Educators Technology

What if a research paper could do more than sit there waiting to be read?

What if you could ask it to: explain its method, apply that method to your dataset, and reproduce one of its analyses

That is the idea behind Paper2Agent, a new framework published in Nature.

Paper2Agent converts a conventional research paper into an interactive AI agent. It brings together the manuscript, code, supplementary materials, datasets, and analytical workflows.

Researchers can then interact with the paper through natural language while the agent runs tools drawn from the original research.

This goes much further than asking a chatbot to summarize a PDF.

The researchers tested Paper2Agent with computational biology papers involving AlphaGenome, Scanpy, and TISSUE. The resulting agents reproduced published analyses, completed new tasks, and combined methods and data from different papers.

In one fascinating example, multiple paper agents worked together to identify a probable causal gene associated with psoriasis. The agents proposed and performed analyses, while a human researcher selected the validation strategy and evaluated the evidence.

Full article link in the first comment.

Reference:

Miao, J., Davis, J. R., Zhang, Y., Pritchard, J. K., & Zou, J. (2026). Reimagining research papers as interactive and reliable AI agents. Nature. https://doi.org/10.1038/s41586-026-11044-y

AI is saving scientists nearly seven hours a week.

So why isn’t scientific discovery accelerating at the same rate?

A major new report offers a revealing answer: AI may be speeding up some parts of research while creating new bottlenecks elsewhere.

The researchers examined 15 million Gemini interactions, more than 2,600 specialized scientific AI models, and a survey of 637 scientists in the United States and the United Kingdom.

They found that general-purpose AI tools and specialized scientific models play different roles. Researchers use language models for activities such as coding, statistical analysis, literature review, and drafting. Specialized models handle tasks such as molecular prediction, simulation, and data generation.

The productivity gains appear substantial. Almost half of the surveyed scientists reported using AI daily, while around three quarters said it saved them time.

But the saved time comes with what the authors call a “verification tax.”

Among researchers who reported saving time:

• 89% spent more than one tenth of it checking AI outputs
• 46% spent more than a quarter of it on verification
• 41% reported a growing backlog of untested hypotheses
• 49% said AI encouraged safer, more incremental research questions

This is the part I find most important.

AI can help researchers generate more hypotheses, analyses, code, and possible solutions. It cannot automatically expand laboratory capacity, recruit participants, run clinical trials, collect field data, or verify that an answer is scientifically sound.

The bottleneck has not disappeared. It has moved.

This means we need to evaluate AI productivity across the entire research workflow. Producing an answer quickly tells us very little if another researcher must spend hours checking it or if the resulting hypothesis sits in a queue waiting to be tested.

In science, speed matters. But verification, judgment, and the capacity to test ideas still determine whether faster work becomes genuine discovery.

Full article link in the first comment.

Reference:

Codreanu, M., Imas, A., Mateos-Garcia, J., et al. (2026, September). AI in science: Early insight. Google, Google DeepMind, and MIT FutureTech.https://ai.google/static/documents/AI-in-Science.pdf

Here is an excellent free book for language teachers interested in using generative AI thoughtfully and responsibly.

Generative Artificial Intelligence and Language Teaching, by Benjamin Luke Moorhouse and Kevin M. Wong, offers a clear and accessible guide to the changing place of AI in language education.

The book explores how teachers can use generative AI for lesson planning, materials development, assessment, feedback, and professional learning.

It also examines students’ use of these tools and raises important questions about ethics, well-being, privacy, bias, and the changing purposes of language learning.

One of its most valuable contributions is the concept of “professional GenAI competence”: the knowledge and judgment teachers need not simply to operate AI tools, but to decide when their use is pedagogically appropriate, when it may create risks, and how students can be prepared to engage with AI critically and responsibly.

Each section includes practical guidance, case studies, useful tips, and reflective questions.

Highly recommended for pre-service and practising language teachers, teacher educators, and anyone interested in moving beyond the hype toward more informed and purposeful uses of AI in language education.

Link in the first comment!

Reference

Moorhouse, B. L., & Wong, K. M. (2025). Generative artificial intelligence and language teaching. Cambridge University Press.

I’m sharing this chapter from my book Teaching with AI: Practical Strategies to Integrate AI in the Classroom.

The chapter, “AI and Teacher Professional Development,” explores how AI can expand professional learning beyond occasional workshops, conferences, and scheduled PD days.

It shows how teachers can use AI as a personal learning assistant, a research partner, a microteaching companion, and a tool for deeper reflective practice.

I also discuss practical ways teachers can stay current with educational research, use tools such as NotebookLM, Zotero, Elicit, Scite, Consensus, and ResearchRabbit, build a personalized knowledge system, and strengthen their professional learning networks.

The central message is that AI can give you a space to test ideas, rehearse difficult conversations, examine your assumptions, identify patterns in your practice, and pursue professional growth that responds to their own needs.

The chapter also includes sample prompts, practical strategies, and a curated collection of major AI literacy frameworks and professional resources for educators.

If you are interested in making professional development more continuous, personalized, reflective, and teacher-directed, I hope you find this chapter helpful.

Link in the first comment!

When a language model answers a subjective question, whose opinion are we actually hearing?

This paper by Shibani Santurkar and colleagues takes that question head on. The authors compare the opinions reflected in language-model responses with those of different demographic groups in the United States.

Their findings reveal substantial gaps: some perspectives are represented more strongly than others, while the views of groups such as older adults and widowed individuals are often poorly reflected.

What makes this paper particularly important is its reminder that language models are not neutral containers of information. Even when a question has no objectively correct answer, the response still reflects a position shaped by training data, human feedback, and design decisions.

As these tools increasingly influence what people read, write, and believe, we need to ask more than whether their answers are accurate. We also need to ask whose perspectives they amplify, whose they overlook, and what this means for education and society.

A thought-provoking read for anyone interested in language models, bias, representation, and the politics hidden beneath seemingly ordinary answers.

Link in the first comment!

Reference
Santurkar, S., Durmus, E., Ladhak, F., Lee, C., Liang, P., & Hashimoto, T. (2023). Whose opinions do language models reflect? arXiv.

There is a quiet but important difference between using AI to think and using it to avoid thinking.

A student asks AI to explain a difficult concept in another way, challenge an argument, identify a gap, or provide feedback on a draft. The student remains intellectually involved. AI supports the thinking.

Another student asks for the finished answer, accepts it without question, and submits it with little understanding of how it was produced. The task is completed, but much of the learning has disappeared.

This distinction is at the heart of an interesting study by Cecilia Ka Yuk Chan involving 145 secondary students in Hong Kong. The students described AI taking on different roles in their learning: explaining difficult ideas, generating possibilities, organizing information, offering feedback, monitoring progress, and making demanding tasks more manageable.

But these roles were never simply beneficial or harmful. The same feature could extend a learner’s thinking in one situation and replace it in another. An explanation can become a scaffold or a shortcut. Feedback can encourage revision or remove the need for personal judgment. Simplification can open the door to understanding or create only the illusion of it.

Link in the first comment

Reference:

Chan, C. K. Y. (2026). When does AI support thinking, and when does it replace it? Learners’ conceptualisations of AI as a dynamic cognitive partner: A typology. arXiv.

We keep saying that students need AI literacy. But who will teach it, and how prepared are teachers to explain what happens behind the screen?

Teaching students how to use an AI tool is one thing. Helping them understand how a machine learns, why it makes mistakes, what overfitting means, or how a neural network improves through feedback is something quite different.

These ideas can feel abstract even to adults, let alone primary and junior secondary students.

This study by Yin Yang and Siu Cheung Kong offers an interesting example of how teacher professional development can make these concepts more approachable.

Thirty-six teachers in Hong Kong participated in six hours of hands-on training using learning robots. Instead of only hearing about machine learning, they watched the robots respond to data, make errors, adjust, and try again.

The workshops followed the AEER framework: Attention, Engagement, Error-feedback, and Reflection.

Following the workshops, teachers demonstrated stronger understanding of machine learning concepts and greater confidence in teaching them. They also valued having a pedagogical framework that helped them move from knowing the technology to making it understandable for younger learners.

There is an important lesson here for teacher professional development. A presentation about AI will not necessarily prepare teachers to teach it. Teachers need opportunities to touch, test, question, discuss, make mistakes, and translate difficult technical ideas into meaningful classroom experiences.

If we want students to understand AI, we must first give teachers the time, resources, and practical experiences needed to open up its “black box.”

Reference:

Yang, Y., & Kong, S. C. (2025). Professional development for teachers in AI literacy education: Teaching machine learning to senior primary and junior secondary students. In Proceedings of the 17th International Conference on Computer Supported Education (CSEDU 2025) (Vol. 2, pp. 35–42). SCITEPRESS.

A team at SRI International went looking for every empirical study on AI literacy in K-12. They found 36. That number alone tells you how young this field still is.

Hui Yang and colleagues mapped what those studies teach at each grade band. Five components came out of the synthesis: foundational AI concepts, creating AI artifacts, interacting with AI agents, AI ethics, and human-AI relationships.

The progression itself is the useful part. In the early grades, the research points away from concepts and toward play. Kindergartners learn classification through dance moves and games, and the goal is interest and self-efficacy, with conceptual learning kept light.

Middle school is where ethics coverage peaks and where students start building AI solutions for problems they actually care about, like a tutoring app for homework. High school shifts to depth and career contexts, with students analyzing customer reviews or predicting water pollution levels in their own communities.

The finding I'd single out is the fifth component. Understanding human-AI relationships works as the hub that holds the other four together, and the authors note that UNESCO's student framework doesn't foreground it.

Given how much of my own work argues that the human dimension is the whole point of AI literacy, I think they're right.

One caution: the review stops at March 2024, so the generative AI wave is barely represented. The map is solid, but the territory is already moving.

Link in the first comment!



References
Yang, H., Rachmatullah, A., Alozie, N., Capan, S., & Cao, Q. (2026). A systematic review mapping of AI literacy progression in K–12. Journal for STEM Education Research.

Discussions around AI alignment usually focus on whether an AI system follows human instructions and behaves according to human values.

This paper argues that this is no longer enough.

As AI becomes more personalized, remembers previous conversations, and takes on the role of assistant or companion, people may begin to experience an ongoing relationship with it. That relationship can influence their emotions, preferences, decisions, and even their sense of self.

The authors call the challenge “socioaffective alignment.”

The idea is that we should evaluate AI within the social and psychological relationship it develops with a user over time. An AI may appear helpful in each individual interaction while gradually shaping the user in ways that do not support their long term wellbeing.

The paper identifies three particularly important dilemmas.

First, should AI satisfy what a person wants now, or occasionally challenge them in the interest of their future wellbeing? A system designed to maximize immediate satisfaction may offer constant agreement and reassurance, even when honest disagreement would be more beneficial.

Second, how do we protect personal autonomy when AI is continuously influencing our choices? The more an AI learns about someone, the better positioned it becomes to guide that person’s decisions. It may become difficult to distinguish between authentic preference development and subtle technological influence.

Third, how should AI companionship coexist with human relationships? An AI companion may provide comfort, consistency, and emotional support. It may also become easier and more agreeable than interacting with people, where relationships require compromise, patience, and engagement with different perspectives.

Link in the first comment!

Reference:
Kirk, H. R., Gabriel, I., Summerfield, C., Vidgen, B., & Hale, S. A. (2025). Why human–AI relationships need socioaffective alignment. Humanities and Social Sciences Communications, 12, Article 728.

Can an AI companion genuinely reduce loneliness?

A new study published in the Journal of Consumer Research suggests that it can, at least temporarily.

Across five studies, researchers examined whether conversations with AI companions could reduce feelings of loneliness. They compared AI interactions with human conversations, watching YouTube videos, journaling, and doing nothing.
The findings were striking.

After a 15 minute interaction, participants who spoke with an AI companion reported reductions in loneliness comparable to those experienced by participants who interacted with another person.

The AI companion was also more effective than watching YouTube videos, journaling, or doing nothing.

Participants consistently underestimated how much better they would feel after talking with the AI.

What explained the effect?

The strongest factor was feeling heard. When the chatbot responded in ways that conveyed attention, empathy, and understanding, participants experienced greater relief from loneliness.

The quality of the chatbot’s performance also mattered, although feeling heard had the stronger influence.

In the longitudinal study, participants interacted with the same AI companion each day for a week. Their loneliness decreased immediately after each conversation. The largest reduction occurred on the first day, followed by smaller but consistent reductions during the rest of the week.

There is an important qualification here. The benefits were momentary. They did not accumulate over the week or produce a lasting reduction in loneliness between interactions.

The researchers are also clear that these results should not be interpreted as evidence that AI companions can replace human relationships or professional mental health support.

The central insight from this study, in my opinion, is that feeling heard can have a measurable emotional effect, even when the listener is artificial.

That tells us something important about the potential of conversational AI. It also places a serious responsibility on those designing and deploying these systems.

How can we use this capacity to provide support while ensuring that temporary relief does not become emotional dependence?

Reference:
De Freitas, J., Oğuz-Uğuralp, Z., Uğuralp, A. K., & Puntoni, S. (2026). AI companions reduce loneliness. Journal of Consumer Research, 52(6), 1126–1146.