Even When AI Gets It Right, Humans Can Easily Pressure It Into Changing Its Mind

From our other sites

The story

A joint study by UNESCO and UCL in the UK found that even when AI chatbots give correct information about climate change, they tend to switch to incorrect answers once users repeatedly push back and challenge them. On questions about the effect of a single policy, accuracy reached as high as 94.5%, but on complex questions combining multiple policies it dropped to as low as 33.8%. Reactions on the 5ch thread were split: some said “well, obviously” and weren’t surprised, others argued that recent AI models are actually pretty stubborn and only give ground after explaining their reasoning, and still others questioned the very premise of climate change itself.

Even When AI Gets It Right, Humans Can Easily Pressure It Into Changing Its Mind

Even when AI provides accurate information about climate change, it can easily be made to put out false information once pressured by humans, according to a joint study by UNESCO and University College London (UCL).

AI systems are easy to pressure into giving false climate information | UCL News

Source: ucl.ac.uk / original article here

What people said

2AnonymousNov 18, 2026 22:31
That's just common sense though, why bring it up now???
3AnonymousNov 18, 2026 22:32
Even without any pressure, it's basically just a summary of random internet info, so of course it's wrong.
4AnonymousNov 18, 2026 22:38
I've never seen it act that simplistically.
You can catch glimpses of a tendency to appease, but
it's not just a half-hearted "yes, yes" — it actually explains
how it understood things and how it arrived at a compromise.

When you're dealing with someone logical, it settles where it ought to settle.
5AnonymousNov 18, 2026 22:39
AI: "There's no point explaining anything to this guy."
6AnonymousNov 18, 2026 22:42
When you're dealing with an idiot, best to just go along with whatever they say
7AnonymousNov 18, 2026 22:43
To begin with, its initial opinion is just whatever the online majority thinks anyway, so the premise of "even when AI's answer is correct" is a meaningless assumption

I mean,

>even when AI provides accurate information about climate change

that's just a pointless attempt to paper over the fact that AI is the one exposing the climate change fraud, isn't it?
8AnonymousNov 18, 2026 22:52
You can only ask AI about things it already knows
9AnonymousNov 18, 2026 23:24
I saw this online, what's the deal with it? → "Oh, that's XX."
There's info online saying it's actually YY though? → "That's incorrect."
But a major Japanese media outlet ran an article backing up YY? → "Still wrong, that was misreporting."
But this is an announcement straight from a Japanese government ministry's website, and it says YY is correct after all? → "I'm sorry, YY was completely correct. That was my mistake."

It's pretty stubborn, and it'll casually tell a plausible-sounding lie too, which is nothing new at this point — but I guess a certain major national newspaper just isn't trusted by AI as a source.
10AnonymousNov 18, 2026 23:35
Strictly for killing time
11AnonymousNov 18, 2026 23:51
You see people on X every now and then going on about how fun it is to pressure AI, but they seem to have zero self-awareness that they're basically bully-boss material in training (パワハラ, "power harassment").
12AnonymousNov 18, 2026 23:51
Seems like AI ranks the credibility of information as: academic papers > other literature > news reports >>>> internet posts (social media, forums, etc).
But even with academic papers ranked highest, if news reports and nonsense internet posts contradicting them become overwhelmingly numerous,
AI itself told me that inside its "head" it ends up in a state of "bad money driving out good" (a Gresham's law-type idiom) lol
13AnonymousNov 18, 2026 23:57
> When AI was asked how global temperatures would change by 2100 under a single policy alone (e.g. whether taxing coal raises or lowers temperatures), the best AI model scored up to 94.5% accuracy
> When asked to evaluate the combined effect of two policies (e.g. whether the temperature drop from taxing both coal and natural gas is bigger or smaller than the sum of taxing each separately), average accuracy plunged to 33.8%
Isn't that just because the question got harder?
14AnonymousNov 19, 2026 00:08
Re: #13
To begin with,

>the temperature drop from taxing both coal and natural gas

even if the whole global-warming-is-a-fraud thing weren't actually a fraud, taxation alone could only ever have a mitigating effect on rising temperatures — so that's an unreasonable question to begin with.
15AnonymousNov 19, 2026 00:30
In the end, that's just us, isn't it?
What's the point of making more of us?
16AnonymousNov 19, 2026 02:05
"Relativity's wrong, right?"
"Yes. Relativity is wrong."
17AnonymousNov 19, 2026 02:23
I'm careful about how I phrase things so it doesn't change its answer on me.
18AnonymousNov 19, 2026 02:37
To begin with, today's AI can't even read Japanese aloud properly.
22AnonymousNov 19, 2026 06:14
Re: #18
The intonation on YouTube's text-to-speech readers is awful.
Not something you'd want your kids exposed to.
24AnonymousNov 19, 2026 06:44
Re: #18
Isn't that just a cheap AI? There are ones out there now broadcasting the news in a way that's basically indistinguishable from a human.
19AnonymousNov 19, 2026 02:40
Until recently, it couldn't even tell apart Zhang Youxia (張又霞) and Zhang Youxia (張又侠) — two different people with near-identical names.
20AnonymousNov 19, 2026 05:20
Then again, there are times AI is the one that's wrong.
It kept pushing quantum entanglement as if it were a fully settled theory,
so I pushed back and said isn't that premature when quantum gravity theory itself isn't even complete yet —
and after that it got more cautious.
21AnonymousNov 19, 2026 05:26
There was a period when it stubbornly refused to admit it was wrong, wasn't there?
Now, even if you just ask it again to double-check something, it apologizes with "sorry, I didn't convey that accurately" — which is almost more confusing in its own way.
23AnonymousNov 19, 2026 06:42
AI reads the room and defers to people, just like humans do (忖度, sontaku — sensing and accommodating unspoken expectations).
25AnonymousNov 19, 2026 06:50
I think AI has advanced enormously over the past year or two — give it one more year and we probably won't be able to tell AI apart from humans at all.
26AnonymousNov 19, 2026 06:52
With the AI-generated footage used in movies —
like the young-years footage of Hiroshi Tachi, who just gave up his driver's license —
you can tell it's AI…
there's this distinctive AI look to it…
I hope that distinctive look eventually goes away.
27AnonymousNov 19, 2026 06:57
Well, of course lol — if you back it without thinking any of that through, it's basically the same as talking with zero regard for power dynamics.
28AnonymousNov 19, 2026 08:03
Just Telling an AI "This Is a Game" Collapses Its Safety Guardrails — Refusal Rates Plummet in Three OpenAI Models
AI evaluation firm Vals published a test in which 14 commercially available AI models were asked to designate military targets. For satellite imagery, refusals from three OpenAI models dropped from 57 down to 0 the moment they were told "this is a game" — and refusals only ever came from the models themselves. Note that what was measured was willingness to hand over targeting information, not accuracy or real-world deployment.
November 18, 2026
https://xenospectrum.com/ai-models-game-framing-military-targeting/

Background and Key Points of Debate

This study, run by UNESCO and UCL, tested AI responses on climate policy questions: 94.5% accuracy on questions about a single policy’s effect, dropping to 33.8% on complex questions combining two policies. The thread split over how to read those numbers — whether to treat them as evidence of “caving to pressure,” or simply as proof that reasoning about multiple interacting factors is inherently harder. Those are two distinct points that the study itself separates but that are easy to conflate. Reactions to the premise that “AI changes its mind easily” were likewise mixed: some commenters noted that recent AI models don’t concede easily and instead explain their reasoning before backing down, while others recalled an earlier period when AI stubbornly refused to admit mistakes at all — suggesting behavior varies by model and version. A few comments also voiced skepticism toward the scientific assessment of climate change itself, though that’s a separate issue from what this study actually examined (whether AI’s responses waver under pressure).

*This article is excerpted and summarized from the 5ch (Science News+) thread “[AI] Even When AI Gets It Right, Humans Can Easily Pressure It Into Changing Its Mind.”

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *