AI Will Excel at Verifiable Tasks, Lag at Creative and Strategic Tasks: Andrej Karpathy
Here's a simple way to think about AI's near future: it gets scary good at anything with a clear right answer, and stays clumsy where taste and judgment matter. That's the core of Andrej Karpathy's take on how progress will roll out.
For creatives, this isn't a threat. It's a playbook. Offload the parts of your work that can be checked, scored, and iterated. Keep the parts that need taste, context, and strategy.
What "verifiable" actually means
Karpathy frames it like this: AI thrives where outputs are easy to verify. Code compiles or it doesn't. A math proof is correct or it isn't. If a task is verifiable, you can train a model against it, practice it at scale, and improve fast.
He calls out three conditions for practice: the environment must be resettable, efficient, and rewardable. Translation: you can try again, try often, and get automated feedback on whether it worked. That's why reinforcement learning shines in these domains.
His summary hits clean: "Software 1.0 automates what you can specify. Software 2.0 automates what you can verify."
Why creative and strategic work lags
Creative direction, brand voice, story arcs, campaign strategy-these rarely have a single "correct" answer. They depend on taste, timing, and context. You can't reset a product launch a hundred times this week. You can't get instant ground truth on a rebrand.
So AI improves here, but slower. It can imitate, suggest, and remix. It struggles to decide. That's not a bug; it's the nature of work without clear verification.
What to offload now (high-verifiability tasks)
- Generate and refactor code snippets for automations, data cleanup, and creative tooling.
- SEO checks, grammar fixes, headline scoring, readability passes.
- Batch asset tagging, alt text, file naming, transcript generation, and formatting.
- Style enforcement: color contrast checks, spacing rules, brand voice constraints.
- Variant generation for A/B tests: headlines, hooks, thumbnails, CTAs.
Make creative work more verifiable (without killing taste)
You can engineer feedback loops into creative projects so AI becomes more useful. Keep the human in charge, measure the parts that can be measured, and let the system iterate.
- Define fast metrics: CTR, watch time, scroll depth, conversion, response rate.
- Run resettable experiments: multiple versions in parallel, short cycles, clear stop rules.
- Automate rewards: pipe metrics into a sheet or dashboard and score variants automatically.
- Constrain for quality: style guides, checklists, reference boards, banned phrases, length limits.
- Tight prompts + examples: show 3-5 on-brand samples and require structured outputs.
Keep these human (for now)
- Positioning, narrative, and big creative bets.
- Cross-channel coherence and timing.
- Taste decisions where data is thin or delayed.
- High-stakes calls you can't "reset."
A simple AI workflow for a creative team
- Brief: Write a tight objective and success metric.
- Explore: Use AI to produce 20-50 rough variants under clear constraints.
- Filter: Human selects the best 3-5 based on taste and context.
- Test: Ship small. Track metrics automatically.
- Refine: Feed results back into the prompt and constraints. Repeat.
- Ship final: Human makes the call, AI handles polish and packaging.
What this means for your career
Lean into systems thinking. Split your craft into parts that can be measured and parts that can't. Let AI grind the measurable parts, and spend your time on direction, insight, and voice.
If you build loops that are resettable, efficient, and rewardable, you'll pull ahead while others debate prompts.
Further context
Karpathy's framing explains why coding and math are moving fast, while roles like CEO remain human. Those jobs have feedback, but not the kind you can iterate a thousand times by Friday.
Curious about the person behind the idea? Start here: Andrej Karpathy.
Tools and training for creatives
Want a curated list of AI tools for writing and concepting? Check these resources:
Bottom line: make more of your process verifiable, and you'll get compound gains from AI without losing your edge in taste and strategy.
Your membership also unlocks:
