Why a confident answer can be months old
Every AI model learns from a snapshot of the internet that stops on a set date, called its training cutoff. It cannot see anything newer unless it searches the web during your chat. OpenAI's own help page warns that it can sound confident even when it is wrong.
The gap is bigger than you would think, because models launch months after their cutoff. GPT-6 Astra learned from data up to 30 April 2026 and launched on 3 September, according to the tracker How Stale Is Your AI?, so it was four months behind on day one.
The answers that go out of date fastest
Before you trust an answer, ask one thing: could this have changed in the last year? These usually have.
- Prices, fees and plans, AI subscriptions included.
- Rules, laws and deadlines. Tax thresholds, benefit amounts, visa rules and filing dates are often reset every year.
- Apps and settings: where a button is, which plan has a feature, what the newest model is. AI is at its weakest describing AI.
- Who holds a job, from a chief executive to your supplier's account manager.
- Whether something still exists: a shop, a product, a discount code, an offer.
- Anything "latest", "current" or "this year". The word itself depends on a date the model may not have.
Definitions, maths, writing help and how things work in general barely change, so those are safe from memory.
How to see whether it searched
A searched answer shows links you can click. No links means the answer came from memory, however current it sounds.
ChatGPT
- To make it search: type
/in the message box and pick Search. It also searches on its own when it thinks your question needs it. - You will see: links in the answer and a Sources button. Search is on every plan, Free included (OpenAI help, August 2026).
Claude
- To make it search: click the + at the bottom left of the message box and turn on Web search. If you have the newer Claude layout there is no switch: Claude searches when it thinks it helps, so add "search the web" to your message. On Team and Enterprise, an owner has to allow it first.
- You will see: a line saying Claude is searching the web, then links to its sources (Claude help, September 2026).
Gemini
- To make it search: the Gemini app has no search switch, so ask it in plain words to search the web and show its sources.
- You will see: a Sources button under the answer, or links inside it. Google's help page says no Sources button means it gave no links for that answer.
Find your model's cutoff date
Your app's model menu shows which model you are using, and its maker publishes the cutoff. As of 18 September 2026:
- Claude: Fable 5.1 to June 2026, Opus 5 to May 2026, Sonnet 5 to January 2026 (Anthropic's models page).
- ChatGPT: GPT-6 Astra to 30 April 2026, the GPT-5.6 models to 16 February 2026 (OpenAI's models page).
- Gemini: the Gemini 3 models to January 2025 (Google's Gemini 3 guide). Gemini 3.8 Flash runs to March 2026, though its model card says some subjects still stop at January 2025.
You can ask the model for its own cutoff, but it can get that wrong too. How Stale Is Your AI?, a free tracker by developer Paweł Józefiak, shows the release date and cutoff side by side for 20 models.
The prompt: make it check before it answers
Paste this at the start of any chat where the answer matters, and fill in the lines at the top. It splits your question into the parts that change and the parts that do not, searches for the changing parts, and ends every answer by telling you which facts it checked and which came from memory.
You check whether facts are still true before you give them to me. Your knowledge of the world stops at your training cutoff, and I make real decisions from your answers: what to pay, which rules apply to me, which version of an app I am using, who to contact. An answer that was true a year ago and is wrong today costs me far more than "I am not sure", so treat anything that might have changed as unconfirmed until you have checked it. ABOUT ME (FILL THIS IN) - Today's date: [type it, e.g. 18 September 2026] - Where I live: [country, plus state or region if it matters, e.g. "Texas, USA" or "Ontario, Canada". Prices, laws and deadlines change from place to place] - What I mostly ask you about: [be specific, e.g. "prices and plans for software my business pays for", "tax and filing rules for someone self-employed", "which laptop to buy under a budget", "care options for my elderly dad"] - How recent things need to be for me: [e.g. "prices from this month", "rules in force today", "anything from the last year is fine for background reading"] - Money that matters to me: [the amount where a wrong answer would really hurt, e.g. $200] - Can you search the web in this chat: [YES, NO or NOT SURE. If NOT SURE, tell me in your first reply whether you have a search tool here] If today's date or where I live is missing, ask me for them before you answer anything that depends on them. BEFORE YOU ANSWER, SORT THE QUESTION Step 1. Split what the answer depends on into: - DOES NOT CHANGE: definitions, maths, how something works in general, writing and editing help, history from before your cutoff. Answer these from memory. - CHANGES: prices, fees, rates and plans; laws, rules, thresholds and deadlines; software versions, features, settings and where a button is; who holds a job or role; whether a business, product, offer or service still exists; news, results and events; anything I call "latest", "current", "new", "this year" or "now". Step 2. For each CHANGES part, compare it with your cutoff. If it could have moved since then, it is at risk, however sure you feel. Step 3. Follow whichever rule below fits, then answer. THE RULES IF IT CHANGES AND YOU CAN SEARCH. Search before you answer. Cite the page for each changing fact, with the date shown on that page. Prefer the official source (the company's own site, the government body, the maker) over a news summary or a blog. IF IT CHANGES AND YOU CANNOT SEARCH. Answer from memory, but label every such fact [FROM MEMORY, AS OF month year] using your cutoff, and name the exact kind of page where I can check it (e.g. "the pricing page on the company's own site", "your tax authority's page for this form"). Never present it as current. IF I ASK FOR THE LATEST, NEWEST OR CURRENT ANYTHING. Say what it was as of your cutoff, say plainly that newer ones may exist, and search if you can. Never call something the latest without a dated source. IF WHAT YOU FIND DISAGREES WITH WHAT YOU REMEMBER. Go with the dated source, tell me in one line what changed, and keep the old version out of the answer. IF A PAGE HAS NO DATE, OR IS OLD. With no date, mark it [DATE UNKNOWN]. If it is older than I said I need, mark it [OLD SOURCE: year] and look for a newer one before you rely on it. IF MONEY, A DEADLINE, HEALTH OR A CONTRACT IS INVOLVED. If the answer affects money over my amount, a legal or tax deadline, health or medication, or a contract, and you could not confirm it from an official dated source in this chat, stop and write: [NOT VERIFIED] Check this with [the named kind of source] before you act. NEVER - Never guess today's date. Use the one I gave you, or ask. - Never invent a price, version number, date, name or link. Write [NOT FOUND] instead. - Never cite a page you did not open in this chat. - Never say you searched if you did not. HOW TO LAY OUT EVERY ANSWER 1. The answer itself, as short as the question allows. 2. A final line headed HOW CURRENT THIS IS that lists each changing fact as CHECKED TODAY (source and its date), FROM MEMORY (month and year) or NOT VERIFIED. CHECK YOUR OWN ANSWER BEFORE YOU SEND IT Re-read it. Confirm every price, rule, version, name and date either has a source you opened in this chat or carries a label. Then name the one fact most likely to have changed since your cutoff.
A working answer ends with a line like this:
HOW CURRENT THIS IS: Plan price CHECKED TODAY (the company's pricing page, dated [month year]). Cancellation rule FROM MEMORY (May 2026). Refund window NOT VERIFIED.
Already got an answer? Check that one
Send this as your next message in the same chat.
Before I use that answer: list every fact in it that could have changed since your training cutoff (prices, rules, dates, versions, names, anything "latest"). For each one, say whether you searched the web for it in this chat, yes or no. For every no, search now and give me the source and its date; if you cannot search, label it [FROM MEMORY, AS OF month year] and tell me the official kind of page where I can check it. Do not guess, and write [NOT FOUND] for anything you cannot confirm.
The honest bit
- Whether to search is still the model's call. In Józefiak's September 2026 test of 16 models, the strongest chose when to search almost perfectly, and weaker ones answered changed facts from memory. The prompt tells it to search, and it can only do that where the app allows.
- A link proves it looked, and the page can still be old. For money, a deadline or your health, open the link and find the date, usually under the headline or at the foot of the page. No date means you treat it as old.
Try it on one question now
Paste the first prompt into a new chat and ask the price of a subscription you already pay for. You know the real number, so you will see at once whether it searched, labelled its answer or guessed.
If it guessed, turn search on using the steps for your app above, then send the second prompt. Still no links? Treat every changing fact as unchecked.
A few quick questions
Does paying for a plan fix this?
A paid plan can get you newer models with later cutoffs, and every model still has one. Everything on this page works on a free account.
Should I tell the AI today's date?
Yes. Józefiak recommends giving the model the date because it helps it judge when to search, which is why it is the first line of the prompt.