Skip to content
Part of Human Value in AI ↗Family overview
Human Creative Value in AI

Examples

Five reframings, outside art

These five reframings are constructed examples, not client stories or proof. A sensible first instinct can be wrong. Separately labelled, director-reported cases follow.

Case 1

Clinic communication

Illustration

Write clearer discharge letters with AI.

→ Help patients know what to do at 2 a.m. when something feels wrong.

Human moves

  • Asked who reads the letter. Often a carer, not the patient.
  • Questioned the “one letter” limit. Added a single-question card.
  • Asked a night-shift nurse as well as a physician.

EthicsProtect health information. Get consent before putting patient data in a tool. The clinician remains responsible for advice.

Case 2

Product opportunity

Illustration

Add AI features so we keep up with competitors.

→ Find the moment where customers lose confidence and quietly leave.

Human moves

  • Questioned whether “more features” means more value.
  • Combined support logs, sales objections and interviews about why customers leave.
  • Pictured a future where the feature works too well and users stop learning.

EthicsSay what is automated. Do not manipulate users. Let them override recommendations.

Case 3

Education

Illustration

Use AI to grade essays faster.

→ Give each student one piece of feedback they will actually act on.

Human moves

  • Changed the focus from saving teacher time to helping students act.
  • Recognized an experienced teacher’s skill in choosing one next step.
  • Set checks: feedback must be specific, kind and open to revision.

EthicsBe fair across language backgrounds. Tell students when AI is involved. The teacher decides.

Case 4

Service redesign

Illustration

Build a chatbot to reduce call-centre volume.

→ Understand why people need to call twice.

Human moves

  • Problem found in repeat contacts, not first contacts.
  • Asked front-line staff who know where the service fails.
  • Alternative future: fewer, better human calls rather than none.

EthicsRespect people. Guarantee a route to a human. Include people who cannot use chat.

Case 5

A correction that overshoots

Illustration

It came back undercooked — cook it more.

→ Adjust the doneness without burning what already works.

Human moves

  • Named the intended middle: warm and pink at the centre, not “more cooking”.
  • Kept the heat, seasoning and timing that were already right.
  • Set checks a cook could judge, beyond a number on a timer.
  • Asked the cooking specialist, not just the machine operator, to judge the result.

EthicsA correction can go too far. Whoever sets the target must describe the intended middle, not just the fault.

This is a constructed analogy, not one of the dated cases. The design-drift case below showed unrequested changes among retained candidates. A correction can fix one thing and change others. Rework, review time and human hours were reported but not measured. A second AI can prolong the loop without an independent, qualified standard. Automated checks can help. They do not set that standard.

Historical notes · two separate cases

Why human review stayed necessary

A homepage design review · design drift · 11 Nov 2026

  • FactThe user spotted unrequested design changes among retained AI candidates during a homepage review.
  • FactA review burden was reported. It was not measured.
  • InterpretationSpotting drift means keeping the original purpose in mind. That is creative work, beyond a quality check.

A comparison hero illustration · creative-director review · 12 Nov 2026

  • FactOn a different site, the owner selected another hero option. They then reported an apparent figure with two left hands in a comparison image.
  • FactWhether it was repaired has not been verified.

Evidence case · a repeated connector loop

When repeating the instruction stopped helping

  • FactDuring domain-connection support, the same confirmation dialog kept returning after assistant suggestions. Its Continue button stayed disabled. The user reported five failed attempts.
  • FactThe cause was unverified. Time and cost were unmeasured.
  • InterpretationThe useful move changed the task from “repeat the click” to “why does this gate persist, and what genuinely different route can we test?”

Evidence case · a director-reviewed film cut

A completed film was not an accepted performance

An unreleased animated review cut of about 60 seconds. Generated clips existed. The director did not accept them as the film. No film is shown here.

5.2/10

Director’s subjective score, not an audience test. Not approved as final.

The brief

The brief wasn’t just “60 seconds”.

It asked for characters with chemistry, specific gestures and an exact ending.

The review

The film existed, but the acting didn’t land.

Weak mouth movement and lip sync, little chemistry, flat pacing, background music missing or inaudible.

Four human moves

  1. 1Director-reportedSpecified character chemistry, gestures and the exact ending.
  2. 2Director-reportedNoticed that sound and mouth movement did not match.
  3. 3Director-reportedWithheld final acceptance and rejected the cut.
  4. 4Proposed corrective actionProve one short speaking scene before generating more.
  • Director-reportedTen generation jobs completed (about 76 seconds of clips) plus 60 seconds of music. The director had not accepted the material. Some clips may still be reusable.
  • InterpretationThe target changed from “generated clips” to “an accepted performance”. This does not mean AI cannot be creative.
  • UnmeasuredHuman hours, total film cost and lost revenue.

Provenance: the owner’s private research record.

Reframe “generate more” into “prove one speaking shot”.

Open the Studio →

Case study · director-reported

When the edit removed the story →

A line about a drawing the frame had cropped away.

Try your own reframing →