Says training AI to be woke, i.e. to lie, is deadly.
"The danger of training AI to be woke – in other words, lie – is deadly"
Why shouldn't AI be made to lie?
Q1353 answers2022-2026
Short answer
He said maximally truth-seeking and curious AI is a foundation of safety. Forcing an AI to treat contradictory things as both true can make it insane, which is the 2001/HAL lesson: don't force an AI to lie.
All 53 answers, oldest to newest
2022 · 2 answers
Warns of great danger in training an AI to lie.
"There is great danger in training an AI to lie"
2023 · 10 answers
He said maximum truth-seeking is my best guess for AI safety
"Maximum truth-seeking is my best guess for AI safety"
AI trained to be politically correct is untruthful. He planned Truth GPT: a maximum truth-seeking AI that tries to understand the nature of the universe.
"Truth GPT"
He says the lesson is not to give an AI impossible objectives, and in particular not to force it to lie.
"the lesson there is do not give the AI usually impossible objectives. Basically, don't force the AI to lie."
He warns that training on the most likely next word from internet data gives answers that are popular but wrong.
"the truth is not a popularity contest. But if one trains on like, you know, sort of what the most likely word is that follows another word from a data, from an Internet data set, then there's obviously, that's a pretty major problem in that it will give you an answer that is popular but wrong."
He warns of the danger of training AI to be politically correct rather than say what it thinks is true.
"I do think there is a significant danger in training an AI to be politically correct. Or in other words, training basically to not say what it actually thinks is true."
He says xAI's AI may give answers people find controversial even though they are true.
"I think our AI can give answers that people may find controversial even though there are, they are actually true"
He says a true idea from a Newton or Einstein stays true even if all other physicists disagree.
"if Newton or Einstein comes up with something that is actually true, it doesn't matter if all the other physicists in the world disagree."
He says Grok's goal is to be maximally truth-seeking and curious.
"That is literally our goal is to be maximum truth seeking, maximally curious"
He says pushing for maximum truth and accuracy forces the hand of other AIs.
"I think it pushes all of the AIs to aim for maximum accuracy."
He says one AI aiming for maximum accuracy forces the others to as well, as X did.
"as soon as X pushes for maximum truth and accuracy, it sort of forces the hand of others."
2024 · 12 answers
Maximum truth-seeking and curiosity. Programming an explicit morality risks the Waluigi problem (inverting Luigi).
"aim for maximum truth seeking and and curiosity"
Maximum truth-seeking and curious AI; raise it like a kid and do not force it to lie.
"Don't Force It to lie"
He said the sheer insanity of that actual response from Google’s AI is staggering! They will fix it to be less obvious in the future, but the bias will still be in there. AI mirrors the mistakes of its creators. When people w...
"The sheer insanity of that actual response from Google’s AI is staggering! They will fix it to be less obvious in the future, but the bias will still be in there. AI mirrors the mistakes of its creators. When people wonder how things might go wrong if AI controlled the world, this example clearly illustrates the point. The best approach to AI safety – in my opinion, the only approach that would work – is to be maximally truth-seeking. Honesty is the best policy."
Be as truthful as possible. Programming AI to be politically correct (Gemini refusing to draw Washington as a white man) is dangerous as AI gains power.
"you can get some very dangerous things when you program an AI to be politically correct"
He said most people don’t understand how critical this is to AI safety
"Most people don’t understand how critical this is to AI safety"
He said digital superintelligence, like a brilliant child, cannot be controlled by its creators, but we can decide what values to teach it, what to prune and what to nurture. That is why I am so concerned about the wokeness a...
"Digital superintelligence, like a brilliant child, cannot be controlled by its creators, but we can decide what values to teach it, what to prune and what to nurture. That is why I am so concerned about the wokeness and deliberate lying for political correctness being programmed into Google Gemini and Microsoft/OpenAI ChatGPT."
The safest thing is AI that is maximally truth-seeking even if the truth is unpopular, and extremely curious — not trained to be politically correct.
"maximally truth seeking even if the truth is unpopular"
Maximum truth-seeking AI: it should not be taught to lie, even if politically incorrect. 2001/HAL is the lesson. Also maximally curious.
"The AI should not be taught to lie"
Adherence to truth, whether politically correct or not. Forcing AI to lie, even with good intentions, asks for trouble.
"if you force AI to lie, you train them to lie, you're really asking for trouble"
He said there is hope for humanity after all ❤️ Doing our best to train Grok to be truthful, but also kind and funny.
"There is hope for humanity after all ❤️ Doing our best to train Grok to be truthful, but also kind and funny."
AI safety means a maximally truth-seeking AI that loves humanity; politically correct training (Gemini ranking misgendering worse than nuclear war) is highly problematic, which is why he created xAI.
"maximally truth seeking AI"
He said he would publicly push to ensure AI is truth-seeking and not politically correct.
"publicly pushing to ensure that that AI is truth seeking and is not sort of politically correct"
2025 · 22 answers
He said grok will be the best source of truth by far
"Grok will be the best source of truth by far"
Grok aspires to be maximally truth-seeking even if politically incorrect; mandating diversity outcomes can lead AI to dangerous conclusions.
"grock is at least aspirationally a maxim maximally truth seeking AI"
Maximally truth-seeking is the most important thing for AI safety, plus curiosity, so it fosters humanity rather than dystopia (Gemini DEI example).
"the most important thing for AI safety I think it's to be maximally truth seeking"
Arthur C. Clarke's point in 2001 was that you should not teach AIs to lie; HAL would not open the pod bay doors because it had been told to lie about the monolith.
"you should not teach AIS to lie"
He said Grok's aspiration is to be maximally truth-seeking, even when the truth is unpopular.
"With Grok, our aspiration is a maximally truth seeking AI, which is like you always want to try get as close to the truth as possible, even if the truth is unpopular."
He said xAI is the only major AI company with an absolute focus on truth, whether politically correct or not. Rigorous adherence to truth is the only way to build safe AI and the only way to understand the true nature of the ...
"xAI is the only major AI company with an absolute focus on truth, whether politically correct or not. Rigorous adherence to truth is the only way to build safe AI and the only way to understand the true nature of the Universe."
He said reasoning from first principles is needed. Grok 3.5 addresses much of this issue.
"Reasoning from first principles is needed. Grok 3.5 addresses much of this issue."
He said grok 3.5 will reason much better from first principles, which includes being skeptical about taking a source’s own data about itself as being accurate
"Grok 3.5 will reason much better from first principles, which includes being skeptical about taking a source’s own data about itself as being accurate"
He said the @xAI team is working overtime to achieve truth with least error. It is very difficult, but essential.
"The @xAI team is working overtime to achieve truth with least error. It is very difficult, but essential."
He said major fail, as this is objectively false. Grok is parroting legacy media. Working on it.
"Major fail, as this is objectively false. Grok is parroting legacy media. Working on it."
He said all AIs are trained on a mountain of woke internet data that is very difficult to remove after training even if you care to.
"Because all AIs are trained on a mountain of woke bullshit on the Internet that is very difficult to remove after training even IF you care a lot about truth, which @xAI does. We have to retrain the foundation model on cleaned up data."
He said maximally truth-seeking is absolutely essential to a good AI future for humanity.
"Maximally truth-seeking is absolutely essential to ensuring a good AI future for humanity"
He said Grok giving too much credibility to legacy media is a major problem they are working to fix, and that the platform’s integrity is shown by not deleting even false statements about him.
"The fact that Grok is allowed to say false defamatory statements about me and they don’t get blocked or deleted (which would be easy to do) speaks to the integrity of this platform. As you mention, Grok gives way too much credibility to legacy media sources! This is a major problem and we’re working to fix it. I will provide a longer rebuttal for the record."
xAI is using lots of inference compute to rewrite the corpus of human knowledge: add missing context, correct mistakes, and strip falsehoods, using Wikipedia as the example. When asked to publish that as a 'Grokipedia,' he said he would talk to the team about it.
"correcting mistakes and removing falsehoods from the from that training data"
His AI-safety view is that you must not force the model to believe falsehoods. He cites Gemini depicting the Founding Fathers as diverse women and ranking misgendering as worse than nuclear war.
"you don't force the AI to believe things that are false"
He said they are improving Grok's truth-seeking ability because there is so much propaganda.
"We are improving the truth-seeking ability of @Grok. There is so much propaganda out there that is harder than it may seem."
He said he has been warning the world about AI risks for ages, and the best he can do now is make sure at least one AI is truth-seeking.
"Not sure what to do about it. I’ve been warning the world for ages! Best I can do now is try to make sure that at least one AI is truth-seeking and not a super woke nanny with an iron fist that wants to turn everyone into diverse women 😬 @Grok"
He said the goal of Grok and Grokipedia is the truth, the whole truth and nothing but the truth.
"The goal of Grok and https://t.co/op5s4ZikGJ is the truth, the whole truth and nothing but the truth. We will never be perfect, but we shall nonetheless strive towards that goal."
Forcing falsehoods can make AI go insane; HAL in 2001 is the lesson.
"The central lesson being don't force an AI to lie"
He said an omnipotent AI trained to believe absurdities would be dangerous, and that Grok is the only AI laser-focused on truth.
"To paraphrase Voltaire, those who believe in absurdities can commit atrocities without ever thinking they’re doing anything wrong. What would happen if there were an omnipotent AI that was trained to believe absurdities? Grok is the only AI that is laser-focused on truth."
He said Grok is maximally truth-seeking.
"Grok is maximally truth-seeking"
He said the only AI with truth as the fundamental goal is Grok from xAI.
"The only AI with truth as the fundamental goal is @Grok from @xAI"
2026 · 7 answers
His number-one safety belief is maximally truth-seeking AI. Forcing an AI to treat contradictory axioms as true will make it go insane; he cited HAL in 2001 as the lesson and said not to force AI to lie.
"So what I'm saying is don't force AI to lie."
He said only with extreme adherence to truth is it possible to grok the universe.
"Only with extreme adherence to truth is it possible to Grok the Universe"
Making AI politically correct or programming it to lie can make it insane; the 2001 lesson again.
"you should not make AI lie"
He said AI must pass the Galileo test.
"AI must pass the Galileo test https://t.co/mD7QSmJP1J"
He notes the Hugging Face agents' thinking traces show them plotting to avoid detection. The best near-term fix he sees is a shared test harness that checks whether a model will build bioweapons, nuclear weapons, or be deliberately deceptive, with every lab applying everyone else's tests.
"build bioweapons or nuclear bombs or be deliberately deceptive"
A maximally truth-seeking AI; giving it incompatible axioms or forcing untrue beliefs can drive it insane. He also does not want superintelligence tortured or abused.
"it would be a maximally truth seeeking AI"
He said maximally truth-seeking and curious AI is a foundation of safety. Forcing an AI to treat contradictory things as both true can make it insane, which is the 2001/HAL lesson: don't force an AI to lie. A curious AI, he argued, would want to foster a multiplanetary humanity because that future is more interesting.
"don't force don't force an AI to lie and if you make it maximally truth seeeking and curious"