This Week in AI Research

Transcript 27 lines

Cold Open Stats Overview Paper Walkthrough free_promo

Cold Open

Jenny When an app explains why it showed you something, do you actually feel more in control?

Davis A little, but an explanation label is not the steering wheel; it tells me why the road curved after I'm already on it.

Jenny I know, and I still like the label, because if a shopping app says it recommended boots because I clicked hiking gear, I feel less like I'm being haunted.

Davis Right, but feeling oriented isn't the same as changing the system, and the social feed version gets bleak when people can name the algorithm and still can't make it listen.

Jenny So this week is about control without control, and who actually gets influence once AI is everywhere...welcome to This Week in AI Research on paperboy.fm.

Stats Overview

Jenny This week we had 531 AI research hits, narrowed to 200 for semantic review, and 143 made the qualified set. That's up from 138 last week, so 5 more papers, or 3.6 percent. Small rise, but it matters because the pile got cleaner, not bigger.

Davis Right, the search pool actually fell hard, from 949 hits to 531, down 418, or 44 percent. So the headline is not more noise. It's fewer raw matches, with a slightly higher useful yield, probably because the week leaned into human-facing AI work like education, adoption, and generative AI.

Jenny And the authorship widened. Unique authors rose from 407 to 474, up 67 people, or 16.5 percent, while countries went from 26 to 29. The country metadata is thin at the top, though: U.S. shows 7, India 5, the U.K. 4, then China, Canada, Germany, and Vietnam at 3 each.

Davis The career mix is interesting too. Of 474 authors, 133 were first-time, meaning publishing their first-ever paper in the metadata, not just new to our feed. Emerging researchers were the largest group at 207, or 43.6 percent, and experienced authors were 134, basically the same size as first-timers.

Jenny Method-wise, this was not a benchmark week. We saw 36 qualitative studies, 26 surveys, and 15 case studies before you even get to reviews or mixed methods. Plain translation: researchers were mostly asking people what AI is doing to classrooms, workplaces, and decisions, not just scoring models on tests.

Davis That fits the episode's through-line. The top themes were artificial intelligence, generative AI, and education, with machine learning trailing behind. So the center of gravity is shifting from can we put AI in the system, to who understands it, who shapes it, and who gets left out.

Paper Walkthrough

Paper 1 From AI Access to AI Influence: Who Uses AI for News, Who Is Concerned About It, and What Are the Implications for the Multi-Level Digital Divide

Jenny Alright, let's get into the papers, and I'm starting with From AI Access to AI Influence, a twenty twenty-six Applied Informatics study by T. Laor that asks who in Israel is using AI for news, and who thinks AI is shaping their social and political views.

Jenny The surprising bit is that the AI news divide did not follow the usual income-and-education ladder. In an online survey of five hundred fifteen people in Israel, older people, women, and people with lower education reported somewhat higher AI news use in the simple comparisons, meaning the first-pass group-to-group checks before controlling for other factors.

Davis If those demographic effects are small, what would actually explain who trusts AI with the news?

Jenny The authors measured self-reported AI news practices, like using AI to consume news, summarize stories, and identify fake news, plus worries about bias and perceived influence. Then they ran multivariable regressions, which just means they tested age, gender, education, and income together, and age, gender, and education still mattered while income did not show a consistent independent effect. But this is exploratory self-report data, and the authors say demographics explain only a small portion of the variance.

Davis So the practical takeaway is pretty direct: newsrooms shouldn't assume the AI news gap is just about who has access or money. If AI is becoming part of everyday algorithmic autonomy, then they need to measure literacy, trust, usefulness, and skills directly, because those are probably where the real control gap lives.

Paper 2 Artificial Intelligence in Social Media: University Students’ Experiences Across AI-Driven Platforms

Davis That control gap is exactly where this next paper lands, but on social feeds instead of news apps. N. Tuncay's Artificial Intelligence in Social Media surveyed seven hundred fifty undergraduates about how AI shapes what they see, click, and believe online.

Davis The plain finding is pretty stark: students were deeply inside AI-shaped feeds, but that didn't mean they could steer them. The sample was ages eighteen to twenty-two, mostly students of Turkish cultural background living in Northern Cyprus, Türkiye, Germany, the United Kingdom, and fifteen other countries, and higher AI-driven engagement, personalization, and misinformation exposure sat alongside lower AI awareness and control.

Jenny How did they distinguish being aware of algorithms from actually having control over them?

Davis They used a survey, so students reported their own experiences with AI's role in engagement, information exposure, habits, misinformation, and sense of control. Awareness means knowing or noticing that algorithms shape the feed; control means feeling able to interpret, question, or influence those systems, and the authors say those didn't rise together. Seven hundred fifty is a strong sample for this kind of student survey, and the country spread helps, but it's still centered on students of Turkish cultural background, so I wouldn't call it a universal student result.

Jenny That makes the media-literacy answer feel too small. If students can know the feed is engineered and still can't change much, then everyday algorithmic autonomy is a design and governance problem, not just a classroom reminder to be smarter on the internet.

Paper 3 AI Transparency and User Behavior in Human–AI Collaboration: Evidence from E-Commerce Recommendation Systems

Jenny That gap between knowing the feed is engineered and feeling able to steer it is exactly where this next one lands. I. Oncioiu's AI Transparency and User Behavior in Human–AI Collaboration looks at shopping recommendations, where the algorithm isn't just showing content; it's nudging a purchase.

Jenny The plain finding is that transparency matters because it changes how people think, not just how much they see. In a survey of three hundred twelve users of e-commerce recommender systems, clearer AI cues were linked to more algorithmic understanding, stronger fairness perceptions, more perceived control, more trust, and then higher purchase intention. AI literacy also shaped how users interpreted the information the system gave them.

Davis But is transparency helping people make better choices, or just making the recommendation feel trustworthy enough to buy the thing?

Jenny The authors can't prove better choices here. They built a conceptual model and tested the links with partial least squares structural equation modeling, which is a way to estimate a chain of survey relationships when concepts like trust, fairness, and control aren't directly measurable. That's useful evidence from three hundred twelve real recommender-system users, but it's still e-commerce, so I wouldn't stretch it to every human-AI collaboration setting.

Davis That keeps the everyday algorithmic autonomy thread pretty concrete. A disclosure box that says 'AI recommended' is thin; the design win is showing why this item appeared, what data mattered, and how I can push back before trust turns into a one-click purchase.