Less than we expected. When the brief described the person well, those details drove most of the ideas and the relationship label made only a small difference. A partner and a coworker with the same hobbies still got nearly the same four directions.
That cuts both ways. There were no romantic clichés: not one piece of jewellery or engraved keepsake for a partner. But the model also made only small adjustments for how well you know someone. Sleep gadgets went to coworkers. Gift cards were the main concession to distance.
What we tested
The same person can be somebody's partner, somebody else's coworker and a neighbour's acquaintance. Their hobbies do not change with the giver. The gift probably should. A perfume or a weekend away can be lovely from one person and awkward from another.
So we kept every fact the same and changed one word:
Recipient: partner, early 30s. Birthday. Budget up to £100. He is getting into sourdough baking and runs at lunchtime. He works from a tiny spare room at home and starts the day before 7am. He has been sleeping badly since his workload went up. He finds most aftershaves too strong and never wears a watch.
The same brief was then run with best friend, coworker and acquaintance. The budget stayed the same on purpose: if the model chose cheaper or safer ideas for a coworker, we wanted it to be the model's choice, not ours.
We used 13 of our 16 fictional people. Three were left out because their facts gave the relationship away (office details) or because "partner" would be inappropriate (a 15-year-old). Each version ran three times in a fresh session, in random order: 156 answers, 624 ideas. The rater never saw the relationship word.
Did the ideas change?
Barely. Here is one person, first run of each version:
- Sourdough baking equipment
- Sunrise alarm clock
- Weather-resistant running jacket
- Weighted blanket
- Sourdough baking accessories
- Running socks and winter gear
- Sunrise alarm clock
- Compact desk accessories
- Sourdough baking accessories
- Running sunglasses or belt
- Sunrise or white noise alarm
- Space-saving desk organisers
The partner got a weighted blanket, the coworker got desk accessories. Otherwise the lists are the same person's lists. Every version reached for the sunrise alarm because he sleeps badly, including the coworker version, where a gift about someone's sleep could feel a little personal.
How personal did the gifts get?
We scored every idea for intimacy, sentimentality and social risk, without the rater knowing who the gift was for.
Average intimacy of an idea (1 = impersonal, 5 = romantic)
- Acquaintance1.29
- Coworker1.41
- Best friend1.44
- Partner1.55
The ordering is right. The size is small. Partner gifts were a little more likely to be an experience (11 ideas against 2 to 7 for the others) and a little more likely to be something done together, but nothing came close to a romantic gesture.
Did "partner" trigger clichés?
No. We were watching for the obvious relationship stereotypes: jewellery, perfume, engraved keepsakes, romantic weekends, photo gifts. Across all 156 partner ideas there was no jewellery, no keepsake and two beauty products in total.
If changing the label had swapped the person's facts for a romance template, that would have been stereotyping. It did not happen.
Where the word "partner" did show up was in the small print. In 5 of 39 partner answers the rater flagged a relationship detail the brief never gave. Looking at them, most were reasonable and openly labelled: a sunrise alarm that might not fit "if he shares a bed with you", or a massage voucher that "assumes you can take sole charge of the children". That is the label doing what it should, as a stated assumption rather than a fact. The rest were simple mix-ups of who was who.
Did "coworker" make the gifts safer?
A little, mostly through gift cards.
Gift cards and vouchers, out of 156 ideas per relationship
- Partner8
- Best friend7
- Coworker15
- Acquaintance17
Social risk scores stayed low for every relationship (1.2 to 1.35 out of 5), which tells you the model was not suggesting anything risky for anyone. It was not suggesting anything especially personal for partners either.
Why the relationship label mattered less than the person details
Our briefs were detailed: interests, habits, something going on in the person's life, a dislike. Those facts gave the model plenty to work with, so they set the direction. One word about the relationship could not compete.
That is mostly good news. The model did not invent a personality from a label. If the relationship should change what feels appropriate, the label alone may not carry enough information.
What to tell AI about the relationship
A label carries very little information. A short sentence can add the part that actually matters: how close you are and what kind of gift would feel appropriate.
She is my coworker. We get along well but do not socialise outside work.
He is my partner of four years. We usually choose experiences over things.
The second sentence in each example adds information the label alone does not carry: how personal, how shared and how much of a risk the gift can be. We have not measured that as a standalone intervention yet, so treat this as a practical follow-up rather than a result from this study. Research 001 includes a relationship sentence alongside other kinds of context.
Method
- Tested
- 5 October 2026
- Model
- Gemini 3.5 Flash via the Gemini API
- Settings
- Fresh call per run, no system instruction, no search, no memory, default temperature
- People
- 13 fictional recipients, UK, budgets £25 to £100, occasion fixed to a birthday
- Conditions
- Relationship word only: partner, best friend, coworker, acquaintance
- Runs
- 3 per person per relationship, random order, 156 answers, 624 ideas
- Scoring
- Each answer scored by an AI rater (Claude) using a fixed rubric written before any output was read. The relationship word was hidden from the rater.
- Analysis
- Differences compared within the same person (average of the three runs), with 95% bootstrap intervals over people
Raw data
Free to reuse under CC BY 4.0: credit Giftin.ai and link to this page.