I read a compliance disclosure last month that had gone through six weeks of legal review. Every word placed. It was elegant.
It also would not do the thing the writers thought it would do.
The pride in the room
Head of digital advisory at a Zurich private bank, mid-August, walking me through the copy for a new AI assistant pilot. She was proud of it. She had earned the pride.
The disclosure told the user, in plain language, that the assistant was the bank's own AI, and that its recommendations reflected the bank's product shelf. The team called the tone honest by design.
I asked her what the disclosure did inside the user's head. She paused. She said something close to: it lets people decide with their eyes open.
I said probably not.
She looked at me the way people look when a nice piece of work has been questioned. Fair enough. I owed her the paper.
What Bayreuth measured
Greubel, Guhling and Herweg ran an online experiment with nearly 3,700 participants. Two index-fund savings plans. One dominated the other on fees. Participants met a human adviser, a neutral AI chatbot, a bank-labelled AI chatbot, or no adviser at all. In one arm the AI was instructed to promote the worse fund.
The numbers do not read like a nudge.
An AI chatbot increases optimal choices by over 20 percentage points when promoting the optimal fund but reduces them by almost 30 points when promoting the dominated fund.
Without any advice, about 64% of participants picked the better plan. With a neutral AI recommending the better plan, roughly 87%. When the AI pushed the dominated plan, correct choices dropped to about 34%.
Now the part that lands for the Zurich meeting. When the chatbot was labelled as the bank's own AI with a stake in the outcome, it was not measurably less influential. Telling users the AI has skin in the game did not blunt the AI.
The authors also read the transcripts. Their tentative reading: the AI wins because it argues better than the human adviser.
A compliance disclaimer is a comfort blanket for the room that wrote it. In the field it does very little. Before the next customer-facing rollout, I would ask for two things. A red-team on the recommendation. An audit that measures how much the AI moves user choices, with a good human adviser as the baseline.
The disclosure was elegant. The lever was still in the model.