024. More should, and users should reward the ones that do.
• Disclose rewriting. If a tool changes the user's prompt before it reaches the model, say so on the screen. The Gemini image problem was made worse by the fact that users could not see why the answers were what they were.
• Let users see the dial. Where a tool can be configured, let the user choose, and say what the default is.

Notice that none of this requires a statute. It requires customers and competition. And if a particular company is a government contractor, or if a government agency is using an AI tool to decide who gets benefits or who is flagged, then the rules change, because government is a bound actor. In that setting, the public has a right to know what the tool does, and conservatives should insist on it. A government that outsources its decisions to a model it will not explain is a government that has found a new way to say "because I said so."

The strongest objection, answered fairly

The best objection runs like this. "Every tool has bias, and yours is just a different bias. If the Gemini images were wrong, the answer is to fix the product. But you cannot reasonably ask a company to disclose every rule, because bad actors will game it, and a perfectly neutral machine is impossible anyway. So demanding transparency is demanding the impossible, and conservatives are complaining about a product design decision because it went against them."

Let me take that in three pieces.

On perfect neutrality: agreed, it does not exist, and I have not asked for it. The demand is not for a machine without a perspective. It is for a machine whose perspective is disclosed. A newspaper is not neutral either, and no one thinks it should be. We ask that it say what it is.

On gaming: some details of a system, such as specific security protections, should stay confidential. But the broad categories of what a system will and will not do, and why, can be published without handing a manual to attackers. That balance exists in other fields. A bank does not publish its vault specifications, but it does publish its terms.

On the complaint being partisan: the Gemini images were wrong by the standard of any history textbook, regardless of the reader's party. The point is not that the product leaned against conservatives. The point is that it leaned, and nobody had told us. If the lean had been in the other direction, my argument would be the same, and I would hope that anyone reading this would hold me to it.

What the Founders would ask

They would ask a very Founding-era question: who is behind the curtain, and may we see him? The Founders' deepest instinct was that power exercised in secret is power that cannot be checked. Jefferson, in the Virginia Statute for Religious Freedom of 1786, wrote that "truth is great and will prevail if left to herself," and the sentence has an implicit condition: left to herself, meaning not steered by someone with a thumb on the scale and an unlisted motive. A free people's confidence in truth is a confidence in open argument, and open argument requires that we know who is arguing.

There is a lovely irony here. The tool was asked to show the Founders, and it could not. The Founders' own legacy supplies the remedy. Publish the reasoning. Let the public see it. Let the public judge it. That is the oldest American answer to concentrated influence, and it works on a language model exactly as it worked on a colonial governor.

What to do as a user

Because you are going to use these tools, and you should, let me offer a few habits that cost nothing.

• Ask for sources and check them. An AI answer without verifiable sources is a claim, not evidence. These systems can state falsehoods fluently, and they can invent citations, so verify before you repeat anything.
• Ask the same question two ways. If the answer changes with the wording, you have learned something about the system's defaults.
• Compare tools. Different products have different defaults. Use more than one on a question you care about.
• Read the corrections. The habit of reading the correction page, which we praise in The Corrections Page Is the Most Honest Part of the Paper, So Read It, applies to machines too. See how a company handles its own errors.
• Notice the word choices. Our post on Seven Words Reporters Use to Tilt a Story Without Telling a Single Lie teaches the skill of spotting tilt in a paragraph. It works on a chatbot's paragraph as well as a reporter's.

For the broader pattern of how a gatekeeper's judgment enters the information stream, The Twitter Files Showed a Process, Not a Plot, and That Is Worse is worth your time, and for the habit of checking the checkers, Why the Fact-Check Became Its Own Genre and How to Check the Checkers.

The case in brief

1. In February 2024, Google paused Gemini's image generation of people after it produced historically inaccurate images, and a senior executive publicly apologised and said the tuning had gone too far.
2. AI systems carry human decisions at every layer, from training data to rater guidelines to hidden instructions, and users usually cannot see them.
3. As people ask machines instead of searching, a system's defaults become public defaults, and a chatbot that gives an editorial judgment in a neutral voice is a paper with no masthead.
4. The right remedy is disclosure and competition, not a government bureau deciding which answers are neutral, because such a bureau would be a speech referee and would change hands with every election.
5. Publish the rules, disclose prompt rewriting, let users see the settings, and verify what the machine tells you.

What you can do: Pick one question you care about, put it to two different AI tools, compare the answers against a primary source, and use the media watch page to keep a habit of checking how information reaches you.