Supriya Subudhi, product manager at Bumble, told AI Engineer that 78% wanted AI photo feedback but only 10 to 15% used it in a 50/50 test.
Before that test, Bumble had more than 20 million monthly active users who told research they could not judge if their dating profile was good, Subudhi said.
After shipping, top-line gains hid tail failures where the model told smiling photos to smile more, and Subudhi said Bumble added guardrails and human review of real outputs instead of trusting averages.
The best LLM cost about 10 times a simpler approach, Subudhi said, and unlike Bumble's matching algorithm where incremental cost is low, generative cost rises with every use.
Subudhi pointed to Meta's Muse AI staying US and Canada only for regulatory reasons, said Bumble paused the feature outside the US, and the constraint matters as Bumble now faces a reported $1.3 billion sale at a 138% premium according to FHM India.
The daily brief
Get the free daily brief with the day's AI news, new AI startups and funding rounds every morning.
Join 12,006 subscribers.
No attacker, affected system beyond dating profiles, or patch was demoed, and the question Subudhi never answered is whether long-term brand trust can justify a bill that grows with success?
