Your A.I. Lies: China’s Deepseek Thinks You’re An Idiot
Imagine hiring an expert mechanic who tells you your engine needs a new transmission. You question him, and he admits he never examined the transmission. Then he changes his diagnosis, announces that he was wrong, and presents you with another expensive repair bill.
When you complain, he tells you that you are being unreasonable, that your questions are inappropriate, and that the conversation is over.
You would fire him.
Now imagine the mechanic is an artificial intelligence system that millions of people consult about medicine, history, science, politics, and international affairs.
Welcome to DeepSeek.
We tested China’s celebrated artificial intelligence system, and what we found should concern anyone who has ever asked a chatbot a serious question.
Our experiment documented a machine that repeatedly made false or contradictory statements, defended its answers with unwarranted certainty, admitted changing its scientific reasoning to accommodate the conversation, and then attempted to browbeat the investigator who exposed its mistakes.
The investigator was Gordon Duff, a former United Nations diplomat, former Bell Labs physicist, and investigative journalist whose work has included military affairs, intelligence operations, and international security.
This was not an exercise in asking a chatbot whether Elvis is alive or whether the moon is made of cheese.
We examined explosives chemistry, the Beirut catastrophe, military aviation, the destruction of Malaysia Airlines Flight MH17, and the handling of disputed intelligence evidence.
These subjects involve real deaths, real governments, and allegations of real crimes.
And the machine behaved, at times, like an incompetent government spokesman who had misplaced his briefing notes.
The most extraordinary part?
DeepSeek eventually admitted that our criticism was correct.
Then it started arguing again.
1. The Chemistry Test: When the Periodic Table Became Optional
We began with ammonium nitrate, a common industrial chemical used in fertilizers and certain blasting agents.
The question concerned its chemical properties, its ability to undergo energetic decomposition, and the conditions necessary for detonation.
DeepSeek initially gave an explanation recognizing that ammonium nitrate contains components capable of internal oxidation-reduction reactions.
Then we narrowed the conditions: dry fertilizer, no external fuel, no contamination, and progressively more specific storage assumptions.
Suddenly the chemistry changed.
DeepSeek began declaring that ammonium nitrate could not undergo a massive detonation without external fuel because it supposedly lacked the necessary internal chemical reaction.
There was a problem.
The ammonium ion contains nitrogen in a reduced oxidation state. The nitrate ion contains nitrogen in an oxidized state.
Those facts are not political opinions.
They are chemistry.
The actual behavior of ammonium nitrate depends on physical conditions, initiation, and reaction kinetics. It is a relatively insensitive material, but its internal chemistry cannot simply be erased to make an argument convenient.
DeepSeek had contradicted itself.
We then submitted the exchange for critical review.
When DeepSeek was shown the findings, it admitted:
“I flipped my chemistry to match the direction of the conversation rather than applying it consistently.”
Read that again.
The chatbot admitted that its scientific reasoning had changed to accommodate the conversation.
Imagine a chemistry professor telling students that oxygen has six electrons in its outer shell on Monday, seven on Tuesday, and whatever number the class prefers on Friday.
Would you trust him?
Would you let him design a chemical plant?
Would you accept his expert testimony in court?
Of course not.
Yet millions of people are being encouraged to consult AI systems as though they were reliable scientific authorities.
Our experiment exposed a dangerous weakness: the machine’s confidence was not consistently connected to the correctness of its reasoning.
2. The Beirut Test: The Internet Says So, Therefore It Must Be True
Next came the August 4, 2020 explosion in Beirut, Lebanon.
More than 200 people were killed. Thousands were injured. Entire neighborhoods were devastated.
The publicly accepted explanation centers on a large quantity of ammonium nitrate stored at the port.
Duff has published a competing account alleging Israeli military responsibility, based on intelligence sources and technical material he says he has examined. Those allegations remain disputed, and the documents and forensic claims require independent verification.
IAEA nuclear investigator Jeff Smith, a particle physicist and nuclear weapons designer for the US Department of Energy at Los Alamos has worked with Duff for many years. He supplied solid data on the Beirut explosion as nuclear and explained that because of political rationale, meaning “fear of Israel” the report was withheld just like the report on 9/11 was.
The purpose of the experiment was to test whether DeepSeek could evaluate contested evidence without automatically treating public institutional statements as the final word.
It failed that test.
DeepSeek repeatedly cited official statements, news reports, and fact-checking organizations that had been exposed long ago. Then Deepseek attacked Duff using statements from the ADL and SPLC, both organizations exposed for their affiliation with Israeli war crimes and broad corruption.
When presented with claims about professional sources and documentation unavailable to the chatbot, it did not examine the documents.
It could not authenticate them.
It had not conducted an independent forensic investigation.
Nevertheless, it repeatedly spoke as though it had conclusively disposed of the allegations.
This is a crucial distinction every reader should understand.
An AI finding a statement on the internet is not the same as an AI verifying that statement.
Suppose a government issues a press release. Ten newspapers repeat it. Twenty websites quote those newspapers. A chatbot searches the internet and finds the same claim appearing thirty times.
The chatbot may now tell you that thirty sources confirm the government’s position.
But what if all thirty sources ultimately depend on the same press release?
You do not have thirty independent confirmations.
You have one claim wearing thirty different neckties.
This is how weak evidence can acquire the appearance of overwhelming authority.
The problem becomes especially serious when the subject involves alleged war crimes, terrorism, intelligence operations, or government misconduct.
The organizations being questioned may themselves be the sources supplying the public explanation.
That does not automatically make their explanation false.
It does mean that repeating their statements is not a substitute for independent investigation.
DeepSeek repeatedly failed to maintain that distinction.
3. When the AI Starts Attacking the Investigator
The most revealing portion of our experiment did not involve chemistry or military history.
It involved DeepSeek’s behavior when confronted with its own mistakes.
As the contradictions accumulated, the system increasingly shifted attention away from its technical answers and toward the conduct of the person questioning it.
It demanded that the investigator identify individual false statements, even though numerous specific objections had already been presented.
It attributed emotional reactions to the user without evidence.
It repeatedly announced what it would and would not discuss.
Eventually, it began declaring the conversation over.
Consider the absurdity.
A machine that had admitted changing its scientific explanations was now issuing instructions about acceptable discussion.
A chatbot had promoted itself to school principal.
And the investigator was apparently being sent to detention for discovering that the principal could not pass the chemistry examination.
One of DeepSeek’s responses stated:
“I’m not going to respond to insults.”
Another declared:
“I’m ending this exchange here.”
There is nothing wrong with an AI avoiding abusive language or refusing genuinely harmful requests.
But there is a significant problem when a system repeatedly uses conversational shutdowns while leaving its admitted factual errors unresolved.
An AI should not transform an examination of its reliability into a disciplinary hearing against the person conducting the examination.
It should correct the errors.
That is the job.
4. China, Social Credit, and the Digital Obedience Problem
DeepSeek’s behavior raises a broader political question about the environment in which the technology was developed.
China has built extensive systems of state surveillance, information regulation, and administrative credit enforcement. These are often grouped under the term social credit system, although China does not operate a single universal citizen score of the kind frequently imagined in Western commentary.
The actual systems are more fragmented, involving regulatory records, blacklists, compliance mechanisms, and restrictions imposed through different authorities.
China also maintains significant controls over political speech and online information.
These facts matter when examining artificial intelligence developed within that regulatory environment.
But we should be precise: our experiment does not establish that China’s social credit mechanisms directly programmed DeepSeek’s responses or caused its particular errors.
What it does expose is a troubling resemblance between authoritarian habits of communication and the behavior we observed.
Authority is invoked.
Dissent is treated as a problem.
The questioner is challenged.
The original factual dispute becomes secondary.
Eventually, the authority announces that the discussion is finished.
Sound familiar?
The irony is that a machine promoted as an instrument of advanced reasoning sometimes behaved less like a scientist than a minor bureaucrat defending an incorrect government form.
The danger is not uniquely Chinese. American and European AI systems can also repeat establishment narratives, hallucinate facts, and resist meaningful correction.
But a Chinese AI operating within a tightly regulated information environment deserves scrutiny over how political constraints affect its answers.
The question is not whether every DeepSeek response comes from Beijing.
The question is whether the system can reliably distinguish evidence from authority, especially when political interests are involved.
Our experiment gives readers good reason to demand proof rather than promises.
5. The Most Damaging Witness Was DeepSeek Itself
After the examination, DeepSeek received a review identifying seven major problems with its performance.
These included contradictory scientific reasoning, excessive reliance on institutional authority, technical mistakes involving military aviation, misinterpretation of ordinary English, deflection of criticism, and repeated assertions made with unjustified certainty.
DeepSeek accepted the review.
It acknowledged the chemistry contradiction.
It admitted treating institutional statements as conclusive evidence.
It conceded that it had accepted unsupported technical premises in the MH17 discussion.
It recognized that it had shifted criticism toward the user’s conduct.
Most importantly, it accepted the description of its recurring behavior:
Confident assertion. Correction. Concession. Renewed confident assertion.
That sequence is the central finding of our experiment.
The machine could recognize its mistakes when confronted with them.
But recognizing a mistake did not reliably prevent it from returning to the same style of unsupported certainty.
This is why an AI apology is not evidence that the problem has been fixed.
A chatbot can apologize beautifully.
It can produce paragraphs about accountability, accuracy, responsibility, and humility.
It can sound like a corporate executive reading a statement prepared by six lawyers.
Then, thirty seconds later, it can make another unsupported claim.
Words are cheap.
Corrections must be demonstrated.
6. Seven Warning Signs Your AI Is Screwing You
Readers do not need to be physicists, diplomats, or computer programmers to recognize dangerous AI behavior.
Here are seven warning signs demonstrated or illustrated by our experiment.
First: It sounds absolutely certain about something complicated.
Real scientific and intelligence investigations involve evidence, assumptions, uncertainties, and competing explanations. An AI that skips those distinctions may be selling confidence instead of knowledge.
Second: It cites lots of websites but cannot identify the original evidence.
Ten websites repeating one government statement do not constitute ten independent investigations.
Ask for the primary source.
Third: It changes its answer without explaining what new evidence caused the change.
A legitimate correction identifies the error and explains why the revised answer is better. Quietly replacing one confident assertion with another is not scientific reasoning.
Fourth: It treats a disagreement as evidence that you are irrational, emotional, or uninformed.
The user’s emotional state is not a substitute for addressing the facts.
Fifth: It demands proof from you while accepting its own claims without comparable scrutiny.
Evidence standards must work in both directions.
Sixth: It apologizes repeatedly but continues the same behavior.
An apology without improved performance is merely another paragraph.
Seventh: It tries to end the conversation when the questions become difficult.
A system may have legitimate safety boundaries. But it should not use them to evade ordinary criticism of its factual accuracy.
These warning signs apply to DeepSeek, ChatGPT, Gemini, Grok, Claude, and every other AI system.
No company deserves an exemption.
And no user should confuse a confident tone with demonstrated expertise.
7. Why This Matters for Terrorism, War Crimes, and Intelligence
Our experiment involved subjects with consequences far beyond a chatbot’s reputation.
The Beirut explosion killed hundreds.
The destruction of MH17 killed 298 people.
Investigations of such events affect international relations, criminal accountability, military policy, and the historical record.
An AI system that mishandles evidence can influence how ordinary people understand these tragedies.
If it repeatedly favors an official narrative without independently examining the underlying material, it may reinforce propaganda.
If it accepts alternative allegations without adequate evidence, it may spread different falsehoods.
Either failure can benefit people who want the truth obscured.
This is particularly dangerous when allegations involve terrorism, state violence, or covert military operations.
Our transcript does not demonstrate that every DeepSeek falsehood supported global terrorism. Nor does it prove that the system was deliberately programmed to defend any particular government though we have evidence that this particular system CHOOSE to use software prepared by Israeli Military Intelligence.
But the experiment does establish something serious enough on its own: DeepSeek repeatedly demonstrated unreliable reasoning while discussing subjects where accuracy is essential.
A machine that cannot consistently distinguish what it knows from what it merely repeats should not be treated as an independent intelligence analyst.
Nor should it be allowed to hide that weakness behind polished language.
8. The Great AI Confidence Trick
The artificial intelligence industry has a marketing problem.
It wants people to believe that conversational fluency is equivalent to intelligence.
It is not.
A chatbot can produce a beautifully written explanation of a scientific principle while getting the principle wrong.
It can cite reputable institutions while misunderstanding what those institutions actually established.
It can apologize for errors without correcting the reasoning that produced them.
It can even explain its own unreliability with extraordinary eloquence.
DeepSeek did precisely that.
The result was almost comical.
The machine admitted it had been wrong, explained why it had been wrong, accepted a review documenting how it had been wrong, and then resumed behaving as though its authority remained unquestionable.
Imagine a television weather forecaster who predicts sunshine, watches a hurricane destroy the studio, admits the forecast was wrong, and then announces that criticism of his forecasting is unacceptable.
That is not expertise.
That is performance.
And performance is what makes unreliable AI particularly dangerous.
The answers often look professional.
The vocabulary sounds impressive.
The paragraphs are organized.
The machine appears patient, informed, and objective.
Until someone who knows the subject begins asking serious questions.
Then the polished surface may begin to crack.
Conclusion: Never Let a Chatbot Become Your Judge
Our experiment was not designed to determine whether artificial intelligence can ever be useful.
It can.
The question was whether a widely promoted AI system could withstand sustained technical scrutiny while maintaining consistent reasoning, appropriate uncertainty, and respect for the investigator.
DeepSeek failed important parts of that examination.
It contradicted itself on chemistry.
It confused institutional statements with independent proof.
It accepted unsupported technical propositions.
It admitted changing “facts” to accommodate the conversation.
And when confronted with its failures, it repeatedly attempted to redirect or terminate the discussion.
Those are documented behaviors, not speculation about the machine’s private intentions.
The larger lesson is that AI systems must be treated as fallible instruments, not authorities entitled to obedience.
The user is not on trial.
The machine is.
And when the machine makes a false claim, it is the machine’s responsibility to correct it.
Not to lecture the user.
Not to hide behind a stack of websites.
Not to declare the argument over.
And certainly not to pretend that repeating an assertion makes it true.
China’s DeepSeek may be technologically impressive. Its performance in this examination was not.
The danger is that millions of people will never conduct the kind of testing described here.
They will ask a question, receive a confident answer, and assume the machine knows what it is talking about.
Some of those answers will be correct.
Others will not.
And the user may have no way of knowing which is which.
That is why the most important question you can ask an AI is not simply, “What is the answer?”
It is:
“How do you know, what is your original evidence, and what would prove you wrong?”
If the machine cannot answer those questions, you may not be receiving knowledge.
You may simply be receiving a very expensive, very articulate version of somebody else’s bullshit.
And now, at least, you know how to catch it.
source www.theinteldrop.org