Notes from Day Two of SIETAR Brasil 2026
Earlier today I wanted to ask a question in Michelle (Mica) Chahin's session, Quando a IA encontra culturas: entre contexto, adaptação e viés. I did not ask it.
My brain was full. At the last minute I was asked whether I could present in English instead of Portuguese. I had written my talk in English, translated it into Portuguese, and spent days learning it that way, so by the time the session started my head was somewhere between the two languages and a bit pifando. I wrote the question down instead.
A gente vê a IA abrindo portas para comunidades que antes não tinham acesso à informação. Mas o que acontece quando a resposta da máquina vira a única fonte de verdade dessas pessoas, sem espaço para checar ou contestar?
We see AI opening doors for communities that previously did not have access to information. But what happens when the machine's answer becomes the only source of truth for those people, with no room to check or contest it?
I did not ask, so I did what I do best. I researched.
I also use AI tools to help find the papers I read, and, honestly, to sort out the SEO. It is handy, which is part of why this question matters to me.
The part I believe
AI does open doors. By now I do not even need to explain how. We both know it does.
The part that worries me
The worry is that answers sound right whether or not they are.
I have seen this before AI. People around me were sharing that eating twelve oranges can help with diabetes. I am not a doctor, and I do not think whoever wrote that on Facebook was one either, but the people reading it believed it. It was on a screen, said with confidence by someone who sounded as if they knew.
Now picture the same thing as a machine that answers in fluent, confident language, for free, at any time, in your own language.
One small study tested language models on questions in English, Kazakh and Mongolian. In the two lower-resource languages accuracy fell by about 14 to 17 percentage points, while fluency held up much better, so the answers still sounded right. It is a preprint with 50 questions and says nothing about Portuguese, so I am not claiming more than that.
Errors happen in practice, too. New York City's business chatbot told users that landlords could refuse tenants because of their source of income, which The Markup reported as wrong. An immigration lawyer in New South Wales used ChatGPT to find cases for submissions in an immigration matter, and some of the cases did not exist. The judge referred him to the state legal services regulator and noted that he had not checked what he filed. If someone trained to check can file invented cases under time pressure, I would not like my chances at a government counter in a language I am still learning.
I went looking for proof that migrants use these tools as their only source. I did not find any. That part of my question is a risk to watch, not a finding.
What I actually think
A right answer and a wrong one can look the same on the screen, so what matters is whether you have somewhere to take it: a friend, a lawyer, a helpline, someone who says "let me look at that with you".
In October 2025 the Parliamentary Assembly of the Council of Europe adopted a resolution on AI and migration. It says AI should support human decision-making and never replace it, and it calls for transparency, accountability and human oversight. My own word was contest, because it names what the village used to be for. It was where you checked things.
I do not know who the people reading that orange post checked with.
Who do you check with when an answer sounds official and you are not sure? I would love to hear about it in the comments. 💛
Jessica Gabrielzyk
