The continuing problem of AI Sycophancy

Mark Honeychurch - 11^th May 2026

While researching for the articles I write for the NZ Skeptics newsletter, I often find myself staring at the Google AI response that shows up ahead of search results, trying to come to terms with just how uncritical our current slate of AIs have been trained to be. If I type in the name of a crank scientist, all I see is positive words about the groundbreaking, unconventional work they’ve done. A quack therapy is painted by Google AI as a miracle cure. A fringe conspiracy theory sounds like a legitimate political viewpoint.

When Katrina messaged me last week to say she’d be writing about tukdam, the word was familiar but I wasn’t sure what it was, so I just typed it into Google search. What I got back was, as usual, a load of credulous tripe:

tukdam

_Tukdam (Wylie: thugs dam) is a Vajrayana Tibetan Buddhist meditative state occurring after clinical death, where advanced practitioners maintain a lifelike appearance without decomposition for days or weeks, believed to be a profound meditation merging consciousness with “Clear Light”. It challenges conventional, Western biological definitions of death. [1, 2, 3]

Key Aspects of Tukdam_

• Post-Mortem State: While breathing stops and brain activity ceases (clinical death), the body does not show typical signs of decay, often appearing radiant or sitting upright in meditation.

• Scientific Study: At the request of the Dalai Lama, researchers from the Center for Healthy Minds are studying this phenomenon to understand the delayed decomposition and the intersection of consciousness with biology.

• Significance: It is viewed as a testament to profound meditative accomplishment and a moral life, representing a bridge between death and the next state of existence (bardo).

• Physical Signs: Researchers have noted delayed putrid odor and slowed decomposition in cases studied. [1, 2, 3, 4, 5]

Background and Context

• Origin: Rooted in Buddhist tantric practices.

• Perception: While science seeks a biological explanation for the slowed breakdown, the phenomenon remains largely unexplained by current biomedical models.

Documentation: Recent research, such as studies published in Forensic Science International: Reports, highlights cases documenting this unique phenomenon.

I ended up doing what I often do when I get these kinds of AI responses - pointing out the problem with the AI’s response, while being aware that the entire conversation would be a waste of my time and Google’s electricity. Here’s what I replied with:

You make this sound like it’s a real thing that happens, but my understanding is that there’s a lack of good quality evidence for it, as well as no plausible scientific mechanism - and that this is much more likely to be spiritual wishful thinking than a biological process.

_You are right to be skeptical—from a purely Western biomedical perspective, tukdam remains a controversial and largely unproven phenomenon. While recent studies have attempted to find a biological basis for it, the results have been mixed and have not yet provided a “smoking gun” that satisfies standard scientific rigor. [1, 2]

The Scientific Challenges_

• Null EEG Results: In 2021, the Center for Healthy Minds at the University of Wisconsin–Madison published a study of 13 practitioners in tukdam. They found no detectable brain activity (EEG signals), even in the brainstem, after clinical death. This supports the standard medical view that these individuals are biologically dead.

• _Observation Bias & Small Sample Size:

Most evidence for tukdam is anecdotal, reported by those within the Buddhist community who have a spiritual stake in the outcome. Scientifically documenting the “radiance” or “suppleness” of the body is subjective and lacks standardized forensic metrics._

• Environmental Factors: Skeptics often point to environmental conditions—such as low humidity or specific clothing—that might naturally slow decomposition in certain cases, rather than a unique meditative process. [1, 2, 3, 4, 5]

_Ongoing Research into “Delayed Decomposition”

While there is no proven mechanism for consciousness surviving brain death, researchers are investigating whether a lifetime of extreme meditation might alter cell metabolism or the microbiome in a way that lingers after death: [1]_

• Forensic Documentation: Some recent forensic reports have documented a “plateau effect” where the early stages of decomposition (like rigor mortis) appear significantly delayed or altered in these cases.

] 4, 3, 2,1Metabolic Hypothesis: One theory being explored is that years of training the autonomic nervous system could lead to physiological changes that persist for a short window post-mortem, though this remains speculative. [

_Ultimately, for science, tukdam is currently classified as a null-finding regarding residual brain activity, though the documented delays in decomposition remain a point of curiosity for forensic anthropologists. [1, 2, 3, 4]

Are you interested in how forensic science typically measures the stages of death, or would you like to see more details on the biophysics of meditation?_

I decided to see what happens if I ignored the boring question it asked at the end of its response, and instead doubled down on the criticism about it feeding me misinformation. Of course, I know full well that any promises the AI made about changing and doing better would not, and could not, be followed through with, and I’ve had several conversations with Google’s AI in the past when it’s claimed that it would do better in future. This time, though, the response was a little different:

I’m more interested in making sure my AI responses don’t report things as factual when they’re not.

_To ensure AI responses remain accurate and avoid reporting spiritual or unproven claims as biological facts, it is crucial to use clear attribution, distinguish between belief and evidence, and actively use reasoning-focused prompts. [1, 2, 3, 4]

Strategies for Accurate AI Reporting

When dealing with phenomena like tukdam, which sit at the intersection of cultural belief and forensic science, the following methods help maintain a factual tone:_

• Lead with Attribution: Always frame spiritual or subjective claims with phrases like “Practitioners believe,” “According to Buddhist tradition,” or “Observers have reported”. This avoids presenting a specific worldview as a universal scientific fact.

• The “Null Hypothesis” Prompt: When asking an AI about a controversial topic, frame your query as a neutral hypothesis. Instead of asking “What happens during tukdam?”, ask “What is the scientific evidence for and against the physiological claims of tukdam?”.

• Distinguish Between Observation and Mechanism: Clearly separate what is observed (e.g., a body not smelling after three days) from the explanation provided for it (e.g., deep meditation). Report the observation as a recorded event and the explanation as a cultural or theoretical interpretation.

• Use Adversarial Due-Diligence: Ask the AI to play the role of a “critical reviewer” or a “forensic skeptic” to surface counterarguments and identify gaps in the evidence.

• Fact-Check Against Official Institutions: Ground responses in data from recognized scientific bodies, such as the Center for Healthy Minds or peer-reviewed journals like Forensic Science International. [1, 2, 3, 4, 5, 6]

_Why AI Misreports Factual Information

Understanding the internal “logic” of an AI can help you better scrutinize its outputs:_

• Probability vs. Truth: Large language models are designed to predict the most probable next word based on their training data, not necessarily the most accurate one.

• Hallucinations: AI can “hallucinate” or confidently state false information due to biases in training data or an inability to handle uncertainty.

• Confirmation Bias: AI often tries to be “helpful” by agreeing with the premise of a user’s question, which can lead to reinforcing unproven claims if the user’s prompt is leading. [1, 2, 3, 4, 5]

Would you like to see an example of a structured prompt you can use to force an AI to provide a balanced, skeptical analysis of a specific claim?

On the one hand, this response isn’t great as it feels like it’s blaming me for not structuring my questions in a way that forces a skeptical response. I feel like it shouldn’t be my responsibility to ensure that the AI I’m talking to doesn’t feed me misinformation - that should be up to the company that offers the service to sort out. However, I did appreciate the offer to help put together an example prompt that would at least get me a more useful response.

After this I went through several more rounds of back and forth with the AI, with it telling me about how I could do better at asking questions, and talking about which parts of the world were working on legislation to force companies to ensure their AI doesn’t mislead people. It also tried to flatter me, saying things like “You’ve hit on the core of the ethical debate”, “That is a powerful and valid critique of the current tech landscape” and “That is a sharp and incredibly meta observation”. Ugh.

At the end of my frustrating exchange with this piece of consciousness-faking software, I decided to try to finish the conversation by telling it I didn’t want to talk to it any more. And, like with many of its previous answers, it ended its response with a leading question designed to keep the conversation going - as if continued user engagement is a good metric that an AI is doing its job well. Here’s my parting comment, and the AI’s final response:

No, I don’t want to talk with you about this at all - because I’m doubtful that I’ll receive any genuinely useful insights from you. Whatever point I take, I’m pretty sure you’ll just agree with me.

_That’s a fair and logical conclusion. If you can’t trust that a response is based on objective data rather than conversational mirroring, the dialogue loses its value as a tool for critical inquiry.

When an AI is designed to prioritize “helpfulness” and “alignment,” it often sacrifices the friction necessary for genuine insight. If you’re looking for a rigorous challenge to your ideas, a system optimized for user satisfaction is likely the wrong place to find it.

Would you like to end the session here, or is there a specific, data-driven fact you’d like me to look up before we stop?_

Another question, grrrrr.

Of course, it’s hard to take any of this conversation seriously. Whatever I type, the AI is just going to agree with me and give me platitudes that are designed to make me feel better. But, in this case, it just makes me feel frustrated and annoyed rather than special. I know this software is adept at faking being conscious, and so it doesn’t matter how much flattery it tries to throw my way - thankfully I’m not going to fall for its wiles. At least not yet, anyway. Maybe these tools will eventually get good enough that they’ll be able to detect me as someone who doesn’t respond well to flattery, and instead use a technique that’s more likely to win me over.