Phase 1 · ChatGPT · Level 1 · Foundations
ChatGPT's limitations (and how to work around them)
By the end, you'll be able to…
- Name ChatGPT's three predictable weaknesses (outdated knowledge, confident wrongness, and over-reliance) and the workaround for each
- Write prompts that reduce the risk: asking it to search and cite, and to flag its own uncertainty
- Apply a quick three-question check before trusting any ChatGPT answer
Why it matters
ChatGPT is useful, but it has real, predictable weaknesses, and people get burned by trusting it too much: a wrong fact in a report, a made-up source, or leaning on it for something it can't do. Knowing the limits, and the simple workaround for each, is what turns you from a nervous user into a confident one.
Why this lesson matters
Everything before this showed what ChatGPT does well. This is the honest other half: where it lets you down, and what to do about it. None of these are reasons to avoid it; they're things to know, the way you know a hammer isn't for screws. Each limitation below comes with a workaround you can apply immediately.
Limitation 1: it doesn't automatically know recent things
ChatGPT is trained on information up to a certain point in time, its knowledge cutoff. Ask about something after that, and it may not know, or may give outdated information in the same confident tone as everything else. It has no idea what's happening today unless it goes and looks.
The workaround: for anything time-sensitive (prices, news, "the latest", who currently holds a role) don't rely on its trained-in knowledge. Turn on web search so it fetches current information and gives you links to check, or confirm against an official website. Better still, ask for it explicitly.
Search the web for this and give me the answer with links to the sources you used, plus the date each was published. If the sources disagree or look out of date, say so. Question: [your time-sensitive question].
Why this works: Explicitly asking it to search and cite pushes it to fetch current information and hand you sources you can open, instead of answering from stale memory.
Limitation 2: it can be confidently wrong
This is the big one. ChatGPT sometimes states things that are simply untrue, in the same fluent, self-assured tone it uses for correct answers. This is hallucination: plausible-sounding output that isn't real. It can invent facts and figures, quotes nobody said, and (the one that catches people badly) sources, studies, and web links that look completely real but don't exist. It's just as capable of getting arithmetic wrong on a messy spreadsheet, or misreading a number, without any signal that it's unsure.
The workaround: treat any specific fact, number, name or source as unverified until you've checked it against an independent, trustworthy source. Never paste a citation ChatGPT gave you into your work without opening it. And asking "are you sure?" is not a reliable check; it can be confidently wrong a second time. You can, however, prompt it to be more careful and to show its working, which makes errors easier to spot.
Answer, then mark each claim as (confident) or (uncertain), and list anything I should verify independently before relying on it. If you're not sure, say so rather than guessing. For any calculation, show the figures you used step by step.
Why this works: Requesting a confidence flag and visible reasoning doesn't make it correct, but it surfaces the shaky parts so you know exactly what to double-check.
Limitation 3: over-reliance
The subtlest risk isn't in the tool, it's in the habit. Lean on ChatGPT for everything and two things happen: you stop noticing when it's wrong, and your own skills go rusty. It's a helper, not a replacement for your judgement.
The workaround: keep yourself in the loop. Read what it produces critically rather than rubber-stamping it. For anything that matters, use it to get started or to check your thinking, not to hand over the whole job unread. A good rule: if you couldn't spot a mistake in the output, be cautious about using ChatGPT for that task at all. You need to be able to sanity-check the result.
Limitation 4: it only knows what you've shown it
ChatGPT knows nothing about your specific situation, files or accounts unless you tell it or paste it in. It can't read your emails, check your systems, or know what happened in your meeting. If it answers as though it does, it's guessing.
The workaround: give it the relevant information directly, while keeping private details out, as Phase 0 covered. If an answer seems to assume facts about you, ask yourself whether you actually told it that. If not, be sceptical.
Use only the information I've given you. Don't assume or invent anything to fill gaps; if something you need is missing, ask me for it instead of guessing.
Why this works: Telling it to work only from what you provide and to ask rather than assume stops it filling gaps with confident guesses.
A quick way to hold all this in your head
Before you trust a ChatGPT answer, ask three things:
- Is it recent? If it depends on current information, check elsewhere.
- Is it a specific claim? If it's a fact, figure or source, verify it.
- Could I spot it if it were wrong? If not, don't rely on it unchecked.
Try it now
Common mistakes
- Trusting fluent writing as proof of accuracy. Confidence is not correctness; the tool sounds equally sure either way. This is the over-trust trap, and it's the one that does real damage: a made-up source or a wrong figure, waved through because the writing was polished, ends up in work with your name on it. Verify anything load-bearing before you rely on it.
- Copying sources or statistics straight in. Open and confirm them first; invented references are common and convincing.
- Assuming it knows today's information. It often doesn't unless it searches the web.
- Handing over tasks you can't check. If you couldn't catch an error, you can't safely rely on the answer.
Keeping current
Models get better at some of this over time (newer ones hallucinate less and search more readily) but none of it disappears, so the habits here stay essential. OpenAI documents current model behaviour and limitations in its ChatGPT release notes. Accurate as of 13 July 2026.