Society & Ethics

Tsing Hua president warns students not to treat AI as infallible

NTHU President John Kao urges universities to teach students to interrogate AI output, citing biased data, data center power and water use, and 10 interdisciplinary research themes.

'AI is not God': NTHU president warns against blind faith in technology
'AI is not God': NTHU president warns against blind faith in technologyjarmoluk / pixabay
By Rebecca Stone6 min read

Why it matters

  • NTHU has identified 10 interdisciplinary research themes, including AI, to encourage collaboration across departments.
  • Kao said data centers consume enormous amounts of electricity and water.
  • NTHU forms peer study groups mixing humanities, political science and software engineering students, plus short courses for the general public.
  • Kao on AI limits: "AI is not God. AI cannot do everything. It makes mistakes."
  • Kao on rankings: "Ranking is important, but it's not everything."

"AI is not God. AI cannot do everything. It makes mistakes," National Tsing Hua University President John Kao said in an interview with The Korea Times, urging students to assess AI-generated answers instead of accepting them at face value.

The interview covered training data, data center resource use, technological sovereignty and research strategy in a single conversation. It lands as universities race to build AI programs and as National Tsing Hua University (NTHU), a research university in Taiwan, presses for international recognition.

Kao framed the caution as a teaching requirement rather than a technical footnote. Rapid progress in AI raises the value of students' analytical skills over their ability to operate new tools, he said. AI now reaches past software development and semiconductors into ethics, privacy, education, national security and environmental sustainability.

Why do AI answers come back incomplete?

The weakness sits in the data. "If the database is biased or if the database is incomplete, the answer that AI will provide or the analysis will be incomplete as well," he said.

He applied the point to the chatbots students already use. "When we ask ChatGPT a question, the answer that we get is most likely incomplete," he said, warning against blindly trusting its responses.

His standard is probabilistic, and that is what makes it a teaching rule rather than a complaint. A partial answer presented without qualification looks like a complete one, so the reader, not the model, carries the burden of checking.

The stakes run through classrooms and labs alike. A university that teaches AI as an authority trains students to inherit the gaps in its training set without checking them, which is the specific failure Kao named.

What does AI cost in electricity and water?

Kao also raised the physical footprint of the systems. Data centers consume enormous amounts of electricity and water, he said, tying adoption to environmental sustainability rather than treating computation as free.

He extended the concern to technological sovereignty. Countries, he said, are increasingly wary of relying on AI systems trained largely on English- and Chinese-language data, and those systems can produce biased results when they are asked to serve other populations.

The mechanism he described for a single query operates at national scale, which makes his data-center point and his language point one argument. Skewed or thin coverage produces skewed or thin answers, and the cost of correcting that shows up in electricity, water and validation work.

For a research university, both concerns settle on the same ledger. Compute determines what an experiment costs and who can afford to run it; language coverage determines whose questions the models handle well.

Can one college handle AI by itself?

Kao said no. "AI is complicated. It's hard to just put it together within one department or even one college," he said.

His prescription was viewpoint diversity. "To fully understand the impact of AI on teaching, learning, research and basically our daily lives, you really need to look at AI from these different points of view," he said.

He named the domains:

  • Ethics
  • Privacy
  • Education
  • National security
  • Environmental sustainability

NTHU answers by organizing researchers around problems rather than around departments. The university draws experts from engineering, natural sciences, humanities, social sciences and education into AI work instead of confining it to a single discipline.

It has also identified 10 interdisciplinary research themes, including AI, to encourage collaboration across departments and strengthen the broader impact of its research.

The count is the administrative version of Kao's argument. Ten themes places AI beside other cross-cutting problems under a shared framework, so a data-center energy study can draw on engineering and environmental work at the same time as an algorithm draws on computer science and ethics.

How does NTHU teach students to question AI?

The same structure reaches the classroom. Beyond course instruction, NTHU assembles peer study groups that bring together students from the humanities, political science and software engineering to examine AI from different perspectives.

The university also runs short courses for the general public. Kao treated both audiences as needing the same skill, the capacity to question an AI-generated answer rather than take it at face value.

His reasoning connects classroom method to research posture. A graduate trained to interrogate an AI answer carries a different tool set into labs and industry than one trained to prompt a model and trust the output.

Peer groups also invert a common hierarchy in technical fields. Software engineering students sit beside humanities and political science students, so the questions asked of a model include consequences, not only accuracy, and the answer gets pressure-tested before it reaches a lab notebook.

Do rankings come before research?

Kao answered no. "Ranking is important, but it's not everything," he said, urging universities to focus on research that tackles complex global challenges and makes a meaningful impact on society.

He tied the two together in one line. "Multidisciplinary collaboration increases the impact of university research, and that drives the rankings," he said, arguing that stronger rankings should follow meaningful work rather than serve as the goal itself.

Rankings, he acknowledged, help universities raise their international profile and attract top students and faculty. Research impact, not position alone, is a key factor in global university assessments, and that is the variable a university can move with its own choices.

The order matters for how hiring and funding decisions get made. An institution that treats a ranking gain as the objective optimizes for the metric itself; one that treats rankings as an outcome of impact keeps the metric downstream of the work.

Kao's account of recognition runs in one direction. Collaboration produces impact, impact produces international recognition, and recognition produces rankings, which in turn widen the pool of applicants and faculty a university can attract.

What changes if universities adopt the standard?

Nothing about the tools. Kao's proposal concerns how AI is taught and how research is organized, not which models to license or how fast to deploy them. The measurable changes sit in course design, peer group composition and the share of research themes that cross departmental lines.

The test he set for students is simple enough to apply: ask the model, then check whether the answer holds up. Applied across a campus, that habit becomes a research culture, and Kao argues that is what produces the recognition universities want.

The bet is that the next generation of researchers will be defined as much by the questions they ask of AI as by their ability to run it, and that the universities organized around that premise will be the ones whose rankings move without chasing them.

Source: feed.koreatimes.co.kr

Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.

754 articles

Related articles

  1. 95 Percent of Organizations See No Measurable AI Gains, University Leaders Warn
  2. 79% of Students Use AI for Homework, Only 47% Get Guidance: OpenAI Flags Education 'Guidance Gap'
  3. Metis Reasoning’s Plan to Verify AI Beyond Model Weights
  4. Global experts back AI in education, demand exam-cheating detection
  5. Why 80 Percent of AI Projects Fail Before the Model Is Ever Wrong

« Previous article