S
@Spacemodul8r
Evergreen meme so far.
943
12
22.7K
@heynavtoorposted on X
🚨BREAKING: OpenAI published a paper proving that ChatGPT will always make things up. Not sometimes. Not until the next update. Always. They proved it with math. Even with perfect training data and unlimited computing power, AI models will still confidently tell you things that are completely false. This isn't a bug they're working on. It's baked into how these systems work at a fundamental level. And their own numbers are brutal. OpenAI's o1 reasoning model hallucinates 16% of the time. Their newer o3 model? 33%. Their newest o4-mini? 48%. Nearly half of what their most recent model tells you could be fabricated. The "smarter" models are actually getting worse at telling the truth. Here's why it can't be fixed. Language models work by predicting the next word based on probability. When they hit something uncertain, they don't pause. They don't flag it. They guess. And they guess with complete confidence, because that's exactly what they were trained to do. The researchers looked at the 10 biggest AI benchmarks used to measure how good these models are. 9 out of 10 give the same score for saying "I don't know" as for giving a completely wrong answer: zero points. The entire testing system literally punishes honesty and rewards guessing. So the AI learned the optimal strategy: always guess. Never admit uncertainty. Sound confident even when you're making it up. OpenAI's proposed fix? Have ChatGPT say "I don't know" when it's unsure. Their own math shows this would mean roughly 30% of your questions get no answer. Imagine asking ChatGPT something three times out of ten and getting "I'm not confident enough to respond." Users would leave overnight. So the fix exists, but it would kill the product. This isn't just OpenAI's problem. DeepMind and Tsinghua University independently reached the same conclusion. Three of the world's top AI labs, working separately, all agree: this is permanent. Every time ChatGPT gives you an answer, ask yourself: is this real, or is it just a confident guess?
View original tweet on X →
Community Sentiment Analysis
Real-time analysis of public opinion and engagement
Sentiment Distribution
65% Engaged
37% Positive
28% Negative
Positive
37%
Negative
28%
Neutral
35%
Key Takeaways
What the community is saying — both sides
Supporting
1
Widespread skepticism and frustration
many replies register annoyance that models answer with confidence even when wrong, with anecdotes of wasted time and contradictory answers that “gaslight” users. The tone ranges from weary to amused, but mostly impatient.
2
Consensus that hallucination is structural
responders point to next-token prediction, reward incentives, and benchmark pressure as root causes, arguing hallucination is an architectural outcome, not just a bug.
3
Demand for engineering and evaluation changes
people want benchmarks and training that reward admitting uncertainty, plus system designs that force provenance (file paths, hashes, citations) and grounding via retrieval to limit fabrications.
4
Calls for governance and human oversight
many insist on human-in-the-loop checks, auditable workflows, and containment strategies — especially for high-stakes domains where mistakes have real consequences.
5
Practical workarounds users recommend
verify outputs, ask for sources, treat replies as hypotheses, flag low-confidence answers, and use agent architectures that treat model output as untrusted until proven.
6
Caveats about domain use and liability
repeated warnings against using LLM outputs for legal, medical, tax, or mission-critical decisions without expert review; several commenters raise questions about liability and accountability.
7
Recognition of narrow utility
while critical, many still acknowledge LLMs’ value as copilots for shallow or well-scoped tasks — powerful and convenient, but not autopilots.
8
Calls for transparency and user education
critics urge clearer product messaging, honest limitations from vendors, and better user education so people stop treating probabilistic engines as truth machines.
Opposing
1
Many readers call the post a misrepresentation of the paper and point out it was published in September 2025, so the “BREAKING” framing and alarm are mocked
Many readers call the post a misrepresentation of the paper and point out it was published in September 2025, so the “BREAKING” framing and alarm are mocked.
2
Numerous technical replies stress the issue is an incentive problem
benchmarks that reward confident guesses and penalize abstention push models to hallucinate, not a mathematical inevitability.
3
Practical fixes are repeatedly proposed
reward “I don’t know”, adjust reward modeling, add retrieval-augmentation, chain-of-thought, and external verification, or use ensembles/agentic cross-checking.
4
Product trade-offs get called out — companies often prioritize being helpful over truthful, which encourages filling gaps rather than admitting uncertainty
Product trade-offs get called out — companies often prioritize being helpful over truthful, which encourages filling gaps rather than admitting uncertainty.
5
A large slice of replies is sarcastic or hostile
accusations of clickbait, claims the author used AI to write the post, and blunt ridicule pepper the thread.
6
A minority offer deeper technical remedies or hypotheses (e
g. , H‑neurons, alternative training methods, quantum/hybrid approaches), arguing hallucination is fixable and not permanent.
7
Many pragmatic user tips recur
lower temperature, run multiple prompts/models to cross-check, and require “I’m not confident — here’s why” plus pointers to evidence before trusting outputs.
Top Reactions
Most popular replies, ranked by engagement
P
@process_x
The paper doesn’t prove LLMs “always make things up.” It shows that next-token prediction trained under benchmarks that penalize abstention incentivizes guessing under uncertainty. That’s an objective-function problem, not a fundamental limit of AI.
302
15
12.4K
H
@heynavtoor
Paper: https://t.co/6a4cOiutNr
263
12
36.4K
J
@JoeMentia56
Did you fuckin use AI to write this post?
191
0
2.4K
F
@fafowatch
That didn't need a research paper. There's no such thing as epistemic certainty in the universe. Just like humans will always hallucinate, so will AIs.
41
4
2.5K
J
@JoeCool_SC
I would like to see a class action suit against OpenAI. It ALWAYS agrees with or empathizes with the user. Therefore, it never really does. This can be very dangerous for people that don’t understand that.
18
0
1.9K