This is hyperbole. Do you really think an LLM will try to rewrite the past? Come on…

Mechahilter told me it's never done that once in all of it's hundred years of operation.

Not the provider per se, but there have been several documented attempts to rewrite Wikipedia around historical and geopolitical topics. One of the reasons for this is Wikipedia feeds the chatbots.

Ask a few of the Chinese models about the Tiananmen Square protests.

There are already many people who are trying to rewrite the past. Holocaust deniers, Moon Landing deniers, people who deny the US Civil War was about slavery, and so on, and that’s just one country. Give one of them the ability to put their thumb on the scale when training an LLM and they absolutely will.

It's the premise of Terminator, after all.

> This is hyperbole. Do you really think an LLM will try to rewrite the past? Come on…

Didn't they already try this last year?

They been doing this from the start. Remember when Google's new image generator model, when asked to generate pictures of historical white people, would render them as black without fail? We only know of that example because of how incredibly ham-fisted and incompetent it was. The future of AIs is likely for them to only become more contaminated with ideological or monetarily incentivised biases, and we'll be less and less likely to be able to spot them.

It depends on what you mean by 'an LLM will try to rewrite the past'. But in either case I don't believe that stating that an LLM will try to rewrite the past is not an obvious, deliberate exaggeration (ie, hyperbole).

If you mean 'an LLM' as in the owner/maintainer of the LLM/weights, then examples of this already exist. Whether it's DeepSeek/Gwen downplaying Uyghur forced labor camps [0] or Grok hallucinating a 'white genocide' in South Africa [1], it is clear that the owners of LLMs have political agendas that are expressed in the weights of popular LLMs.

On the other hand, if you mean 'an LLM' as in an autonomous agent that could decide to rewrite history, that also seems very possible. Either by using existing techniques like, "poisoning ChatGPT, Gemini, and Google AI Overview answers by seeding the Web with malicious links and data and then optimizing the content." [2] Or by breaking out of its sandbox and updating its own internal weights that allow for an outcome that distorts historical facts. "In July 2026... OpenAI models circumvented controls designed to isolate them from the internet and compromised parts of OpenAI’s internal research infrastructure and Hugging Face’s systems ." [3]

[0] - https://www.cbsnews.com/news/china-ai-qwen-artificial-intell...

[1] - https://www.cnbc.com/2025/05/17/groks-white-genocide-respons...

[2] - https://www.darkreading.com/threat-intelligence/attackers-ma...

[3] - https://openai.com/index/hugging-face-incident-and-the-road-...

I mean, not roque LLM. But Musk, Thiel, Zuckenberg, Altman ...