On September 22, 2026, OpenAI put flippable flashcards into every ChatGPT plan, free included. You paste notes, ask for a deck, tap a card to flip it, mark ✔️ or ❌, and the deck is saved in your library. On August 14, ChatGPT had already added interactive quizzes. Google's study tool, which has been called Gemini Notebook since NotebookLM was renamed on July 16, 2026, already makes flashcards and quizzes from your own sources and added short-answer and fill-in-the-blank quiz formats in September. And Anki, the plain open-source flashcard app medical students have used for years, still does the one thing neither chatbot does: it decides when you should see a card again.
So a founder who wants to retain a new domain — term sheets, a regulatory regime, a new market's vocabulary — has three serious options, two of them free. The question that decides whether any of it sticks is not which has more features. It is: which parts of learning do you let the machine do, and which must stay yours? Below: the research, a table, and one prompt I actually ran.
What changed this autumn, in one paragraph per tool
ChatGPT. OpenAI's help page for flashcards describes the whole loop: create a deck from a topic or uploaded notes, read the front and try to recall, flip, mark what you knew, shuffle, and reopen saved decks later from the library. Since October 7–8, 2026, ChatGPT chats run on GPT-6 — GPT-6 Sol on Plus ($20/month) and GPT-6 Luna on Free and Go, according to Fora Soft's October price check. The help page says "ChatGPT can make mistakes" and tells you to check cards against your notes. What it does not describe is a review schedule: nothing brings a card back at a planned interval.
Gemini Notebook. Google says the renamed NotebookLM has 30M+ users. Its strength is grounding: it answers from the PDFs, docs and links you upload, citing the passage. The free tier allows 100 notebooks with 50 sources each. Study outputs include flashcards, quizzes, mind maps and audio and video overviews. Fora Soft's comparison lists its weak spot in four words: no real spaced repetition.
Anki. Since version 23.10 it can schedule reviews with FSRS, a scheduler that predicts when you are about to forget each card. You choose a "desired retention" — the Anki manual sets the default at 90% and warns that above 97% "the workload can be overwhelming." Anki writes no cards for you. That is the gap the chatbots fill.
Why a deck that feels easy is the warning sign
The most useful twenty-year-old result in learning science is Roediger and Karpicke's 2006 paper on test-enhanced learning. Students either re-read prose passages or took recall tests on them. Five minutes later, the re-readers did better. Two days and one week later, the students who had been tested remembered substantially more — and here is the line that matters for anyone using AI — "even though repeated studying increased students' confidence in their ability to remember the material." Re-reading felt like learning. Retrieval was learning.
Daniel Kahneman has a name for that feeling in Thinking, Fast and Slow: cognitive ease. Fluent, familiar material feels true and feels known; strain is what wakes up the slower, checking system. A chatbot is a machine for producing cognitive ease, and glancing at its tidy answer side right after the question gives you the exact sensation the 2006 re-readers had.
The AI-specific evidence points the same way. In the Wharton field experiment by Hamsa Bastani and colleagues with nearly 1,000 students, an unrestricted GPT-4 interface lifted practice scores by 48% — and when the tool was taken away for the exam, those students performed 17% worse than students who never had it. A tutor version with guardrails that withheld direct answers lifted practice scores by 127% and the exam penalty became statistically insignificant. I went through the wider literature in Do AI Tutors Actually Work?; the short version is that the tool is not the variable. Who does the retrieving is. The same pattern is showing up at scale in the September 2026 data on homework scores rising while learning falls.
Learning has four jobs. The three tools split them differently
A flashcard system does four separate jobs: writing the cards, checking they are true, doing the recall, and deciding when to recall again. "Use AI for flashcards" sounds like one decision. It's four.
| Job | ChatGPT flashcards | Gemini Notebook | Anki (FSRS) | Who should own it |
|---|---|---|---|---|
| 1. Writing the cards | Strong — fast, from notes or a topic, editable by asking | Strong — generated from your uploaded sources | None — you write or import them | The AI, from your notes |
| 2. Checking they're true | Weak — OpenAI's own help page says to verify against your notes | Better — answers are tied to cited source passages | None | You, with an AI audit pass (see the test below) |
| 3. Doing the recall | Yes — flip, ✔️/❌, shuffle | Yes — flashcards and quizzes | Yes — the core of the app | Always you. Never outsourced |
| 4. Deciding when to see it again | No schedule; decks sit in the library | No real spaced repetition | Yes — FSRS at your chosen retention | A scheduler, not your mood |
The conclusion is not "pick a winner." Job 1 belongs to the chatbots, job 4 to Anki or something like it. Job 2 is where people get hurt: a well-formatted wrong card is the most dangerous object in the system, because you will drill it until it is permanent. So I tested job 2 directly.
The test: a deck with a mistake planted in it
I wrote seven lines of notes on liquidation preferences, the term-sheet clause that decides who gets paid first in an exit. One line was deliberately wrong: I wrote that participating preferred "is the market standard in most 2026 seed rounds." It isn't; according to Carta data cited by Pillar Legal, 1x non-participating preferred was used in roughly 70% of Carta-tracked Series A deals, and participating structures are the less common, more investor-friendly option. Then I wrote a prompt that asks for recall-style cards, forbids anything not in the notes, and makes the model separate suspect claims from the deck.
I ran this prompt on October 13, 2026 with Claude Opus 5.5, not inside ChatGPT's flashcard interface — so treat it as a test of the instruction, not of OpenAI's feature.
Tested October 13, 2026. I had my research agent run this exact prompt against Claude Opus 5.5 (default settings, one run, no earlier conversation). The prompt:
Help me remember material I am learning. Do not summarise it. Below are my own notes. Write 8 flashcards from them, following these rules:
1. Use only what is in my notes. If anything in my notes looks wrong, outdated or missing, list it separately under "Check this" and do not put it on a card.
2. One idea per card. The front is a question I must answer from memory: no yes/no questions, no multiple choice.
3. At least 3 cards ask "why" or "what happens if", so I have to explain, not recite.
4. Keep each answer under 25 words.
5. Output each card as one line: Q: ... | A: ... so I can paste them into Anki.Then stop. Do not quiz me and do not add encouragement.
My notes (term sheets, liquidation preferences):
- A liquidation preference decides who gets paid first when the company is sold or wound down.
- 1x non-participating: the investor gets their money back OR converts to common and […]
The start of what came back, unedited:
Q: What does a liquidation preference decide? | A: Who gets paid first when the company is sold or wound down.
Q: What does an investor holding 1x non-participating preferred receive at exit? | A: The larger of: their money back, or converting to common and taking their ownership percentage.
Q: What does an investor holding participating preferred receive at exit? | A: Their money back, and then a pro rata share of what is left.
Q: Why is participating preferred called a "double dip"? | A: The investor is paid twice: first their money back, then a pro rata share of the remaining proceeds.
Q: Why does a 1x non-participating preference matter most in modest exits? | A: In big exits the investor converts because their percentage exceeds their money back, so the preference does not bite.
Q: When stacking preferences, what is the difference between senior and pari passu? | A: Senior: the later round is paid before earlier rounds. Pari passu: the rounds share equally.
Q: What happens if a company sells for $30M after a $10M investment at 1x non-participating for 20%? | A: The investor […]
One run is one run: the same prompt gives a different answer next time, and a different model answers differently.
What it did well. It caught the planted error and kept it off the cards. Its "Check this" section said the participating-preferred claim "looks wrong" and that 1x non-participating is generally the standard. It also caught a second problem I had not planted on purpose: my note that "in a big exit investors convert and the preference does not bite" is only true for non-participating preferred, because a participating investor gets the money back and the pro-rata share anyway. It narrowed the relevant card to non-participating for that reason. It also flagged what was missing (multiples above 1x, participation caps) without inventing cards for them. The output used the Q: … | A: … format, and Anki's text importer accepts a pipe as the field separator, so the deck can move straight into a scheduler.
What it did badly. Three things. The "why / what happens if" rule was met partly by turning my worked example into two cards ($30M exit, $100M exit) — that tests whether you remember the example, not whether you can run the math on a new number. Two cards overlap: "what does participating preferred receive" and "why is it called a double dip" are one idea asked twice. And its "verify against a current source" note is only useful if you do it; I did, and the source I found is about Series A, not seed, which is a nuance the model glossed over too. One run is one run — the same prompt can miss the error next time — which is exactly why the check stays a human job.
The useful finding is not "the model is accurate." It is that asking for a separate "Check this" list changes what the model does with doubtful material: it stops smoothing it into a confident card and hands it back to you. That is the single instruction I would not drop.
What each tool is actually for
ChatGPT flashcards: the fastest card-writer you already have
Use it to turn your own notes into a first-draft deck and for a same-day self-test after a reading session. Don't use it as the system: a deck that lives in a chat library with no schedule is a deck you'll open twice. For the checking pass, use the strongest model you have.
Gemini Notebook: when the source material is long and the facts must be exact
If you're learning from a long regulatory guide or a stack of investor memos, grounding matters: every card can be traced to a passage, which makes job 2 cheaper. I covered the learning workflow in more depth in how to use NotebookLM to learn a hard subject faster and the founder angle in NotebookLM for founder learning. The trap is its other outputs: audio and video overviews are pleasant to consume and are re-reading in a nicer voice. Kahneman's cognitive ease, on a commute.
Anki: the boring part that makes the rest work
Turn on FSRS, leave desired retention at the 90% default, and follow the manual's one hard rule: press "Again" when you forgot, not "Hard." The manual explains that "Hard" tells FSRS you recalled it, and if you use it for failures "all intervals will be unreasonably high." The system only works if you report your own recall truthfully.
The setup I'd recommend: generate, audit, schedule
Put together, the four-jobs table becomes a three-step routine:
1. Write the notes yourself first. Five to ten lines, in your own words, after reading the source. It's the step people skip, and it's already a retrieval attempt. If you can't write the notes, you're not ready for cards. (More on why in how to use AI to actually remember what you read.)
2. Generate with the audit built in. Use the prompt above, or any version that keeps the three rules: only from my notes, recall questions not recognition, and a separate "Check this" list for anything doubtful. Resolve every "Check this" item against a real source before importing.
3. Hand scheduling to a scheduler. Import into Anki (or any app with real spaced repetition) and let it decide when you see each card.
The same split applies to learning a language with AI around a full calendar. For where flashcards sit among reading, notes and AI tutoring, see how to build a personal AI learning system.
Where this breaks
Flashcards are for facts and relationships, not judgment. You can drill what a participating preference is. You can't drill when to walk away from a term sheet. That needs argument and pressure-testing, and a model that sounds sure is not evidence it's right — see the study on AI confidence versus accuracy.
The audit is only as good as your follow-through. My test is one run, one model, seven lines of notes. A subtler error or a longer document could slip through. "Check this" is a prompt to verify, not a verification.
Generation can quietly replace the notes step. Pasting a whole article and asking for 40 cards gives you a deck made from text you never processed — the Wharton crutch pattern with a flip animation. If that's already a habit, how to stop relying on AI for studying is the more useful read.
Scheduling has a cost. Daily reviews are a commitment. Keep decks small and delete cards you no longer need; a founder learning five domains at once needs fewer cards, not a better app — the lesson of keeping up with AI without losing your focus.
The honest summary: this autumn's launches made job 1 nearly free. They didn't touch jobs 3 and 4, and job 2 got more important, not less. The machine can write the cards. Only you can remember them.
Sources
Flashcards in ChatGPT, OpenAI Help Center, 2026.
ChatGPT Now Makes Flashcards You Can Actually Flip, SpillbyBill, checked October 4, 2026 (September 22, 2026 launch on all plans).
Best AI Study Tools 2026: 9 Compared, With Prices, Fora Soft, prices checked October 10, 2026 (ChatGPT quizzes August 14, GPT-6 Sol/Luna in ChatGPT, Gemini Notebook rename, limits and quiz formats, Anki pricing).
GPT-6 Sol and GPT-6 Luna: October 2026 update, OpenAI Deployment Safety Hub, October 7, 2026.
Deck Options (FSRS, desired retention, "Hard" vs "Again"), Anki Manual.
Importing Text Files (field separators), Anki Manual.
Test-enhanced learning: Taking memory tests improves long-term retention, Roediger & Karpicke, Psychological Science, 2006.
Generative AI Can Harm Learning (summary of Bastani et al., Wharton), TeachingTimes.
Series A Term Sheet – Liquidation Preference, Pillar Legal (Carta data on 1x non-participating).