Stimulate, Don't Replace: How AI Design Shapes Human Creative Effort

Why we wrote it

The European Commission is preparing an AI strategy for the cultural and creative sectors on the principle that AI should complement and not replace human creativity. We asked what the evidence says about how that principle can be built into a tool, and what it implies for public policy. The paper reviews seven bodies of research and reports only findings from sources we opened, with their statistics as stated. It was submitted as evidence to the Commission’s call in September 2026.

What the evidence says

1. 01Working with generative AI raises individual output and lowers collective diversity.

   Across the studies reviewed, ideas produced with AI were rated better (g = 0.27) while the diversity of ideas across people fell sharply (g = −0.86). A second meta-analysis puts the homogenisation effect at d = 0.33.

2. 02The interface decides which effect you get.

   An AI that questions the person about their own idea preserves diversity and ownership. An AI that rewrites the idea loses both (d = 0.76 and 0.57 for the loss).

3. 03Answers help now and hurt later. Hints do no harm.

   In learning, an answer-giving AI raised assisted performance and lowered unassisted performance by 17%. A hint-giving AI did no harm.

4. 04High-information feedback beats corrective feedback.

   Feedback that explains has about twice the effect of feedback that corrects (d = 0.99 against 0.46).

5. 05The evidence against a simple Socratic thesis is reported too.

   One randomised trial found a strategy-giving AI coach left users worse off than no AI at all. Another found no difference between Socratic and non-Socratic tutors. There is no reliable neural signature for divergent thinking.

What it means for how CineCoach is built

• The student writes before the AI speaks. Nothing is generated from a blank page.
• The AI never writes the scene. It asks, it names the problem, it points to the next step.
• Strong work gets no suggestions. Feedback comes per dimension, in words, and ends with what to do next.
• Fifteen languages, each with its own natively analysed film curation. Eleven are official EU languages.

What we do not claim

• CineCoach was built in February and March 2026. The evidence was reviewed in August. The design is consistent with the evidence, not derived from it.
• The system runs on a United States foundation model. The European layer is the expertise, the languages, the curations and the governance.
• There is no controlled study of CineCoach yet. The paper says what would be needed to run one.

Recommendations for the European strategy

1. Fund the design, not the technology. Score “the human generates and the AI questions” in calls for artist-facing AI, and require every funded system to say which design it is.
2. Measure diversity of output as an outcome, alongside adoption.
3. Put artist-facing AI under the control of the schools and studios that use it.

The paper makes six recommendations in full. These are the three carried into the Commission submission.

Cite as

Bakshi, A., Briquet-Laugier, V., & Wiser, P. (2026). Stimulate, Don't Replace: How AI Design Shapes Human Creative Effort. The CineCoach Case and Recommendations for European Policy. Working paper v4. Zenodo. https://doi.org/10.5281/zenodo.22514980