The book, titled “AI Through the Experts’ Eyes: Communicating Complex Ideas” and published by the University of Pittsburgh Press, examines the daily work of AI and machine learning researchers and the communication practices that shape public understanding of the field. Gallagher, who is affiliated with the university’s School of Information Sciences and is teaching two courses on artificial intelligence this semester, set out to help experts communicate about their work without hype, and to help nonexperts, particularly writing teachers and educators wrestling with the role of AI in their classrooms, think more critically about the technology. The project began as an effort to deepen his own understanding of machine learning and natural language processing techniques for his research, but when the COVID-19 pandemic derailed his plans to visit AI research centers in person, he pivoted to a different methodology: interviewing the researchers directly about their work and how they describe it.

That pivot produced an unusually rich dataset. Gallagher interviewed more than 100 AI and machine learning experts, most of them based in academia and drawn primarily from the disciplines of computer science and physics. What he heard from them stands in stark contrast to the picture of artificial intelligence that dominates headlines and popular culture. The public, he notes, may envision killer robots when they think about AI, but the reality is mostly mundane: the GPS software that guides a vehicle through traffic, the spam filter that quietly sorts an email inbox, the recommendation systems that shape what people see online. These are the actual products of the field, and creating them is not a matter of conjuring intelligence from silicon but of methodical, repetitive engineering work: designing a model, obtaining feedback on its performance, and revising it over and over again.

“It’s boring. It’s math, a lot of math, and meetings and day-to-day normalcy,” Gallagher said, describing what the researchers told him about their working lives. The gap between that reality and the public image is not accidental. Gallagher observed that the leaders of AI companies tend to talk about “fanciful answers and magic in the sky, not linear algebra,” a rhetorical choice that serves marketing purposes but leaves the public with a distorted sense of what the technology is and how it works. Linear algebra, the branch of mathematics concerned with vectors and matrices, underlies nearly every modern machine learning system, from the neural networks that classify images to the large language models that generate text. Understanding that foundation, Gallagher argues, is the first step toward demystifying the field.

One of the most technically substantive concepts the book explores is what researchers call the alignment problem, and Gallagher is candid about why it concerns him more than the cinematic scenarios of Terminator or HAL 9000. “I’m less concerned with Terminator and HAL 9000 and more concerned with the systematic error that didn’t get picked up as a bug in the code, called the alignment problem,” he said. “The program still works but it’s not exactly what you wanted it to do, and it’s now doing something kind of bad.” In other words, the danger is not a malevolent machine but a well-functioning one that has been optimized toward the wrong objective, producing systematic errors that no single line of faulty code reveals.

To make the alignment problem concrete, Gallagher turns in the book to a real example: a garbage-detection program called TrashCan, designed to help clean up the ocean garbage patches that accumulate plastic waste across vast stretches of open water. The challenge of such a system illustrates how difficult it is to teach a machine what counts as trash. If the model is trained on objects that are not meant to be in the water, it can learn to identify and collect garbage. But edge cases multiply quickly. “Maybe you have a buoy that’s supposed to be there and is not trash, until it sinks onto the ocean bottom and it is trash,” Gallagher explained. “That’s an alignment problem if it collects buoys that are supposed to be there.” Training data introduces further complications: if a model is trained on sea life, what happens when it encounters land animals that have ended up in the water? It might classify them as trash and collect them, with potentially harmful consequences. “The problem is not an evil robot,” he said. “It’s bad programming leading to an unintended result.”

If the technology itself is less dramatic than the headlines suggest, why does the hype persist? Gallagher locates much of the answer in the hypercompetitive atmosphere of a field that is changing at breakneck speed. Researchers in both academia and industry are expected to publish frequently, and the dominant venue for that publication is the conference rather than the peer-reviewed journal. Conferences move far faster than journals, which can take many months or even years to release a paper, and that speed creates pressure. Under such conditions, Gallagher argues in the book, researchers may be pushed to overstate their findings or understate their limitations, inflating results to stand out in an environment where thousands of papers compete for attention.

The pressure does not end with publication, and this is where the book’s analysis of scientific communication becomes especially pointed. Nearly everyone Gallagher interviewed mentioned the expectation of building a public relations campaign around each paper. “Now you must advertise the paper on social media. It also needs a blog post, a GitHub repository of the code, shared results on LinkedIn, Bluesky, even YouTube videos,” he said. “You can be a researcher and you also have to be a content creator and influencer, producing a conference paper and also six other content genres.” The consequence, as Gallagher wrote in the book, is that “the current AI publication landscape may lower publication quality while possibly sensationalizing scientific results.” A system that rewards volume, visibility, and self-promotion over careful, hedged, peer-scrutinized claims is a system structurally inclined to produce hype, regardless of the intentions of individual researchers.

Against that structural drift, the book offers a set of concrete recommendations for countering the hype cycle. First, Gallagher said, researchers should stress that AI is a type of automation designed to serve a specific need, whether that means detecting trash in the ocean or making sense of vast amounts of scientific data, rather than a general-purpose intelligence hovering above ordinary engineering. Second, AI and machine learning researchers should be trained and encouraged to translate their research for many different audiences, including nonexperts, and to discuss their work in the media, including explaining the technical, granular aspects of their field on long-form platforms such as YouTube, where nuance has room to breathe in a way that a social media post does not. Third, Gallagher calls for changes to incentives in academic publishing, particularly for conference proceedings, so that venues are no longer pressured to produce a high quantity of articles within a short time frame at the expense of quality and accuracy.

Finally, the book has practical advice for nonexperts who want to understand artificial intelligence more accurately: seek out technical AI researchers rather than business leaders promoting their products or podcasters seeking an audience. Many researchers, Gallagher noted, maintain YouTube channels devoted to explaining the technical details and core concepts of the field, and engaging with those primary sources can demystify the technology in a way that secondhand commentary cannot. The broader lesson of the book is that the story of artificial intelligence is not a story of magic or menace but a story of people: mathematicians and engineers iterating on models, wrestling with alignment problems, and navigating a publication culture that often rewards exaggeration. Listening to those people, Gallagher argues, is the most reliable way for the public, and especially for educators shaping the next generation, to develop a clear-eyed, critical understanding of what AI can do, what it cannot, and why the gap between the two matters so much.

Subject of Research: How artificial intelligence researchers communicate their work and the hype surrounding AI

Article Title: Book counters AI hype by examining how researchers talk about their work

Article References: Book counters AI hype by examining how researchers talk about their work. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, machine learning, AI hype, science communication, alignment problem, academic publishing, John Gallagher, University of Illinois, natural language processing, research interviews, TrashCan, AI education

Glawatzki’s core theoretical move is to connect two concepts that researchers have usually treated separately. Anthropomorphism, the attribution of human characteristics to non-human entities, has mostly been studied as a psychological phenomenon occurring in individual users who chat with bots and start to feel a social presence. Sociotechnical imaginaries, a term from science and technology studies, describe collectively held visions of desirable futures made possible by science and technology. The study proposes a synthesis of the two: “anthropomorphising imaginaries,” futures that are structured from the ground up by human-like descriptions of machines. In these imaginaries, generative AI is not merely useful; it is framed as performing at a human level and deserving human-like trust, which positions it as powerful in the societies being imagined.

To reconstruct these imaginaries, the researcher conducted a qualitative content analysis of 35 corporate documents published between late 2024 and spring 2025 by three companies: OpenAI, Meta, and the French startup Mistral AI. The sample included product announcements, research updates, company pages, case studies, and user documentation, selected because they either explained the abilities and features of a model or discussed its contexts of use. The choice of companies was deliberate: OpenAI with roughly 800 million weekly active users, Meta with its openly released Llama models, and Mistral AI, often described as Europe’s most promising competitor to U.S. systems, together represent a diverse but culturally proximate slice of Western AI development. Using a structured qualitative approach, the analysis coded the documents for attributed abilities, characteristics, roles, and the distribution of responsibility and power between humans and machines.

The results show that companies anthropomorphize generative AI along four dimensions: cognition, communication, sociality, and personality. Cognitive attributions are the most prominent. Terms like “chain of thought” frame algorithmic processes with human cognitive metaphors, and models are credited with knowledge, systematic understanding, and reasoning that produce “accurate answers.” Some rhetoric even naturalizes these abilities: Mistral AI claims that large language models, “by nature,” are not good at calculations, a phrasing that positions the technology as an independently evolving, almost organic entity rather than a human-built product with material costs and consequences. Communication skills are anthropomorphized too, with Meta describing Llama as able to switch casually between humor, empathy, intellectualism, creativity, and problem-solving. At the extreme, Meta calls the model “an expert conversationalist” that is “companionable and confident,” granting it a narrated personality that transcends its status as software.

From this material, three distinct imaginaries emerge, each present across all three companies. The first imagines generative AI as a tool with human capabilities. Here, human cognitive abilities—thinking, learning, understanding—are portrayed as technically encodable inside a seemingly neutral, human-controllable product that will boost productivity and democratize skills. Intelligence becomes a commodity sold by private companies, and the vision carries a distinctly neoliberal flavor: humans augmented by AI will remain competitive and productive, while non-use is framed as a competitive liability. The study warns that this framing raises awkward questions about whose intelligence gets encoded into the tool, potentially importing racist and ableist conceptions of intelligence, while concealing the often exploitative human labor of data collection and verification behind the machine.

The second imaginary casts generative AI as an empowering helper, a cooperative assistant integrated into every corner of life. Companies oscillate between attributing competence to the technology and reassuring users that humans remain in control. In this vision, AI manages daily tasks, offers professional support, and even provides emotional care: Meta describes Llama as recognizing whether a user is looking for chit-chat, emotional support, humor, or venting. The study notes that this rhetoric subtly devalues the human, implying that people are exhausted, limited, and in need of technological enhancement—a premise the researcher identifies as transhumanist in spirit. Strikingly, the imaginary also recasts users as teachers: OpenAI compares a GPT model to “a junior coworker” needing explicit instructions, reframing the unpaid labor of training and prompting as pedagogical care that ultimately optimizes corporate products.

The third and most far-reaching imaginary envisions generative AI as an anthropomorphized superintelligence that will reach and surpass human-level intelligence through self-directed evolution. Meta’s claim that one powerful model was the perfect choice to “teach” smaller models constructs a future in which AI progress unfolds independently of humans. Models are framed as experts and “thought partners,” with OpenAI comparing reasoning models to “a senior co-worker” whom you can trust to work out the details. The study argues that this attribution of expertise to a system that generates text through probability calculations devalues situated human knowledge and normalizes the opacity of AI systems. Yet the same imaginary flips into threat mode: OpenAI’s preparedness framework anticipates self-improvement and autonomous replication capabilities, evoking a Frankenstein narrative that conveniently positions the companies themselves as humanity’s protectors.

Each imaginary generates a corresponding construction of AI power. Instrumental power corresponds to the intelligent tool: high performance, but under human control. Cooperative power corresponds to the helper: a shared power emerging from human–AI collaboration, cultivated through both functional and affective trust. Autonomous power corresponds to the superintelligence: power located mainly within the machine itself, justified by narratives of inevitable technical evolution. The three types build on each other, progressively shifting power away from humans, and each promotes distinct forms of trust—functional reliability in the first case, emotional benevolence in the second, and epistemic confidence in the third.

The normative implications are sobering. According to the study, these constructions foster uncritical acceptance of AI, frame its use as inherently desirable while foreclosing discussion of alternatives, depoliticize the technology by hiding the exploitative labor conditions of its creation, normalize post- and transhumanist ideologies, and legitimize AI companies as authoritative shapers of the future. This matters because companies possess what the paper calls infrastructural power: their research ideas spread more contagiously than academic ones, their innovations routinely precede regulation, and in an AI race driven by speed and resources, futures get realized not through democratic negotiation but through corporate momentum. Sociotechnical futures, as prior research quoted in the study puts it, become commodified by hegemonic technology companies and their expanding infrastructures.

The study is candid about its limits: the analysis covers only Western companies, and firms from other cultural contexts, such as China’s DeepSeek, may articulate different imaginaries that increasingly influence Western discourse. But its central warning travels well beyond the sample. Anthropomorphism, the paper concludes, is not merely an exaggerated or misleading metaphor to be corrected by careful speakers. Embedded in sociotechnical imaginaries and deployed strategically, it becomes an instrument of hegemonic power that shapes how AI is researched, developed, appropriated, and regulated. The words “it thinks” may seem trivial, but whoever controls those words is, in a very real sense, writing the future.

Subject of Research: Anthropomorphising sociotechnical imaginaries of generative AI constructed by AI companies

Article Title: Constructing AI power: anthropomorphising imaginaries of generative AI by AI companies

Article References: Glawatzki, A. (2026). Constructing AI power: anthropomorphising imaginaries of generative AI by AI companies. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03367-1

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03367-1

Keywords: anthropomorphism, generative AI, sociotechnical imaginaries, AI power, OpenAI, Meta, Mistral AI, transhumanism, AI hype, large language models, critical AI studies, corporate power