Blog

How Relying on AI Kills Creativity (and How to Bring Back Real Innovation)

Written by Colin Nelson | Jun 25, 2025, 10:24:29 AM
 

 

AI can make you faster—and blander. The art is knowing where machines should amplify people, not replace them.

The paradox

A French automaker invited 10,000 engineers to propose new digital features. The prompt was broad; the response was huge: 1,500 ideas and 3,500 comments. Humans couldn’t read it all, so popularity became the proxy. Many original, fringe ideas likely died unseen.
Lesson: At scale, human-only review fails—but AI-only ideation flattens originality. The answer is a deliberate human-in-the-loop model.

What AI is good for (today)

Think in four verbs:

  1. Augment – spark and enrich human creativity (seed ideas, generate visuals).
  2. Assist – digest and structure large volumes (cluster, deduplicate, summarize).
  3. Accelerate – speed the pipeline (faster triage, faster synthesis, quicker tests).
  4. Automate – remove repetitive toil (translation, routing, tagging, notifications).

The real headaches AI can relieve

  • Inspiration on demand
    Give teams curated “starter kits”: adjacent examples, patterns, quick AI visuals. Idea quality rises when minds have something concrete to push against.
  • Overlap and duplication
    Use semantic clustering to merge near-identical ideas expressed in different words, languages, or formats. Stop seven teams solving the same problem.
  • Volume triage
    Let models score by novelty, feasibility markers, and thematic fit—so humans spend time on the right 10–15%, not the loudest 10–15%.
  • Language barriers
    Live, high-quality translation allows people to submit and collaborate in their native language without losing nuance.
  • Curation of the “innovation lake”
    Mine years of ideas, insights, and experiments to:
    – Resurface shelved concepts now made viable by new tech.
    – Spot emerging themes across business units.
    – Auto-connect people with shared interests across time zones.
  • Partner scouting
    Crawl public signals to find startups/scaleups and set standing searches that alert you when something relevant appears—without manual sleuthing.

Where AI falls short (and why that matters)
  • Data you can’t see, you can’t use
    Much meaningful knowledge sits behind firewalls or in tools the model can’t access.
  • Tacit knowledge lives in heads
    The “click” between experience and weak market signals often isn’t written down. Humans still outperform on vague, early cues.
  • Assets & competencies aren’t codified
    If you haven’t documented what you’re truly good at (beyond current products), AI can’t recommend credible adjacencies.
  • Mid-tier idea bias
    Out-of-the-box model outputs tend toward “competent average,” not breakthrough.
  • Decision context
    Final portfolio calls depend on constraints (capacity, regulation, brand risk) that models typically don’t internalize well.
  • Multi-company projects
    IP, culture, and security realities limit cross-org data sharing; automation helps, but doesn’t erase governance work.

 

A practical operating model (people + machines)

1) Design the front door

  • Write prompts/briefs that are specific enough to control volume and protect IP.
  • Require a minimum viable data set for every submission: effort, investment, expected impact, time horizon, confidence.

2) Machine first, human final

  • AI clusters, scores, summarizes.
  • Humans review clusters, not individual ideas; they make the advancement/kill decisions.

3) Build the “Innovation Shelf”

  • Park “good but not now” concepts with lightweight metadata; routinely re-scan with AI as tech/regulations change.

4) Make signals a habit

  • Quarterly cadence: trend/tech scans + customer executive dialogs → update focus themes → refresh prompts.

5) Guardrails

  • Clear data boundaries, opt-in sources, audit trails, and explainable scoring.
  • Recognition for stopping as well as shipping (to avoid sunk-cost bias amplified by automation).

 

30/60/90-day rollout

Days 1–30 (crawl)

  • Enable translation, clustering, and ranking on one active challenge.
  • Pilot AI image generation to visualize top ideas.

Days 31–60 (walk)

  • Ingest historic idea/insight data; create a searchable map.
  • Stand up the Innovation Shelf and a monthly “reactivation scan.”

Days 61–90 (run)

  • Launch automated startup scouting for 1–2 themes.
  • Add a portfolio view with weighted impact (impact × confidence × time).

 

What to measure

  • Throughput: time from submission → human decision.
  • Uniqueness score: % of ideas not clustered into existing themes.
  • Reuse rate: shelf → reactivated concepts progressed.
  • Partner hit-rate: scouted → qualified → piloted.
  • Decision quality: concept survival to next gate; post-mortems on kills.
  • Engagement equity: contributions and wins across languages/regions.

 

Do this, not that

  • Do use AI to find, group, and visualize; don’t outsource judgment.
  • Do seed with examples and images; don’t let models set the brief.
  • Do codify assets/competencies; don’t assume products = strengths.
  • Do reward prudent kills; don’t equate persistence with performance.

 

Bottom line

AI doesn’t kill creativity—uncritical reliance on AI does. Treat models as powerful exoskeletons for human imagination: excellent at volume, structure, and speed; reliant on people for context, leaps, and taste.