Practical principles for using AI at work, with Ethan Mollick

Executive overview

Most people have barely scratched the surface of AI — a few prompts on the free tier, no revelation moment. The gap between casual use and genuine understanding is vast, and closing it requires sustained experimentation with the best available models.

AI has a jagged frontier: it excels at things you wouldn't expect (writing sonnets, passing bar exams) and fails at things that seem trivial (counting exact words). The only way to map that frontier for your own work is to bring AI to everything and use it where you already have expertise.

The core insight: treat AI like a brilliant alien colleague — give it a role, converse with it naturally, and stay in the loop to catch what it gets wrong.

The jagged frontier of AI ability

• A BCG experiment with 8% of global staff showed 40% quality improvement, 26% speed gain, 12.5% more output — on tasks AI handled well.
• One task was deliberately designed to look AI-solvable but wasn't; participants using AI made more mistakes by falling asleep at the wheel.
• AI scored 90th percentile on the bar exam; GPT-3.5 scored 10th — capability can jump dramatically between versions.
• It can write a perfect sonnet but struggle to write an exact 25-word sentence (it sees tokens, not words).
• Knowing what AI is good or bad at requires hands-on experimentation, not reading tips.

Always invite AI to the table

• The "P" in GPT stands for pre-trained — it already knows a vast amount; the question is what it can do for you.
• Bring it to everything you legally and ethically can: meetings, brainstorming, creative tasks, writing.
• You don't have to use its output; the goal is to learn the shape of its capabilities.
• Experimentation is cheap for individuals and expensive for organizations — this is your advantage.
• Viral prompting tips are mostly conditional; don't over-engineer, just start conversationally.

Treat AI like a person — but tell it what kind of person

• The best prompters are managers, teachers, and parents — people skilled at working with others, not coders.
• Give it a role and context: "You are an expert marketer who works at X" before asking anything.
• Iterate like a manager: give feedback, say what was wrong, ask it to improve a specific section.
• Humanities majors and people-oriented professionals often outperform technical users because the models emulate human communication.
• Anthropomorphizing has risks (reduced critical thinking, susceptibility to AI mimicking people) — hold the model as an "alien person," not a human.

Be the human in the loop

• Originating from control systems (NASA, military): always keep a person positioned to intervene.
• AI tends to do the boring parts of your job better than you do — that's a feature, not a threat.
• Whatever you're in the top 1-10% at, you still beat any current AI system.
• The goal is to concentrate on high-value work and outsource the rest, not to abdicate judgment entirely.
• Staying in the loop means you'll see signs of AI's growing capability earlier than those who avoid it — giving more time to adapt.

Hallucination and verification

• Hallucination rate on medical citations: 80% (GPT-3.5 free) → 20% (GPT-4) → lower still with web search enabled.
• The danger isn't frequency — it's plausibility. Wrong answers feel correct.
• Use AI in domains you know well so you can catch errors immediately.
• Starting in unfamiliar areas is high-risk: you can't tell a bad recipe from a good one if you can't cook.

Privacy in practice

• Recording anyone requires their knowledge and consent — that's a separate, non-AI ethical question.
• ChatGPT-4 has a private-mode toggle in settings; paying more gets private-by-default.
• HIPAA- and FERPA-compliant versions are available from Anthropic and Microsoft.
• AI learns patterns, not exact copies; the risk of specific data being recalled from one or two interactions is minimal.
• Privacy concerns shouldn't be the barrier — solutions exist for most use cases.

Assume this is the worst AI you will ever use

• Treat current limitations as temporary; the technology has improved dramatically in one year and shows no sign of stopping.
• Organizations and individuals treating this as something to "wait and see" are already falling behind.
• GPT-4 (the best available model at time of recording) is free in 169 countries via Bing — competitive advantage is not gated by access.
• Urgency matters: others are experimenting now.

Starting if you haven't yet

• Begin with work tasks in your area of expertise — not areas where you can't evaluate the output.
• Use conversational prompting over complex prompt engineering.
• Tell it who it is, give feedback, iterate — that's the loop.
• The revelation moment (three sleepless nights) comes from sustained use of the best available model, not a brief free-tier try.