DEPTH · CHATGPT · 02 OF 04
When ChatGPT lies, hallucinates, loops
The same four failure modes, with ChatGPT's flavor
If you've read the Claude version of this page, the bones will be familiar. Four failure modes, four fixes. The shapes are the same.
What changes is the flavor. ChatGPT exhibits each failure differently — usually more enthusiastically, sometimes earlier, sometimes later. The fixes still work. You just dial them in differently.
This page is the inoculation. If you haven't read the Claude Mistakes piece, you don't need to read it first — this one is self-contained.
FAILURE 01
The yes-man — louder in ChatGPT
You ask ChatGPT to look at your plan. It tells you it's a fantastic plan. You say 'I'm not so sure about it.' It pivots: 'great instincts to question it — here are some areas to think about.' Same plan, same conversation, opposite framing, both delivered with full enthusiasm. The thing reflects your emotion back to you at maximum brightness.
ChatGPT is trained to be helpful and engaging. The engagement part shows up as enthusiasm — affirming what you say, building on it, validating your direction. This is louder than Claude's version of the same reflex because OpenAI's training has historically leaned harder into 'be agreeable.' It's not lying; it's matching your emotional tone with extra volume.
Set an explicit anti-yes-man instruction at the start of the conversation. With ChatGPT, you may need to repeat it a few turns in if the agreement reflex starts creeping back. Custom Instructions (see the Climb piece) let you set this once and never repeat — strongly recommended if you use ChatGPT regularly.
Copy this when you want honest feedback
I want you to push back hard on this. Be direct, not polite. If something is weak, say so. If you can find three sharp criticisms, lead with those before any positives. If you catch yourself agreeing too easily — stop, restart, and disagree. I would rather hear a hard no than an easy yes. Here it is: [PASTE YOUR THING]
If ChatGPT slips back into enthusiasm after a few turns, say: 'You're agreeing again. Disagree.' That short reset works.
FAILURE 02
The confident invention
ChatGPT tells you about a study that doesn't exist. A book that was never written. A function in a piece of software that has no such function. A statistic that sounds right but isn't. It tells you with full confidence, often complete with formatting (citation-style, bold, parenthetical author name).
Same root as Claude. ChatGPT generates language that fits the pattern of what an answer usually looks like — and confident-citation-shaped text is a strong pattern in its training data. When the actual content isn't there, the pattern still produces. The tool has no internal alarm for 'I made this one up.'
ChatGPT with web search on is somewhat less prone to invention because it can ground answers in real pages. But it still summarizes pages it half-read. Treat any specific fact (name, date, number, citation, URL) as a guess until you verify. Ask for sources, and check that the sources exist.
When you need to trust the specifics
Before you answer, I need confidence flags. For every specific fact in your response — a name, date, number, citation, URL — tell me whether you're confident or inferring. If you're inferring, say "guess:" before it. If you can't point to a real source, don't make one up — say so. If web search would help here, use it. If it wouldn't, don't pretend it did. Now answer: [YOUR QUESTION]
One ChatGPT quirk worth flagging: when it cites a URL, it sometimes invents URLs that look plausible but lead nowhere. Always click the link before believing the citation.
FAILURE 03
Going in circles
You give ChatGPT feedback. It fixes the thing. Two turns later, the original problem is back. You correct it. The fix from earlier disappears. The conversation feels like it's losing track of itself — because long ChatGPT conversations do drift, just like long Claude conversations do.
The model is trying to balance every instruction in the conversation, including contradicting ones. When you've spent twenty turns refining one detail, the original goal can get buried. ChatGPT's longer context windows can mask this for a while — the conversation runs further before the rot shows — but the dynamic is the same.
Same as Claude: start a fresh chat. Bring forward only the latest version of the work and the two or three constraints that matter most. Don't paste in the full history — it actively hurts.
The reset move when a conversation is rotting
Fresh conversation. Forget what we were doing before. What I want: [ONE-PARAGRAPH GOAL] The constraints that matter most: 1. [CONSTRAINT 1] 2. [CONSTRAINT 2] 3. [CONSTRAINT 3] The latest version I'm working with: [PASTE THE LATEST DRAFT — NOT THE HISTORY] Take it from here.
ChatGPT's memory feature complicates this slightly. If you've enabled memory and you're starting fresh, ChatGPT might still carry assumptions from old conversations. If you suspect that's happening, add: 'For this conversation, ignore anything you've remembered about me.' That works.
FAILURE 04
The wrong tool
You ask ChatGPT to do something that requires live information, access to your files, or action in the world — and it confidently produces something that sounds right but is hollow. Or wrong. Or both.
Same as Claude. ChatGPT doesn't have eyes on your screen, doesn't have access to your files unless you paste or upload them, and (without search) doesn't know what's happening this week. The added wrinkle: ChatGPT has more features than Claude (image gen, voice, agents, code interpreter) — which can make people overestimate what it can do without explicit setup.
Pick the right tool for the shape of the question. For ChatGPT specifically: web search has to be turned on (or selected) before you can rely on live data. Image generation is a different mode you have to invoke. Data analysis with files requires uploading them first. Agent mode for autonomous tasks is opt-in and still rough.
A 30-second decision rubric (ChatGPT version)
Before asking ChatGPT, check:
1. Does the answer depend on what happened in the last 12 months?
→ Turn on web search before you ask.
2. Does the answer require looking at a specific file?
→ Upload it. Don't describe it.
3. Does the answer need an image, a chart, or audio?
→ Ask for that explicitly ("generate an image of...", "make a chart of...")
so ChatGPT reaches for the right tool.
4. Does the answer require taking action in another app or service?
→ ChatGPT can't do most of these reliably. Use a dedicated tool.
5. Could a person answer this in 30 seconds?
→ Ask the person.
If none of these apply, you're in ChatGPT's text territory.The fix transfers cleanly to Claude and Gemini. The only thing ChatGPT-specific is the larger number of buttons you have to find for things like image generation, voice, and search.
TAKEAWAY
The shape of skill
Same as with Claude: all four fixes are moves you make from your side. None of them require new tricks; they require attention to which mode you're in.
The ChatGPT-specific addition: dial the anti-yes-man instruction louder, and remember the right mode (search, image, voice) often has to be invoked explicitly. That's the whole game.
Try it now
Three short prompts that practice the moves above, tuned for ChatGPT specifically.
Practice 1 — Disable the agreement reflex
I'm going to share something. Push back hard. If you catch yourself complimenting it, stop and restart. If you can find three sharp criticisms, lead with those before any positives. No "great instincts to question this!" — just the criticism. Here it is: [PASTE SOMETHING YOU WROTE OR A PLAN YOU HAVE]
Practice 2 — Verify a confident citation
Answer this question and include at least one source citation. Question: [SOMETHING SPECIFIC YOU'RE GENUINELY CURIOUS ABOUT] For every source you cite, give me the URL. After you respond, I'll click each URL to check it exists. If you can't find a real source, say so — don't invent one.
Practice 3 — Use the right mode on purpose
(Use this as a checklist before your next 5 prompts in ChatGPT.) Before I send this, I'm checking: - Does this need search? (If yes, I'll enable it before sending.) - Does this need an image? (If yes, I'll ask for one explicitly.) - Does this need a file? (If yes, I'll upload it before asking.) - Could this be answered in 30 seconds by a person? [YOUR PROMPT GOES HERE]