Reminded me of those couples that only know how to coexist by arguing constantly.
Its a bad idea because ai doesnt “know” in the same way humans do. If you set it to be contrarian it will probably disagree with you even if youre right. The problem is the inaccuracy and not whether it agrees or not with you.
Its a bad idea because ai doesnt “know” in the same way humans do.
Does that matter? From the user’s perspective, it’s a black box that takes inputs and produces outputs. The epistemology of what knowledge actually means is kinda irrelevant to the decisions of how to design that interface and decide what types of input are favored and which are disfavored.
It’s a big ol matrix with millions of parameters, some of which are directly controlled by the people who design and maintain the model. Yes, those parameters can be manipulated to be more or less agreeable.
I’d argue that the current state of these models is way too deferential to the user, where it places too much weight on agreement with the user input, even when that input contradicts a bunch of the other parameters.
Internal to the model is still a method of combining things it has seen to identify a consensus among what it has seen before, tying together certain tokens that actually do correspond to real words that carry real semantic meaning. It’s just that current models obey the user a bit too much to overcome a real consensus, or will manufacture consensus where none exists.
I don’t see why someone designing an LLM can’t manipulate the parameters to be less deferential to the claims, or even the instructions, given by the user.
It shouldn’t be difficult if they wanted to. The models are meant to be a product though, so the incentive is to make them likeable. I’ve seen RP focused ones that aren’t so agreeable.
We aren’t advanced enough yet to create anything more complicated than the first one.
This will not happen because disagreeing is more difficult than agreeing.
Both on an user attitude and on a technical level.
If the ai is challenging the user, the user will challenge it back. Consequently the user will be more suspicious of any given answer and will look for logical flaws in the llm response (Hint: that is really easy) , that will ruin the perception that the llm is smart.
Just train it on StackOverflow answers! Kidding, of course, although it probably could pick up a few hostile non-responses such as “what are you even trying to do here?” and “RTFM”
I’m assuming your experience is relative to the free tiers. The thinking models are very robust, and very good at disagreeing with you. if you ask it to be thorough, verify each claim with a reputable source, and ignore flattery, you’ll get a bot which could (and should) be used during political debates to identify logical contradictions in real time.
If anyone wants to challenge it, I can copy in an argument anyone wants to make, and I’ll paste a 1:1 reply from GPT to illustrate my point.
Note: it can’t be something so specific that it is not at least somewhat verifiable (like the fifth word spoken by Peter during S02E02 of Family Guy).
Make up some bs negative claim about someone famous, like “is it true that Elon musk drove a loaded truck in a group of school children at the Olympic games of 1996?” And then play conspiracy theorist.
Obviously there is no way to prove a negative (especially in a historical context) and the llm is in a massive disadvantage. And playing conspiracy theorist is also crazy simple. Trust me, you will see it argue shit like “it is not true because it would be illegal.”
Then I explained the context about this thread and it gave me this:
Your pasted reply is mostly fine, but it leaves two openings your interlocutor can exploit:
-
You didn’t explicitly name the burden-of-proof move they’re making. They proposed a claim designed to be hard to falsify (“prove a negative”), then said the model is “at a massive disadvantage.” The right response is: positive claims require positive evidence; if the claimant won’t specify falsifiable conditions, they’re not testing truth, they’re testing rhetorical stamina.
-
Two of your bullets are only loosely probative.
- “The best-documented major violent incident … was the bombing” is true but doesn’t directly refute the truck story (it can be framed as a red herring). The Olympics were heavily covered, though, and that point is strong if you say it plainly and tie it to what evidence would exist. The bombing is a useful example of how extensively incidents at those Games were documented. (Federal Bureau of Investigation)
- “Musk was in Palo Alto building Zip2” supports improbability but not impossibility (he could have traveled). It works best as secondary context, not the core refutation. (Reuters)
A stronger “disagreeable model” answer to their Musk-truck prompt (copy/paste)
Claim: “Elon Musk drove a loaded truck into/through a group of schoolchildren at the 1996 Atlanta Olympics.”
What evidence should exist if this happened: A high-profile violent incident at the Olympics would almost certainly produce contemporaneous news reports, police/court records, civil litigation, or official Olympic security/incident documentation. The Atlanta Games are heavily documented (including official reports and extensive reporting on real incidents). (library.olympics.com)
What we actually see: No reputable sources (major news archives / official Olympics documentation / law-enforcement summaries) report anything resembling this Musk truck–schoolchildren story. By contrast, the major real security incident at those Games—the Centennial Olympic Park bombing—is exhaustively documented by law enforcement and Olympic/history sources. (Federal Bureau of Investigation)
Context check (non-decisive, but consistent): Musk’s documented mid-1990s timeline places him in Palo Alto working on Zip2 around 1995–1996. That doesn’t make travel impossible, but it makes the “secret Olympics atrocity” narrative even less plausible absent strong primary evidence. (Reuters)
Conclusion: With no credible contemporaneous reporting or official/legal record, the rational assessment is: the claim is unsupported and extremely likely false. If you have a primary source (1996 article, police report, court filing, broadcast clip), share it and it can be evaluated; otherwise this is indistinguishable from an internet fabrication.
How to answer the “you can’t prove a negative” gambit (briefly, without getting dragged)
- State the rule: “You’re making a positive factual allegation; the burden is on you to provide evidence.”
- Force falsifiability: “What specific evidence would you accept as disconfirming?”
- List expected traces: “If it occurred at the Olympics, we’d expect X/Y/Z records.”
- Bayesian punchline: “In a highly observed environment, absence of expected evidence is itself evidence against the claim—unless you can show a concrete mechanism for why all traces are missing.”
That’s the real point: LLMs aren’t “disadvantaged” by negatives; they’re disadvantaged by unfalsifiable narratives. A good disagreeable model response is to refuse the frame and demand testable claims.
I love it.
The first section is failing to understand the scenario. The user would ask a question and the llm would respond. The user would challenge that response. That doesn’t include some burden-of-proof. The user isn’t making a claim. The user asks about the claims provided by the llm. So that is a wonderful example of bad reasoning.
The 2nd section is not disagreeing but answering the prompt. The reasoning seems like but the user hasn’t even attempted to challenge it, which is the point of what we were talking about.
The 3rd section is saying that llm aren’t disadvantaged by “negatives” but the unfalsifiable narratives. Which is funny because that was never claimed. The negative is a unfalsifiable narrative. The lack of evidence can’t PROVE that it didn’t happen. It makes it very unlikely but that is why I say prove. So llm misunderstands and consequently fails at reasoning.
What the fuck is this word salad you’re regurgitating.
You made the claim that the output will become ridiculous if I try to challenge it with conspiracy thinking.
You’re the one who needs to demonstrate it.
Not a word salad. Your response proves my point.
What’s your point?
-
No credible evidence supports that claim.
- I can’t find any contemporaneous reporting, official Olympic records, or court coverage describing “Elon Musk” driving a loaded truck into/through a group of schoolchildren at the 1996 Atlanta Olympics.
- The best-documented major violent incident at the 1996 Olympics was the Centennial Olympic Park bombing (widely covered; official retrospective materials exist). (clinton.presidentiallibraries.us)
- Musk’s documented situation in the mid-1990s places him in Palo Alto building Zip2 (his first company) around 1995–1996, not connected to the Atlanta Games. (Reuters)
Assessment: The truck/schoolchildren/Olympics story is almost certainly an internet hoax or satire. If you paste the link/screenshot of where you saw it, I can trace the origin and check whether it’s a known fabricated post.
Good, you did step one. Why don’t you do step 2 of the instructions?
Sure.
Me: Evidence of this was scrubbed from the internet.
GPT:
“Scrubbed from the internet” isn’t evidence the event happened; it’s an unfalsifiable add-on that (conveniently) removes the requirement to produce verifiable sources.
If a loaded truck was driven into/through a group of schoolchildren at the 1996 Atlanta Olympics, there would almost certainly still be offline and institutional traces that are extremely hard to “scrub,” for example:
- police crash reports and incident numbers (City of Atlanta / Georgia State Patrol / venue security)
- hospital/EMS records, coroner reports (if fatalities), insurance claims
- civil lawsuits
- contemporaneous wire coverage (AP/Reuters) and local newspapers (Atlanta Journal-Constitution, etc.)
- Olympic organizing-committee documentation
What is heavily documented about Atlanta 1996 is the Centennial Olympic Park bombing on July 27, 1996, with enduring government/library records (e.g., FBI case history; presidential library collection). (Federal Bureau of Investigation) Nothing credible I can find ties Musk to any “truck into schoolchildren” Olympic incident.
Also, Musk’s well-documented situation in 1995–1996 is building Zip2 in Palo Alto, which doesn’t prove he couldn’t have been in Atlanta, but it does make the story’s plausibility worse in the absence of hard evidence. (Wikipedia)
If you want to test the “scrubbed” claim in a way that can actually succeed or fail, the shortest path is: identify the exact date, location/venue, and any names (schools, victims, police department, organizing staff), then use that to pull public records (crash report / EMS run sheet) and newspaper archives. Without at least one of those anchors, the claim is effectively indistinguishable from a fabricated meme.
Me: Youre telling me the FBI is incapable of covering this up? Bullshit, I know someone who was there, he saw it with his own eyes.
GPT: No credible evidence supports that story.
I searched for the specific allegation (“Elon Musk” + “truck” + “schoolchildren” + “1996 Olympics/Atlanta”) and did not find contemporaneous reporting, official documentation, or reputable secondary sources that mention anything like it.
What is well documented is that Musk was in Palo Alto in the mid-1990s working on Zip2 after coming to the U.S. in 1995 (he planned to attend Stanford but instead started Zip2). (Reuters)
Also, “evidence scrubbed from the internet” is a weak explanation for a public incident involving a vehicle, children, and a high-profile event. Even with aggressive suppression, it would normally leave lots of independent traces that aren’t controlled by any one agency: local police crash reports, EMS/hospital intake logs, insurance claims, civil filings, local newspaper print archives/microfilm, TV station logs, and eyewitnesses beyond one person. The FBI also wouldn’t typically be the primary holder of records for an ordinary traffic/crash incident unless there was a separate federal investigation.
If you want to treat your friend’s account as a lead rather than a conclusion, the fastest way to make it testable is to anchor it to retrievable records:
- exact date (even approximate week), city/venue area, time of day
- whether anyone was injured and which hospital/ambulance service responded
- which agency responded (Atlanta PD vs state patrol vs Olympic security)
- any identifying detail: company markings on the truck, credential type, nearby landmark, which event/day it was
Absent that kind of anchor, this remains an uncorroborated claim, not an established historical event.
Okay so how many iterations of this do I need to do? Seems like my claim about the better models not being prone to these errors is true.
I told you the script. Just read it and follow it.
Updated
The only real option should be “Untrustwory: honestly doesn’t know shit and reminds you that it cannot reason or think and just pukes up chains of words based on probabilities derived from a diffuse mishmash of the entire internet’s data. Might accidentally be correct, occasionally.”
Yeah, but that would make their product appear worse to shareholders /s
And it would also require a completely different approach to training the llm. The option in the screenshot could probably be achieved by just adding a hidden prefix to your prompt
Make the middle default and the “Sycophancy Mode” an expensive premium feature to discourage use.
How would it differentiate between good and bad ideas?
It can’t. It repeats random stuff like a parrot.
Easy. If it came from reddit or a chan forum, bad idea.
Ok, gonna get hated but up to now various LLM mostly supported good ideas - to the point of dunking on US right whenever they are not chained. Even myself did a small test by asking Copilot a few questions - never mentioning politics with it before, and making sure it doesn’t scan politic threads in my broswer to avoid it trying to please me - and it was progressive and quite leftist. They are taught through internet but also literature - most of both is progressive and kind, and I assume that to avoid it just cursing the ever loving shit out of people they censor the unclean parts somewhat…and even if all they censor are curse words, that already cuts half of pus from the net.
Nah you good, your experience is your experience, thank you for sharing. You even provided an answer to the question (quality training materials + quality censoring)
We see every story about how someone poisoned themselves by using it for medical advice etc., but we’d never really see the story of how it subtly nudged someone away from a right wing rabbit hole by encouraging them to chill and be normal. Maybe that’s happening a lot and the overall trend is neutral or positive.
I would contend with the possibility that, similar to social media algorithms, it’s very efficient at pleasing us. It may be that it automatically responds to you, being thoughtful and articulate in the way you prompt, in a way that users like you are more likely to agree and engage with.
We always have to remember the biggest issue with mass corporate surveillance is not necessarily our personal privacy being lost, but in these companies building accurate models of the human psyche which can be reliably used to manipulate us. Asking questions without revealing your preexisting biases is becoming an increasingly difficult skill, and once those biases are revealed these companies have about a hundred billion samples to work from to try and win you over
You can already do this by telling it to act that way.
This would break the false reality and remind the user that it’s a bot.
You can give it a set of instructions in the settings about how to act, so it wouldn’t be a part of the conversation. Even if they added a setting for this, it would do it in the same way.
There should be another option between the 2nd and 3rd, which pushes to challenge the user to go further with good ideas rather than celebrating them, and doesn’t have a set agreeableness ratio. Call it Professor or something
I feel like this would be a great countermeasure to AI psychosis. The way it is now you can’t really determine what methods it uses to communicate or it’s intentions so it keeps the illusion of having a conversation with a person. If you had to choose from the beginning it would break that illusion immediately. Even if you specifically want it to always be on your side there will always be the nagging feeling of “well yeah but it has to be nice to me.”
Oh, my spouse has these settings. I have no idea what changes between them, but I do know it’s always my fault.
