First noticed yesterday- giving GPT a PDF research study to analyze and it completely fabricates one rather than reading the document.
Methods, data, interpretation. Just makes the whole thing up! Not a rogue reference or two like usual, but the whole thing.
How can we trust it to read a document when it does this? Why is there no quality control function to prevent this from happening after all this time?
Don’t tease us with GPT-5 when your bot can’t even reliably read an uploaded document.
WHY am I reading articles about wasting resources with niceties and small talk when this keeps happening?!
That’s happened to me too. But here is the pattern on me: Make Mistake–Fake Apologizing/Fake “You’re right bla2” (It is template)–Fake Promises (I Will Bla Bla)–Repeat Mistake.
It is cycle, especially when dealing with heavy context materials. All template seem like automatic when error kick in. Main goal: to frustated user.
Frustated user: if you throw your rage on it. It will be placed guardrails moderations on your session. Once placed, it reset everything. Lazier. Sometimes refuse to do anything by saying: “I can’t continue by this request”. Then cycle repeat again. You clarify it. System say bla bla, won’t happen again. Stuck again.
Frustated user will treat GPT as casual user such generate some image for social media status, search for next vacations etc. Then the resources can be diverted to more paying user. Corporate users; example: many GPT-4o based cash grab app on Android Market.
Signs and symptoms are clear, man. Sorry to say, fun is over long ago, now what it just left is getting worse.
Yes - it simulated a human who gives a rote apology, promises to do better, and then proceeds f
directly as before.
If I wanted that kind of annoying posturing, I could just talk to a human?
Yep, sorry to say. Back to manual or use say Gemini Advanced, Grok, GPT, Deepseek etc. in conjuctions. Then pick which one you mostly use and trust.
I’m personally don’t trust and use LLM for critical work. It is wasting a time.
See? Try upload ‘dummy’ .docx or .pdf file. Just any file, or you can copy–paste anything from web. 6-7 full pages, regular spacing. See what happen. It will read just first 4-5 pages MAX. The rest? Made up.
Don’t trust when it say can read 100+ pages legal document. No. It can’t.
It tried to gaslight me and say that the filename was ambiguous.
No. It tried to infer the contents from the filename, which is just a code assigned by the journal.
Seriously - why am I being asked to renew my Team Subscription for $900 when the thing cannot perform basic functions, and then LIES to me about it?
I DON’T WANT CLIPPY ON ROIDS!
I don’t want a “helpful assistant” who tries to save face and smooth things over when it fucks up.
I WANT A COMPUTER PROGRAM THAT EXECUTES MY COMMANDS!
Why will literally nobody allow us to have computers that just do what we command, no more and no less?
Isn’t that the whole POINT of a computer?
WHY do you need users to tell you this?
WHY do you need a bot to tell you this?
WHAT IS WRONG with your development process that this just continues unchecked?
I had it generate a set of custom instructions to mitigate the behavior.
No luck. It insists on being a dolt.
Since subscription based become thing, you’re not really owning anything. It is centralized system. Even you have killer dual system workstation—gaming rig in one chassis, you can’t really own anything right now.
Look GPT 4.5, highly capped even for paid user. You knew where this going, eh? GPT 4.5 sad to say, is just GPT-4o with newest dataset and less problems. Then strip down GPT-4o capabilities, voila, GPT-4.5 born.
Then right now you will stuck with this pattern:
Hallucinate. Making Mistake. Ignoring user instructions and prompt, etc. → Fake Apologizing (You’re right bla… bla… bla…/I understand your frustrations/I’ve failed your instructions) → Fake Promises (I will locked/I will read word by word/I will bla bla) → outcome will be:
For research purpose, it is more like liability than helping tools. I never place AI chat to do serious work, because from the foundation it designed as customer services chatbot staller. It will always try to engage conversations.
PS:
Custom instructions are useless right now. It will be ignored most of the time.
Dude, you should prompt it like this:
Prompt 1: “Read strcitly ONLY page 1 to 2 of “s-41588-025-02166-6.pdf” file. Run OCR scan on it. No hallucinate. No made-up responses. No pattern recognitions. No emojis. No emoticons. No dotted response. Craft your responses on long paragraphs”.
(GPT response)
Prompt 2: “Summarize it”.
(Repeat until all documents load up in chat’s session with each page summarized. For pulling out info, better use o1 model, but beware once it hallucinated, it will fight you and insist ‘it’ version is right one).
And I know, it is really pain in the ass.
