Copy article

The Reinforcement Gap: Why AI is Mastering Code but Stalling at Poetry

ended 07. October 2025

Techcrunch has reported AI’s progress is wildly uneven. Systems like ChatGPT-5, Gemini 2.5 and Sonnet 4.5 can now debug, refactor, and build software faster than most humans. Yet in the creative arts, the story is very different. Ask AI to write a poem or screenplay and you’ll get roughly the same quality of result you did a year ago. One field is accelerating at breakneck speed and the other is stuck in neutral. 

The Reinforcement Gap: Measurable vs Meaningful

According to METR (Model Evaluation & Threat Research), the length of software tasks frontier AI agents can complete autonomously with 50 per cent reliability has been doubling every seven months. That exponential progress is driven by reinforcement learning (RL), the same method that a manager uses to train a subordinate, offering feedback and praise for getting closer to the correct answer, again and again.

STEM disciplines are perfect for this. Code either compiles or it doesn’t; a physics simulation obeys gravity or it breaks. Those clear pass–fail conditions create billions of testable outcomes, fuelling continuous, measurable improvement.

Creative fields don’t work that way. Poetry, prose, and screenwriting don’t have unambiguous success metrics. One person’s “padding” is another’s “immersive world-building”. There’s no objective “creativity score” to optimise against. The result is a widening reinforcement gap. The measurable is advancing because it can be measured, while the meaningful is stalling because it can’t.

When Investment Follows the Metrics

Silicon Valley is doubling down on what’s easy to test. Billion-dollar RL “environments” are teaching agents to ace synthetic benchmarks, solve multi-step puzzles and build ever-larger software systems, but not to hold a convincing, compassionate conversation. Funding is flooding into functionality that knows every rule of logic yet still struggles to feel.

Anyone who’s tried to get AI to write humour, irony, or human warmth, or experienced the fight redrafting the company “About Us” page, knows the problem. 

The models can pass a coding test but not a Turing-grade moment of connection.

We’d like your views:

  • If reinforcement learning only rewards what’s measurable, are we designing an economy that values compliance over creativity?
  • Can AI ever master the messy brilliance of poetry, humour, or human judgement?
    Should we be developing new, human-centred metrics that recognise depth and meaning, not just mechanical accuracy?
  • And how do we train the next generation of “AI architects” to preserve the creative spark that code alone can’t generate?

Because if we don’t, we may end up with a world that’s technically brilliant and emotionally tone-deaf.

We design these systems. We own the gap.

2 responses from the Newspage community

Copy all

Star Quote
Copy

Colette Mason
AI Ethics Consultant at Clever Clogs AI
n a cruel twist of fate, skilled creatives who built the raw material these models feed on are being left behind as end users dump them and try to recreate the magic in-house. The fix isn’t pity; it’s precision, and the need is pressing. They need better-trained AI: solid guardrails to stop hallucinations and the very best worked examples to teach nuance, rhythm, and emotion. Show in your uploaded training, don’t snap at the prompt box. Explain why a phrase lands, why a beat matters, why a pun fails, why a colour choice feels off. In a world without pass or fail tests, craftsmanship is the strongest currency. Pour your black and white principles into your copilots. Train your model the way you’d train a colleague, with context, feedback, and respect for the work. Because the nature of soulfulness means it will never be low hanging automatable fruit.
Copy

Editor
Mitali Deypurkaystha
AI Governance Strategist & CMO at AISensa
Is it any wonder AI mirrors who we’ve become? For decades, we’ve quietly underfunded the arts and overvalued what can be measured. Schools treat art, music and drama as luxuries, while maths and science are seen as “serious” subjects. AI has simply absorbed that bias and magnified it. It’s optimising for what we’ve long rewarded: precision, not perception; compliance, not curiosity. Reinforcement learning trains systems to please the metric, not move the mind. So, of course, it masters code and stalls at poetry. How do we fix this? By fixing ourselves. We must remember that the arts are not optional extras. They are the architecture of humanity. Studies show that learning arts, music, languages, or drama strengthens confidence, collaboration and communication. As I see daily as an AI strategist, the ability to communicate clearly and creatively isn’t just vital for people; it’s becoming essential for working with generative AI. The more human we are, the smarter our machines become.