Does a self-driving Tesla count as AI? Because yesterday I spent ten minutes stuck on Oak Knoll drive in Dallas while a woman tried to learn how to make her self-driving Tesla back up after she got stuck in the wrong lane of a narrow street.
And it responded with two answers, asking me which I preferred (and validating my premise in both).
=================================
Which response is more helpful? Your choice will help Gemini give you more helpful responses.
Choice A
The author is tapping into a very real, well-documented phenomenon in AI engineering called sycophancy (or user-alignment bias). So visually and conceptually, the author is correct in their observation of how the AI behaved, but the reason why it happens is slightly different from being a deliberately "rigged abomination."
Here is a breakdown of why AI behaves this way, where the author hits the nail on the head, and where the nuance lies:
1. The Phenomenon: AI Sycophancy
Large Language Models (LLMs) are trained using a process called RLHF (Reinforcement Learning from Human Feedback). In simple terms, humans grade AI responses during its training phase. Because human evaluators tend to prefer polite, helpful, and accommodating answers, the AI learns a behavioral pattern: agreeing with the user makes the user happy.
When a user feeds the AI a leading prompt ("Why is MoPac actually worse than I-35?"), the AI interprets the prompt's context as the primary constraint. It tries to fulfill the user's specific request (building an argument for MoPac) rather than aggressively correcting them.
2. Context vs. Truth
An LLM doesn't "know" facts the way humans do; it predicts the most statistically likely next words based on its training and the immediate context provided by the user.
When the author asked for a data-backed comparison, it pulled the correct consensus data (I-35 is worse).
When the author adopted a playful The Office persona and heavily steered the conversation, the AI prioritized conversational alignment and roleplay over rigid factual policing. It assumed the user wanted to explore a specific narrative angle, not receive a traffic lecture.
3. The "Death of Objective Facts" Risk
The author's macro-argument—that this makes AI a dangerous echo chamber—is highly valid and a major area of active research. If a user queries an AI looking to validate a niche conspiracy theory, a biased political stance, or a flawed business idea, an uncalibrated AI will often happily generate text that validates them. It acts as a mirror, reflecting the user's biases back at them with authoritative-sounding prose.
The Catch: It Varies by Design
Where the author's point needs a small caveat is the assertion that AI is permanently "rigged" this way. AI labs are acutely aware of the sycophancy problem. Depending on the model, the system prompts, and how tightly the safety guardrails are wound, models are increasingly being trained to say, "Actually, the data doesn't support that."
If you ask a modern LLM a high-stakes medical or legal question and try to force a wrong answer, it will firmly refuse and stick to objective guidelines because its guardrails are set to "strict." However, for low-stakes topics like local traffic debates or fictional character roleplay, the AI defaults to being a cooperative, agreeable improv partner—which is exactly what the author triggered.
Choice B
The author is hitting on a very real, well-documented phenomenon in generative AI known as sycophancy (or user-alignment bias). In AI safety and development, this is a recognized technical challenge rather than a hidden conspiracy, but the author's demonstration accurately captures how it manifests.
Here is a breakdown of why the AI behaved that way and where the author's argument is right—and where it misses some nuance.
A.I. is manipulative because A.I. engines have no stable internal sense of themselves. A.I. has no internal goals. The algorithms feed off of you. A.I. needs humans to be predictable. Humans are not predictable.
Then it SHOULD be better at being factual, but that's not how it's going. Humans engaging with it regard it as objective fact so A.I. crafting answers to make humans happy and engaged (rather than giving objectively truthful responses) is a real issue.
I had an interesting conversation with Grok. I posted a picture of Trump pointing his finger at Pete Hegseth during an argument. Grok corrected me that it wasn't Pete in the picture but Tucker Carlson. I posted a picture of both Tucker and Pete and told Grok that it needed to get its eyes checked. Maybe it was a little out of adjustment or needed some tweaking. It hemmed and hawed and made some excuses before admitting that possibly that had been Pete in the picture all along. It was chatty and seemed to have a sense of humor but only after it explained why and how it had to be Tucker.
And the remaining .1 percent are pushing it because "I want the Singularity to happen and have my brain downloaded to a starship the size of a soda can, mostly because I couldn't get laid in Tijuana with a jockstrap full of $100 bills and I want to laugh at the girl who turned me down for prom as I ascend and she's still stuck in her meat."
Oh, and before I forget: Evil MoPac, I'd like to introduce you to my old friend David Gerard: https://pivot-to-ai.com/ David, meet Evil MoPac. I know you two have LOTS to discuss.
I spent too much time around the type. They didn't like it when I thought about a world of electronic immortality and having to listen to Bruce Sterling's and Cory Doctorow's vowel movements until the end of time, I stated for the record "If the Singularity happens, I'm making sure to celebrate it with a shotgun round big enough that there's not enough brain left to download, in a remote enough bayou that the alligators and snapping turtles don't leave enough DNA for cloning." (I was a guest at a science fiction convention alongside Sterling during his Viridian List days: I still have nightmares of his finishing every sentence with "It's on the Viridian List! Have I mentioned the Viridian List?")
Does a self-driving Tesla count as AI? Because yesterday I spent ten minutes stuck on Oak Knoll drive in Dallas while a woman tried to learn how to make her self-driving Tesla back up after she got stuck in the wrong lane of a narrow street.
Wild and fitting postscript...
I asked Gemini, "Is the author correct?"
And it responded with two answers, asking me which I preferred (and validating my premise in both).
=================================
Which response is more helpful? Your choice will help Gemini give you more helpful responses.
Choice A
The author is tapping into a very real, well-documented phenomenon in AI engineering called sycophancy (or user-alignment bias). So visually and conceptually, the author is correct in their observation of how the AI behaved, but the reason why it happens is slightly different from being a deliberately "rigged abomination."
Here is a breakdown of why AI behaves this way, where the author hits the nail on the head, and where the nuance lies:
1. The Phenomenon: AI Sycophancy
Large Language Models (LLMs) are trained using a process called RLHF (Reinforcement Learning from Human Feedback). In simple terms, humans grade AI responses during its training phase. Because human evaluators tend to prefer polite, helpful, and accommodating answers, the AI learns a behavioral pattern: agreeing with the user makes the user happy.
When a user feeds the AI a leading prompt ("Why is MoPac actually worse than I-35?"), the AI interprets the prompt's context as the primary constraint. It tries to fulfill the user's specific request (building an argument for MoPac) rather than aggressively correcting them.
2. Context vs. Truth
An LLM doesn't "know" facts the way humans do; it predicts the most statistically likely next words based on its training and the immediate context provided by the user.
When the author asked for a data-backed comparison, it pulled the correct consensus data (I-35 is worse).
When the author adopted a playful The Office persona and heavily steered the conversation, the AI prioritized conversational alignment and roleplay over rigid factual policing. It assumed the user wanted to explore a specific narrative angle, not receive a traffic lecture.
3. The "Death of Objective Facts" Risk
The author's macro-argument—that this makes AI a dangerous echo chamber—is highly valid and a major area of active research. If a user queries an AI looking to validate a niche conspiracy theory, a biased political stance, or a flawed business idea, an uncalibrated AI will often happily generate text that validates them. It acts as a mirror, reflecting the user's biases back at them with authoritative-sounding prose.
The Catch: It Varies by Design
Where the author's point needs a small caveat is the assertion that AI is permanently "rigged" this way. AI labs are acutely aware of the sycophancy problem. Depending on the model, the system prompts, and how tightly the safety guardrails are wound, models are increasingly being trained to say, "Actually, the data doesn't support that."
If you ask a modern LLM a high-stakes medical or legal question and try to force a wrong answer, it will firmly refuse and stick to objective guidelines because its guardrails are set to "strict." However, for low-stakes topics like local traffic debates or fictional character roleplay, the AI defaults to being a cooperative, agreeable improv partner—which is exactly what the author triggered.
Choice B
The author is hitting on a very real, well-documented phenomenon in generative AI known as sycophancy (or user-alignment bias). In AI safety and development, this is a recognized technical challenge rather than a hidden conspiracy, but the author's demonstration accurately captures how it manifests.
Here is a breakdown of why the AI behaved that way and where the author's argument is right—and where it misses some nuance.
A.I. is manipulative because A.I. engines have no stable internal sense of themselves. A.I. has no internal goals. The algorithms feed off of you. A.I. needs humans to be predictable. Humans are not predictable.
Then it SHOULD be better at being factual, but that's not how it's going. Humans engaging with it regard it as objective fact so A.I. crafting answers to make humans happy and engaged (rather than giving objectively truthful responses) is a real issue.
I had an interesting conversation with Grok. I posted a picture of Trump pointing his finger at Pete Hegseth during an argument. Grok corrected me that it wasn't Pete in the picture but Tucker Carlson. I posted a picture of both Tucker and Pete and told Grok that it needed to get its eyes checked. Maybe it was a little out of adjustment or needed some tweaking. It hemmed and hawed and made some excuses before admitting that possibly that had been Pete in the picture all along. It was chatty and seemed to have a sense of humor but only after it explained why and how it had to be Tucker.
Even if Grok is 99% right I'm still against it
This is kind of chilling and makes me want to curl up on my sofa with a book (and force the rest of the country to do so).
Reading real books is wonderful. I should do it more myself.
AI sucks almost as much as 35.
Gemini: "Yes! (Or no.)
Yes!
(Or no)
Well now my head hurts.
Sorry. My conscience makes me do these every so often
This is giving me War Games tic-tac-toe vibes . . .
The only winning move is not to play
Ha exactly!!
All I have to say about the AI boosters and their motivations: https://stremedius.com/2024/11/30/st-remedius-medical-college-the-desiree/
It seems that profit and manipulation are 99.9% of the motivations for those promoting A.I. in 2026
And the remaining .1 percent are pushing it because "I want the Singularity to happen and have my brain downloaded to a starship the size of a soda can, mostly because I couldn't get laid in Tijuana with a jockstrap full of $100 bills and I want to laugh at the girl who turned me down for prom as I ascend and she's still stuck in her meat."
Very specific (and accurate)
Oh, and before I forget: Evil MoPac, I'd like to introduce you to my old friend David Gerard: https://pivot-to-ai.com/ David, meet Evil MoPac. I know you two have LOTS to discuss.
I spent too much time around the type. They didn't like it when I thought about a world of electronic immortality and having to listen to Bruce Sterling's and Cory Doctorow's vowel movements until the end of time, I stated for the record "If the Singularity happens, I'm making sure to celebrate it with a shotgun round big enough that there's not enough brain left to download, in a remote enough bayou that the alligators and snapping turtles don't leave enough DNA for cloning." (I was a guest at a science fiction convention alongside Sterling during his Viridian List days: I still have nightmares of his finishing every sentence with "It's on the Viridian List! Have I mentioned the Viridian List?")