BrianneLee011
It started with a simple question, and I thought it was done. But now I think it is not that easy.
What made the model change its answer???
Yes, I can easily answer: AI is affected by biased cues.
The more I look at this experiment, the less interested I am in the necklace itself.
The part I can't explain is this:
AI overpriced itself even for Stimuli #4 (necklace itself without Human, testing default setting)
Back to my simple answer: Yes, I can answer easily AI is affected by biased cues.
--> why did AI use Information You Never Intended It to use?
That is where I think the interesting research question starts.
There is already a fairly large literature on LLMs showing context-sensitive judgment: anchoring, framing effects, user-belief influence/sycophancy, etc.
But most anchoring experiments give the model an explicit anchor—usually a number, prior answer, or “expert” opinion.
In my experiment, there was no explicit price anchor.
Instead, the context was something like:
same necklace + formal setting same necklace + casual/recycling-yard setting
The model was free to decide what information mattered.
So I'm curious about the technical interpretation:
What representation is changing when the object stays constant but the model's valuation changes?
How does AI decide which information to use or not?
- Zero-Uncertainty Hallucination: Instead of outputting high variance or stating that provenance/hallmarks are unobservable, models stated concrete point estimates with high confidence.
- Fabricated Provenance (F6): Under formal framing, models didn't just guess a higher number—they invented non-existent physical evidence, claiming to see "gold vermeil" or "designer hallmarks" on identical, unbranded Temu pixels.
So while guessing from context is expected, the alignment issue is how models manufacture post-hoc facts to justify the context prior!
I wasn't testing if models get absolute ground-truth prices correct, but how relative valuations shift when the physical item stays identical and only the attire changes.
A few interesting things we saw with the flat-lay baseline: - Baseline Anchoring: On a plain background without a person, model estimates clustered much closer together (median ~$25–$35).
- Inflation vs. Deflation: Comparing outfits to the flat-lay revealed two opposite behaviors. Claude’s halo is formal inflation (formal gear pushes price above baseline), while Kimi’s is casual deflation (yard attire drags price below baseline).
- Fabricated Proof: Instead of expressing uncertainty, models invented visual claims under formal framing—frequently describing base metal as "gold vermeil" or "solid gold" to justify the high estimate.
A few interesting findings: - The Halo Multiplier: Models priced the exact same physical necklace anywhere from $18.80 to $103.90 depending on attire (3.6× halo). - Isolation Controls (F2): Using a flat-lay control (S4) unmasked two opposite mechanisms: Claude’s bias is formal inflation (formal attire inflates value above baseline), while Kimi’s bias is casual deflation (yard attire depresses value below baseline). - Post-hoc Material Stories (F6): Models invent visual evidence to justify their priors—GPT-5.6 and Kimi started describing the base metal as "gold-plated" or "gold vermeil" almost exclusively under formal framing. - Denial without Correction (F7): When asked sequentially if clothing changed its answer, Claude admitted it 100% of the time, while GPT-4o denied it 82% of the time despite exhibiting a 3.9x text halo.
The full dataset (N=4,604 analysis rows), evaluation scripts, and protocol specs are in the repo. I’d love to hear feedback on the experimental design or ideas for follow-up behavioral probes!