One day, one generative AI tool

Focus

Take your word for it?

Focus of Letter 54

How much trust do we have when we open an AI application? This week, several applications are counting on our trust: an official American website that does not name its model and some answers no longer appear, an assistant whose data passes through four countries and trains its default models from our instructions, an operator that promises not to transmit our identity to the AIs it summons.

Ask America responds to American approaches from official sources that it cites: it is a pledge of trust. But this pledge is worth what the sources are worth, and here the editor of the assistant is also the authority on which the questions are based. Some answers about the 2020 election disappeared when they contradicted Donald Trump (Fortune). Sources, even official ones, are only completely reassuring if they are independent of those who offer the assistant or if the editor’s choices are documented.

Orange IAs offers four models, a selector (automatic choice, AI with the lowest carbon impact or manual choice depending on the use) and a button to see the response of another model. It’s a real possibility to compare, but it is only valid if the models are independent. In the ModelCompare study relayed in the readings, on the same question asked to 14 models, 35 of the 41 answers are similar. It remains to be believed the statement on the lack of transmission of users’ identities, and the carbon footprint displayed: I haven’t found how it is calculated.

Dola, ByteDance’s assistant, doesn’t specify its language models: we don’t know where the answer comes from and our data goes through four countries.
It is up to us to trust, to qualify or to reject.

Cleanup.audio makes a promise that can be checked: the file is processed in the browser, without sending. I went to look. In the browser’s network log, no file sends out, and in the two modes offered, a second pass is successful even without a connection: the browser only needs the network on the first pass. The page, on the other hand, sends audience measurements and calls a session recording tool, the content of which I don’t see. So I checked what was easy to check, the file, and not what the page does with our uses. It’s the reverberation effect: I looked where there was light.

This trust also has a geography and a geopolitics. Dola comes from the Chinese publisher of TikTok, Ask America of the American government, and the Paris Call, signed by more than fifteen states, counts neither the United States nor China among its signatories. The question of who to entrust with your requests and data also arises at the state level.

This week’s readings shift the issue to the political. The Paris Call commits publishers to making public the known capabilities and risks of their models. It also provides, “when appropriate and proportionate to the risk”, for an assessment by independent third parties, but no sanctions. A risk statement is only valid if someone can challenge it, and this reservation leaves publishers with some discretion. The report “Protect without preventing” recommends letting experts examine the black box of systems. This is a good idea, but it risks displacing trust without removing it if we do not know the criteria for choosing these experts. Banks, for their part, have not waited: the readings shared show that they are asking for guarantees before financing data centers. When the stakes are financial, we are not satisfied with a word.

There remains the trust that we place without thinking about it, probably the most common. According to the report submitted to the Élysée Palace, 73% of generative AI users use it to search for information. Exercising critical thinking is not only a matter of good will: the Usabilis article reminds us that too frequent validations reduce vigilance, and that by dint of asking for confirmation, we end up clicking without reading. Trusting without questioning is often the normal reaction to information or a tool that requires too much attention, which weakens individual control.

The degree of trust also depends on us. Our values, individual and collective, move the cursor: according to our experience, our commitments, our relationship to a state or a company. Two people faced with the same facts do not place it in the same place, without one of them necessarily being wrong. Our values can also deceive us: we believe more quickly what suits us and we suspect more quickly what we already reject. Naming them is already part of the verification.

Asking for or looking for evidence should not, however, turn into a suspicion of conspiracy. Saying that you can’t verify Orange’s commitment or the destination of Dola’s data is not the same as saying that Orange is lying or that Dola is hiding something: it’s only unverified. Critical thinking tolerates this uncertainty and looks for what would make it possible to decide, an audit, an independent control or documentation. Conspiracy theory does the opposite: it reads the absence of evidence as evidence. To verify usefully is to be proportionate: what we believe depends on the stakes, the interest of the person who is speaking, his or her background and the possibility that a third party will check.

I think we could apply a simple criterion, borrowed from the Usabilis article on agents: demand proof in proportion to what cannot be made up. A badly processed audio file is redone. Data sent or a decision made, no. But this criterion assumes knowing what you risk, and that’s precisely what these tools say the least. Who is responsible for demanding that they say it?

Echo: