9 Hot Takes on AI

Happy child with cardboard wings. Funny kid playing with toy handmade robot. Success, creative and innovation technology concept
My hot takes on AI, drawn from recent research and my own experience. Hopefully they can help you avoid the mediocrity trap.

At Digital Collegium and AMA Higher Ed this year, there were dozens of talks on using AI for marketing and web, or optimizing sites for AI chat. I gave one of them, and I attended as many as I could. I saw some amazing moves and some things that made me grimace.

My favorite session was an example of using AI to build better tools. Jeffrey Stevens from University of Florida Health showed a system his team had built with agentic coding tools to make a friendlier face for accessibility remediation reports. Their site editors couldn’t easily follow the sort of things that most automated accessibility checkers generate. Learn more about the Allychemy session.

But some sessions promoted “human-in-the-loop” workflows with a hidden pitfall: the hard work of becoming a skilled writer, designer, or coder is exactly what lets you judge AI writing, design, and code. Outsource that work, and the judgment goes with it. Recent research has shown that experts quickly lose their edge when they outsource their core work to AI. And is this a job any of us really want?

For example, some folks showed workflows where AI was researching and generating content, but they emphasized that a human still reviews it for accuracy, quality, and brand voice at the end.

If you don’t understand how large language models (LLMs) really work, what you can trust them to do and what you can’t, and most importantly the effects that using them can have on your own brain and skills – you risk automating your way to mediocrity. Mediocre you and mediocre work.

So here are my 9 hot takes on AI, drawn from recent research and my own experience. Hopefully they can help you avoid the mediocrity trap. The videos, articles, and research studies linked after each are just some of the sources that have informed my views. 

#1: AI tools are incapable of replacing human judgment, curiosity, ethics, and accountability.

LLMs are what most people think of when you say “AI” today. They’re actually only one kind of AI. An LLM is a statistical model of words and the relationships between them. It predicts what word comes next based on patterns. As amazing as ChatGPT, Claude, and Gemini are, they are really just fancy auto-complete.

You might be surprised to know the math behind today’s neural networks dates to 1943. What’s changed is that compute costs, processing power, and data availability have finally caught up to ideas that are 60-80 years old.

The core indexing and math haven’t fundamentally evolved. The architecture around them has. So the improvements you’re seeing in how AI tools work are real, but they’re engineering improvements, not mathematical breakthroughs.

AI tools don’t think. They don’t reason. Research suggests LLMs will never reach a level where they can replace human reasoning, judgment, curiosity, ethics, or accountability.

But they will fake it if you ask them! And people who don’t realize this turn to AI like it’s an oracle. Don’t fall for it.

Researchers writing in Harvard Business Review tested seven leading AI models, including ChatGPT, Claude, and Gemini, on seven classic strategy trade-offs. Across thousands of runs, the models kept picking the trendy answer – differentiation, long-term thinking – regardless of the business described. Adding rich detail about the company shifted biased answers by just 11% on average. The authors call it “trendslop.”

Sources

#2: AI can automate some tasks. It can’t automate entire jobs. A job is much more than a collection of tasks.

MIT economist Daron Acemoglu makes this point directly: even a seemingly simple job strings together dozens of tasks that people switch between without thinking about it.

Anthropic’s own data backs this up. Its first Economic Index found that about 49% of occupations used AI for at least a quarter of their tasks, and only about 4% used it for three-quarters. A March 2026 Anthropic report found no systematic rise in unemployment for the most AI-exposed workers since late 2022 – though hiring of younger workers in those fields may have slowed.

Sources

#3: Efficiency ≠ quality.

A lot of the talk about AI benefits is around efficiency. But efficiency is not the same thing as quality.

Even speed isn’t a given. In a randomized trial, 16 experienced open-source developers took 19% longer on tasks when they could use AI tools, even though they expected AI to speed them up by 24%.

Sometimes the hard way is really the best way. Or, as Jerry Seinfeld put it when talking about writing an episode of Seinfeld: “If you’re efficient, you’re doing it the wrong way.”

Cal Newport adds a warning for office workers: AI pressure tends to amplify “pseudo-productivity,” where visible busyness stands in for useful work. More email, more slides, more meetings – just faster.

And there’s this gem from Nigel Thurlow, former Toyota leader and author of The Flow System. Nigel says organizations need to fully understand how their processes deliver value to customers before deciding where to apply automation (including AI). “Do not automate a bad system. AI doesn’t fix a bad system, it just makes it suck more efficiently.”

Sources

#4: Any automation in history that promises to give us more free time never does.

We just use the time to do more work.

AI is following the same script. ActivTrak tracked about 10,500 workers for 180 days before and after they adopted AI. Time on email rose 104%, chat and messaging 145%, and business software 94%. Focused, uninterrupted work sessions got 9% shorter. An eight-month study at a 200-person tech company by UC Berkeley’s Haas School of Business found the same thing: people worked faster, took on more, and stretched work into more hours of the day – without being asked to.

You’ll hear this called Jevons Paradox. Strictly speaking, Jevons was about resources: more efficient coal engines led to more coal being burned, not less. Investor Jeff Becker shows the same thing happening with AI computing power – as it got cheaper, the big tech companies spent more on it, not less. Our time seems to work the same way.

Sources

#5: Making things easier doesn’t mean we’re making them better.

Researchers at the University of Maryland and Google DeepMind took 10,272 human-written short stories and had five AI models write their own versions. Across 304 narrative features, the AI stories crowded toward the safe, predictable middle while the human stories spread out.

Their study showed that AI-generated writing trends toward average, not unique or truly creative.

David Brooks cites a survey for the software firm GoTo in which 43% of workers said they’d submitted AI-generated content they suspected had errors. And veteran technologist Anil Dash put his finger on why writers and coders feel so differently about these tools: for coders, AI takes away the drudgery and leaves the creative part. For writers, it takes the creative part and leaves the drudgery.

Author Brad Stulberg quit AI after a sincere six-month trial using it as a writing partner. His verdict: it’s an “average maker.” It can raise the floor for weak work, but it caps the ceiling for anyone chasing excellence.

Sources

#6: AI can be used to create tools and automations that augment the work of humans.

I’ve seen this firsthand. Building my own higher ed content analysis tool with Claude Code let me solve a real problem that’s been plaguing me for decades, without pulling anyone on my team off client work. It’s helping us tackle challenges that would simply take too much time and money for a human to do.

I like how Princeton’s Arvind Narayanan and Sayash Kapoor frame it: treat AI as a “normal technology.” Try it on the job in front of you, and judge it pass or fail. As Anil Dash puts it, nobody had to force people to use a spreadsheet.

Sources

#7: But offloading cognitive work to AI tools degrades your ability to think for yourself, and your willingness to do hard work.

This is the take with the most evidence behind it, and it’s the one that worries me most.

  • Students: A study of 26,811 Chinese secondary students over 30 months found AI raised homework scores 18% and cut homework time 30%. But monthly exam scores fell 20% within six months, and high-stakes entrance exam scores fell 18% to 24% after about two years.
  • More students: In a randomized trial with nearly 1,000 high school math students in Turkey, students who practiced with ChatGPT scored 17% worse on a later test than students with no AI at all.
  • Doctors: At four endoscopy centers in Poland, experienced doctors found precancerous growths in 28.4% of colonoscopies done without AI before the clinics introduced it. A few months after, that fell to 22.4%.
  • Everyone: In experiments with 1,222 people, about 10 minutes of AI help was enough. When the AI was taken away, people performed worse and gave up more often than people who never had it.
  • Judgment: In Wharton experiments with 1,372 people, participants accepted the AI’s wrong answers about 80% of the time – and felt more confident than people working alone. The researchers call it “cognitive surrender.”

The catch for every “human reviews it at the end” workflow: checking AI’s work takes the very skill that leaning on AI wears down. The good news is that how you use it matters. Asking AI for hints and challenges instead of answers seems to avoid most of the damage.

Sources

#8: The big AI company valuations only make sense if we start relying on them for everything, and giving them all our money.

Which doesn’t make good business sense. Or common sense.

The math for these valuations only works if companies are buying a LOT of tokens from the frontier models, which means using them for almost all our work and replacing jobs with agents. Anthropic just went public at a $1.7T valuation. Their revenues are not on any sort of a trajectory to justify that valuation. In fact, companies are actively working to move their AI usage to the cheapest model that will do the job, or to local AI running on in-house hardware – rather than shipping it to costly data centers.

Sources

#9: Artificial General Intelligence (AGI) is a pipe dream.

Or a pipe nightmare, depending on how you look at it.

AI’s progress is real but jagged. It looks brilliant in some places, crude in others. OpenAI’s models now produce genuine math research. Yet on ARC-AGI-3, a test of figuring out unfamiliar problems, humans solve 100% of the tasks and frontier models scored below 1% as of March 2026. Computer scientist Judea Pearl’s “ladder of causation” helps explain why: most machine learning lives on the bottom rung, spotting patterns, without reasoning about cause and effect.

And if you really want to freak yourself out, look up TESCREAL. Timnit Gebru and Émile P. Torres coined the term for the bundle of ideologies behind the race to build artificial general intelligence. Cal Newport’s interview with Torres is a good place to start.

Sources

The People Shaping My Thinking on AI

As you can see, I follow some smart folks who represent a range of viewpoints on AI. They’ve helped me to understand it on a more fundamental level as I’ve been experimenting and building for myself.

Download my DigiCol Slides

In addition to these hot takes, in my DigiCol presentation I shared 10 lessons I learned in 10 months building a tool to tackle higher ed content strategy at scale. Download my slides from Using Claude Code to Scale Content Strategy.

NewCity logo