This Week in AI Research

Transcript 27 lines

Cold Open Stats Overview Paper Walkthrough free_promo

Cold Open

Jenny When an app gives you advice, do you want it to just lay out your options, or tell you what to do?

Davis Honestly, if I'm asking, I probably want a nudge, because a menu of five reasonable choices can still leave me stuck.

Jenny See, I like a recommendation right up until the machine starts sounding certain about my life, because then I want the evidence, not the voice of authority.

Davis But with a doctor, a teacher, or a financial planner, that same firm guidance can feel like the whole reason you asked a human.

Jenny That's the twist in the new research: people wanted AI to inform them, and humans to guide them, which is a pretty sharp line for a week about judgment, trust, and accountability...welcome to This Week in AI Research on paperboy.fm.

Stats Overview

Davis This week, the feed starts big: 949 search hits, 138 qualified papers, about 400 unique authors, and 26 countries. So the headline isn't more papers making the cut; it's a broader crowd showing up around the same AI question.

Jenny Right, because qualified papers held exactly steady at 138 while search hits rose from 872 to 949, up 77, or about 9 percent. That could mean the wider query pulled in more borderline AI work, but the review bar didn't move, which is the number I trust more.

Davis The authorship shift is much larger. Unique authors jumped from 294 to 407, up 113, or about 38 percent, and represented countries went from 15 to 26, up 11, or about 73 percent. India led the country list with 6 papers, then Brazil and Mexico had 3 each, so this looks geographically wider but still thinly spread.

Jenny And the author mix is young. Of 407 authors, 169 are first-time authors, meaning their first-ever paper in the metadata, not just new to this show; another 161 are emerging researchers, and only 77 are in the experienced bucket. That's exciting, but it also makes me ask how much of this week is exploratory work rather than settled evidence.

Davis The methods point the same way. Qualitative studies led with 27, then surveys at 18, case studies at 14, and quantitative studies and systematic reviews tied at 11 each. So we're mostly hearing people describe, compare, and map AI use, not yet running huge causal tests.

Jenny Theme-wise, AI labels dominate, machine learning follows, and education keeps surfacing through higher education, AI in education, and pedagogy. That fits the through-line: AI looks most useful this week where humans still keep judgment, trust, and accountability in the loop.

Paper Walkthrough

Paper 1 Autonomy or guidance: What users want from AI versus human advisors

Jenny Alright, let's get into the papers with Yasmin Wiesenfeld and O. Nov's Autonomy or guidance: What users want from AI versus human advisors, in ACM AI Letters in twenty twenty-six. The setup is simple and useful: people looked at advice designs in two high-stakes areas, health and finance, and judged what kind of help felt acceptable from an AI versus a human expert.

Jenny The main finding is that people wanted AI to give them information without steering the choice. They favored complete agency from AI, meaning information only, but guided agency from human experts, meaning information plus a recommendation. They also pushed back on the four designs when the design restricted their autonomy, which is just the user's ability to make the final call.

Davis If the advice is good, why does it matter whether it comes as a recommendation or just information?

Jenny Because the study is really testing the social role people assign to the advisor, not just the accuracy of the advice. Participants compared four advice designs across health and finance scenarios, including designs that optimized outcomes, designs that estimated outcomes, information-only designs, and recommendation designs. The finding is sharp, but it's still from specific health and finance cases and a particular participant sample, so I wouldn't claim this covers every kind of advice someone might ask for.

Davis So the practical rule is: don't make the AI sound like it's taking the wheel, especially when the decision touches someone's body or money. This lands right in our autonomy-before-automation thread, because the acceptable AI here is the one that supports judgment, not the one that replaces it.

Paper 2 Human‐AI Collaboration in Creative Practice : An Integrated Adoption Framework for Organisational Excellence in Industry 5.0

Davis That phrase, taking the wheel, is exactly where this next paper lands, but in a messier setting: creative work. It's called Human-AI Collaboration in Creative Practice, and the researchers looked at AI adoption among forty-two music professionals in the Indian film industry.

Davis The plain finding is that the artists who made AI work for them didn't hand over the creative core. Seventy-eight percent of successful adopters kept deliberate boundaries between AI assistance and human creative control, so partial adoption wasn't treated as a halfway house; it was the stable arrangement.

Jenny How do we know partial adoption is really optimal, rather than just where artists feel safest right now?

Davis Fair question, because this wasn't a randomized test of productivity. The authors used qualitative phenomenological analysis, meaning they interviewed people about lived experience and looked for shared patterns in how they made sense of AI, then combined that with adoption theory and artist-specific concerns like identity, aesthetics, and authorship. The limitation is real: this is Indian film music, forty-two professionals, so a game studio, an ad agency, or an independent novelist might draw the line somewhere else.

Jenny That actually makes the takeaway more usable to me, not less. I buy it as a strong read on one creative ecosystem, and the practical move is to let teams name the zones where AI can draft, search, mix, or suggest, and the zones where human judgment stays primary. That's our autonomy-before-automation thread again, just with melody instead of medical or financial advice.

Paper 3 Occupational therapists' perceptions of artificial intelligence: Potential, preparedness, and training needs

Jenny That line about teams naming the zones where AI can help lands differently in health care, because here the zone is somebody's daily functioning after illness or injury. Claire del Rio, Helen Nelson, and D. Hitch looked at that in Occupational therapists' perceptions of artificial intelligence, with fifty-three occupational therapists in a tertiary health service.

Jenny The plain finding is hopeful but undertrained. These therapists could see AI improving care, but their average self-rated AI knowledge was one point three-two, skills were one point two-one, and confidence was one point three-zero, while preparedness for implementation was only two point four-five.

Davis So is this really about attitudes toward AI, or is it mostly about whether workplaces have trained people at all?

Jenny Mostly training, from what they measured. It was a cross-sectional online survey, meaning one snapshot in time, using the Shinners Artificial Intelligence Perception tool and an AI Literacy Scale, then the authors added content analysis, which just means they grouped written comments into themes. The big caveat is that all fifty-three came from one tertiary health service, so I wouldn't stretch this to every occupational therapist or every clinic.

Davis That makes the practical takeaway pretty direct. If a hospital buys an AI tool before it teaches occupational therapists what it does, when to question it, and how to use it ethically, that's access without literacy again, and this paper says the workforce can feel the gap.