Every founder I talk to is running some version of the same experiment: "Can I compress a hard new skill into a fraction of the time using AI?" Corporate strategy, hiring frameworks, financial modeling, prompt engineering, contract law, the specific domain of whatever partnership they are about to sign. The urgency is real — you cannot delegate what you do not understand at the founder level — and the tools are visibly more capable each quarter. So the question is not whether to learn with AI. It is: what actually works, and what quietly makes you worse.

I have been running this experiment on myself for about eighteen months, first as an AI architect prototyping systems, then more deliberately as a coach watching other founders do the same. I want to share the workflow I actually use, the tools that survive July 2026 scrutiny (not the ones that trended last year), and the one MIT finding that made me tear up half of what I was doing.

The trap: AI as a shortcut versus AI as a tutor

In June 2025 a team at the MIT Media Lab ran a study called Your Brain on ChatGPT. Fifty-four adults wrote essays across several sessions, split into three groups: ChatGPT-assisted, Google-assisted, and brain-only. Electrodes recorded neural activity. The ChatGPT group showed the lowest neural connectivity, weakest memory of what they had written, and — the detail that stayed with me — 83% of ChatGPT users could not quote a single line from an essay they had just produced. The Google group scored in the middle. The brain-only group had the strongest and most widespread neural networks. When ChatGPT users were later asked to write without the tool, their brain engagement stayed low. The authors called it cognitive debt.

The follow-up work in 2026 pointed the same direction. A separate MIT Media Lab study on fake-news detection found that participants who leaned on AI assistance saw their unassisted detection accuracy fall roughly 15 percentage points by the end of the trial. The tool helped in the moment; the underlying capacity atrophied.

This is not a case against AI-augmented learning. It is a case against the default mode most people use — dumping a question into a chatbot and reading the answer. That mode is a shortcut, and shortcuts do not build skill. They build the illusion of skill, which is worse than not knowing, because you stop asking.

Cal Newport's Deep Work framed the mechanism years before ChatGPT existed: skill is myelinated pattern-recognition, and myelination requires effortful retrieval. If AI does the retrieval, no myelination happens. The frame I use with founders now: the tool has to make me think harder, not less. Every workflow below is judged against that bar.

The four learning modes I actually use

I sort AI-assisted learning into four modes. Only two of them build durable skill. The other two are useful for other things — knowing which is which is the whole game.

Mode 1: Answer extraction (the shortcut — do not confuse this with learning)

You ask ChatGPT or Claude a question, you read the answer, you close the tab. Fast, satisfying, forgotten by Thursday. Fine for looking up the difference between an ISDA and a GMSLA at 11pm. Not learning. If this is your default, expect the MIT result: no myelination, no retention, growing dependence.

Mode 2: Socratic dialogue (real learning, if you resist the temptation to skip)

OpenAI shipped ChatGPT Study Mode in July 2025, available on every plan including free. Instead of answering, it asks. It quizzes one question at a time, layers explanations from simple to detailed, and checks your understanding before moving on. It uses a Socratic frame — the same frame Sir John Whitmore built Coaching for Performance around and the same frame every good executive coach runs. It is, structurally, a coach.

The catch: Study Mode only works if you actually try to answer before it shows you. If you flick past the prompts to see what it "thinks," you are back in Mode 1. I set a personal rule: no scroll until I have typed my answer, even if it is wrong. Especially if it is wrong — the productive struggle is the point.

Mode 3: Deep-source synthesis (where NotebookLM earns its keep)

Google's NotebookLM, now powered by Gemini 3.5 with secure cloud code execution and audio transcription (added January 2026), does one thing better than every general chatbot: it grounds itself in your sources. Upload the ten papers, three books, and the Q3 board deck; ask it questions; the answers cite the specific page in the specific source. When you disagree, you can go to that page in seconds. Its Audio Overview feature turns any notebook into a fifteen-minute two-host conversation — I use these on runs, and the compression forces me to notice which parts my brain skims over.

For founders learning a new domain fast, NotebookLM is the tool I recommend first. It works because the source discipline forces you to build a real corpus (which is itself half the learning) and because knowing you can verify every claim makes you willing to argue with the AI. Arguing is where the myelin gets laid down.

Mode 4: Long-context reasoning partner (Claude Projects for the hard problems)

Claude Projects — with the current Sonnet 5 / Opus 4.8 line and a one-million-token context window — is my third layer. I keep a persistent project per domain (Fundraising 2026, Compensation Design, Delaware C-Corp, etc.) with the primary sources, my running notes, and my open questions. I do not ask it for answers. I ask it to pressure-test mine. "Here is the term sheet I would push back on and why — where is the argument weakest?" That is red-team learning, and it works because Claude has enough context to challenge specifics, not generalities.

Anthropic's July 2026 launch of Claude for Teachers is a signal about where the tooling is going, but that product is aimed at K-12 educators; the general Projects surface is what founders should be using and it is available on the standard paid tier.

A comparison for founders who want to pick one tool

If you are going to add exactly one AI-assisted learning tool to your week and stop reading roundups, this is the honest breakdown as of July 2026. Prices and features drift; verify before you subscribe.

ToolBest forCost (July 2026)Weakness
ChatGPT Study ModeLearning a defined topic from zero (frameworks, exam-style knowledge)Free tier includes it; $20/mo PlusOnly useful if you resist the "just tell me" instinct
NotebookLMDeep dive into a specific corpus (a domain, a company, a body of research)Free; Plus tier via Google One AI Premium ~$20/moOnly as good as the sources you upload; garbage in, confident garbage out
Claude ProjectsOngoing reasoning partner for a hard domain you keep coming back toPro $20/mo; Max tier $100/mo for heavier useEasy to slide into "let Claude decide" — needs disciplined red-team framing
PerplexityFresh-information lookups with citations (news, prices, papers)Free; Pro $20/moLookup tool, not a learning tool — use Mode 1 rules

My own stack: Study Mode for defined topics, NotebookLM as the primary long-form learning workspace, Claude Projects as the pressure-test layer, Perplexity for lookups. I do not use ChatGPT's regular chat for learning at all — it is too easy to slip into Mode 1 there.

The DRILL protocol I actually use

The protocol below is what I run when I have to genuinely learn a new domain in weeks rather than months. It is named to be remembered, not to be clever. It has survived twelve or so real applications now — fundraising instruments, board dynamics, LLM eval design, HR compensation bands, EU AI Act obligations — and it is the workflow I hand any founder who asks.

D — Define the outcome, not the topic. Before opening any tool, write one sentence: "In four weeks I need to be able to do X specific thing." Not "understand fundraising." Rather: "In four weeks I need to be able to walk through our seed term sheet clause by clause with a lawyer and know which two clauses I would fight for." Outcome-defined learning is the difference between a founder who ships and a founder who reads.

R — Retrieve the corpus into NotebookLM. Ten to fifteen primary sources — books, papers, model documents, a couple of good long-form podcasts transcribed. Not blog posts. The act of choosing sources is 20% of the learning. Generate an Audio Overview and listen once for the shape.

I — Interrogate in Study Mode. Open ChatGPT Study Mode and work through the twenty questions a novice would ask, in order. Type your best guess before reading the response, every single time. If you cannot even guess, that is a signal to go read a source in NotebookLM first.

L — Load a Claude Project. Same sources, plus your Study Mode notes and your open questions. This becomes your persistent workspace. Ask Claude to pressure-test your emerging positions — never to tell you what to think.

L — Live-test in one real conversation. Within a week, have one conversation with a real practitioner (lawyer, operator, investor) where you use what you have learned. The retrieval-under-social-pressure is where the myelination completes. AI-assisted learning without a real conversation stays fragile.

The protocol is deliberately slower than "just ask ChatGPT." That is the point. Kenneth Stanley's argument in Why Greatness Cannot Be Planned — that novelty and depth come from following genuine curiosity through steppingstones, not from optimizing for a defined endpoint — applies here in a specific way. The optimizing-for-fast-answers workflow flattens your understanding. The DRILL sequence forces the branching that makes the knowledge yours.

What I would not do

Three specific traps I have watched founders fall into over the last year.

Do not use AI for the first pass of your own thinking. If you are trying to form a view — on a hire, a market, a partnership — write your draft view first, in your voice, before you show it to any AI. The MIT ownership finding matters here: 83% of ChatGPT users could not quote their own writing because the writing had never really been theirs. The same happens with strategy. Form the view, then have Claude challenge it.

Do not confuse breadth for depth. Being able to converse about twelve trending frameworks is not learning; it is a party trick. Pick one domain per quarter that you go deep on with the DRILL protocol. Everything else stays Mode 1 — lookups, not learning.

Do not skip the human conversations. Every founder I have seen level up fast with AI learning also did the real reps: the coffee with the practitioner, the operator dinner, the awkward call to the board member with a hard question. AI compresses the reading; humans still compress the judgment.

The uncomfortable summary

AI-assisted learning works, and it is a significant edge for founders who use it well. But "using it well" turns out to mean the opposite of what the marketing suggests. It means slower, not faster; more effortful, not less; more source discipline, not less; more argument with the model, not less. The founders I have seen genuinely upskill with AI in 2026 all describe it the same way — it is like having a very patient, very well-read intern who will disagree with you if you ask them to. The founders who plateaued all describe it as an oracle. The oracle mode is the trap.

Pick one domain this quarter. Run DRILL on it. Have three real conversations that use it. See what happens.

Sources

If this helped, these related answer pages go deeper on adjacent questions:

Share this post