On AI, Data Science, & Causal Models
Log of insights and news – Gorkem Turgut (G.T.) Ozer
I’m terribly sad to learn that the first consumer humanoid robot marketed to load the dishwasher (finally!) is essentially a proxy operated remotely by a human (oh no). The automation it offers is akin to hiring a teenager to mow your lawn remotely, yet it introduces privacy and latency nightmares.
How much longer must we keep loading the dishes? Until Nvidia’s valuation hits $10 trillion? Let’s buy more stonks. I’m losing my patience here.
Is our society increasingly rewarding conformity? Is AI accelerating this process, and is it simultaneously stripping work of deep meaning? After all, thanks to LLMs, many now define creativity as merely a probabilistic recombination of matrices derived from a training set.
Is science also contributing to this potential lack of deviation in culture, education, arts, architecture, and business? Here’s a take on it (not mine):
You can spot this scientific bland-ification right away when you read older scientific writing. As Roger’s Bacon points out, scientific papers used to have style. Now they all sound the same, and they’re all boring.
Maybe science is supposed to be boring? What happened to the style though?
For definitive answers to these questions, look elsewhere. For a compelling set of data (plus plenty of causal speculation) on many aspects of contemporary society and scientific style, check out this compilation and the essay here.
The greatest thinkers in science (and business) are often prolific authors. They write books, blogs, and copious emails to sharpen ideas. Richard Lewontin, E.O. Wilson, and Paul Graham are but three examples. Dorothy Hodgkin’s scientific correspondence and papers, stacked together, extend 25.85 meters in length. Great thinkers, in other words, write all the time.
Researchers are evaluated by simple measures of productivity or influence — number of papers published, citation count, and grant dollars. In such an environment, it has become exceedingly difficult for scientists to take stylistic risks in their academic writing or to devote significant amounts of time to other forms of creative writing.
I first came across this paper while writing the Machine Learning Using IV chapter of the Causal Book. Revisiting it today, I remain struck by its central finding: about 95% of the empirical TSLS (Two-Stage Least Squares) models surveyed here claim to estimate the Local Average Treatment Effect (LATE), but they fail to meet the necessary conditions to do so.
The failure is mainly due to not controlling for covariates nonparametrically. That is to say, in attempting to correct for selection bias (endogeneity) using IVs, causal modelers inadvertently introduce significant specification bias, thereby theoretically nullifying the LATE interpretation.
On a different note, I’ve resumed work on Causal Book. Updates are on the way!
Today, my feed is full of speculation linking the recent AWS layoffs, driven by increased AI automation, to yesterday’s outage. In reality, we don’t really know if AI caused any of it.
What do we know? I read two articles this morning, and one thing that struck me is that AWS was reportedly not able to diagnose the core issue for 75–90 mins. That’s an absurdly long time.
If this timeline is accurate, the extended delay is compelling evidence that critical expertise was either absent or inaccessible when it was most needed, for whatever reason.
Credit for the image goes to Emil Lendof/WSJ.
Here’s an update on using LLMs for OCR without having to use the same hammer (generic model) for all nails. DeepSeek has released an OCR-focused model: https://github.com/deepseek-ai/DeepSeek-OCR
Check out the deep parsing mode, which is parsing images within documents through secondary model calls. Very useful for data extraction. The results are pretty impressive too:
Our work represents an initial exploration into the boundaries of vision-text compression, investigating how many vision tokens are required to decode 𝑁 text tokens. The preliminary results are encouraging: DeepSeek-OCR achieves near-lossless OCR compression at approximately 10× ratios, while 20× compression still retains 60% accuracy. These findings suggest promising directions for future applications, such as implementing optical processing for dialogue histories beyond 𝑘 rounds in multi-turn conversations to achieve 10× compression efficiency.
The data on education call for attention:
– 33% of eighth graders are reading at a level that is “below basic”—meaning that they struggle to follow the order of events in a passage or to even summarize its main idea.
– 40% of fourth graders are below basic in reading, the highest share since 2000.
– In 2024, the average score on the ACT, a popular college-admissions standardized test that is graded on a scale of 1 to 36, was 19.4—the worst average performance since the test was redesigned in 1990.
The article speculates on several causal links to explain the declining trend in the metrics, ranging from the effects of COVID to the influence of smartphones and social media.
The point that truly resonates with me as an educator, though, is this: a pervasive refusal to hold children to high standards. Standards are about values, not technology or tools. No tool causes the fading emphasis on rigor.
The article discusses other important aspects, such as the disparity between school districts, the heterogeneity in outcomes based on affluence, and the potential role of AI as a democratizer, but keeps returning to the same line: declining standards and low expectations. And that’s for a good reason:
“Roughly 40 percent of middle-school teachers work in schools where there are no late penalties for coursework, no zeroes for missing coursework, and unlimited redos of tests.”
This is potentially the most important problem facing our society today, and it warrants far more attention.
It didn’t take long for Tom Fishburne’s cartoon to come true: Generative AI is increasingly treated as a magic trick. Most tricks aren’t really useful or truly “real,” yet they remain entertaining until we figure out, often the hard way, how the illusion is performed.
My conversations show that some executives are cautiously optimistic, integrating only what is truly useful into their business processes and project workstreams, while others are applauding the tricks and wanting more. The former group is already AI-literate, using machine learning and algorithms to augment and automate processes in the broader, more accurate definition of AI. The latter, more easily impressed group, seems to lack this foundation.
We can expect that more of the applauding executives will eventually join the cautiously optimistic ones once the magic show ends and we move past the peak of the hype cycle.
This is an (almost) complete picture of LLMs, highlighting the underlying math for probability calculations. It’s a great companion for the 64-picture visual summary I curated for a talk earlier this year.
Good reminder that “Life’s most important questions are, for the most part, nothing but probability problems.” Pierre-Simon Laplace in Théorie Analytique Des Probabilités
I once quoted Abraham Maslow in my note “Using predictive modeling as a hammer when the nail needs more thinking“, a problem that has since eased as focus shifts to causal modeling and optimization.
Now, if the problem is extracting handwritten digits from a PDF or an image, what’s the solution? I guess good old reliable OCR? There is nothing generative about extraction, so using an LLM may be moot, if not actively detrimental (due to hallucinations). OCR is widely available through cloud like AWS Textract and Google Cloud Vision, some of which also offer human validation.
Instead, the Reuters team intriguingly chose the shiniest hammer, an LLM, to extract data from handwritten prison temperature logs. They uploaded 20,326 pages of logs to Gemini 2.5 Pro, followed by manual cleaning and merging. One big problem with this approach is the inevitable made-up text (hallucinations), which required the team to hand-code 384 logs.
So why not OCR but an LLM? All I can think of is, LLMs may be useful in extraction when the text is highly unstructured and contextual understanding is needed, neither of which seems to apply here. Surprisingly, the methodological note doesn’t even mention OCR as a solution.
Putting the tool choice aside, though, Reuters asks an important question here: “How hot does it get inside prisons?” and the answer is “Very.” I applaud the effort and data-centric journalism, and I recommend reading the story.
Credit for the image goes to Adolfo Arranz.
How do you get to new ideas when data is always looking back?
In this latest Data Duets post, we discuss a case from United Airlines and share five key lessons from an interview with Patrick Quayle, Senior VP of Global Network Planning.
The post explores how to go beyond historical data by using:
Our Director’s Cut expands the discussion to offer insights for retail merchandising.
Enjoy the read and feel free to leave your comments on LinkedIn here.
[Click title for image]
This is so hilarious I had to share. A major issue with using LLMs is their overly obsequious behavior. They aren’t much help when I’m right; I don’t want to be right, I want to be corrected.
This project uses a Python script to count how often Claude Code says you’re “absolutely right.” The script doesn’t seem to normalize the counts by usage, which might be a good next step.
In January, I wrote a short note based on one of my talks: “How to use LLMs for learning in 2025.” In that note, I differentiated between using LLMs (1) to learn and (2) to do. With the new semester now underway, I’ve checked some usage numbers and read the Ammari et al. (2025) paper on how students use ChatGPT. I was particularly interested in the second RQ: “Which usage patterns correlate with continued or increased reliance on ChatGPT over time?”
An over-reliance on any tool, regardless of what it is, is a potential red flag for persistent learning, especially when the goal is comprehension. For example, understanding derivatives and calculating them using a computer are two distinct learning objectives. If the reliance on a tool substitutes for understanding, long-term implications may not be a net positive.
The article does not really answer the reliance part of the question. It does, however, report some interesting correlations between LLM behavior and student engagement. Notably, when ChatGPT asks for clarifications, provides unintended or inconsistent answers, or communicates its limitations, students are less likely to continue using it.
Plausible, but what these correlations mean for learning and comprehension is unclear. What is the next step after disengagement? Do they switch to another LLM to get a direct answer without having to answer follow-up questions, or do they go back to figuring it out on their own?
Class of 2029, I guess the answer lies with you. Welcome!
A recent New York Times story offers insights.
Using LLMs as a “second pair of eyes” or as a fallible assistant seems to work well. Automation also works effectively when the instructions are clear and the objectives are defined unambiguously. In both cases, human agency remains central.
Use case #15 in the article, “Review medical literature,” reminded me of a study I shared earlier (How do LLMs report scientific text?). The study showed that LLMs systematically exaggerate claims they found in the original text. The user in this case is a medical imaging scientist and is aware of the danger. When a tool isn’t foolproof, the user’s expertise and awareness make all the difference.
The high-demand use cases are quickly scaling into independent businesses with more standardized output, often with LLMs as the core wrapper. I suspect some are marketed as “magic,” and to resist that hype, users will need a combination of expertise and awareness.
If you’re a visual learner looking to deepen your understanding of AI and language models:
I’ve just made the deck from my recent talk, “AI in 64 Pictures,” available. It’s a visual journey through language processing: from word embeddings to RNNs, attention, and transformers.
Understanding AI models better helps us discover more use cases and navigate their limitations. And if you’re looking to dive deeper, you can follow the links in the slides.
Nearly half of all AI agent projects are set to fail (as Gartner predicts here). Why? Unclear business value, inadequate risk controls, and escalating costs.
As I see it, much of this is fueled by hype, leading to existing solutions being relabeled as “Agentic AI” without any rethinking of business processes.
Human creativity is missing in this picture. It’s this creative thinking that should move agent use beyond just automating or augmenting individual tasks with LLMs, leading instead to the redesign of business processes and a vision for how humans and AI can truly complement each other.
The risks and costs are more straightforward to resolve:
– Managers who are most excited about AI agents often do not fully understand the risks and limitations of LLMs. They should invest as much in understanding these models as they do in using them.
– The true cost of scaling proof-of-concept GenAI solutions is often underestimated. This is on selecting the right vendor. Gartner estimates only about 130 of the thousands of agentic AI vendors are real.
Andrew Gelman’s latest “rant” is worth a read. Everybody lies, but why, even when data clearly refutes it?
It’s interesting to think a little bit more and understand why and how people lie, especially when it comes to scientists, medical doctors, and law enforcement officials. Spoiler is, the answer is not always money.
[Click title for image]
Turns out Anthropic literally purchased and scanned millions of print books to train its language model Claude.
The judge finds that the scanning of these purchased books is “fair use” even without licensing, and goes on to say:
Everyone reads texts, too, then writes new texts. They may need to pay for getting their hands on a text in the first instance. But to make anyone pay specifically for the use of a book each time they read it, each time they recall it from memory, each time they later draw upon it when writing new things in new ways would be unthinkable. For centuries, we have read and re-read books. We have admired, memorized, and internalized their sweeping themes, their substantive points, and their stylistic solutions to recurring writing problems.
This is quite interesting, as it draws an analogy between how humans read, comprehend, and write text and how a language model operates.
[Click title for image]
One way is hitting that accept button until the code works. This is like gambling: it works until it doesn’t, and you never know why in either case.
Another way is intentional use. Intentional use is:
This means treating LLMs as over-confident, lightning-fast assistants. They are incredibly helpful in boosting productivity, quickly looking things up and delivering operational code. They can search a 100-page API documentation and find a solution to your problem in seconds.
Unless we see a structural breakthrough in how language models are built, this is also the best we can have: an over-confident assistant. LLMs don’t think or reason; they achieve (shallow) deductive closure at best. While the debate over whether LLMs “think” can be unproductive, there is a practical implication: LLMs make profoundly inhuman coding mistakes.
The errors LLMs make aren’t the errors a human assistant would make, so working with LLMs takes another perspective shift. Understanding this distinction is key to effectively using them: our coding assistant is deeply inhuman.
Otherwise, LLM-driven coding will inevitably lead to more failures in data science. Expect to hear more stories about models breaking unexpectedly.
This post was inspired by the write-up “Here’s how I use LLMs to help me write code,” which nails many other crucial points. It’s worth checking out.
LLMs excel at search and discovery. Why not use them to find IVs for causal models?
In a new section in Causal Book, Using LLMs for IV discovery and data, we offer a prompt template to help discover candidate IVs and their actual data. We tested it with the latest Gemini (2.5 Pro Preview 06-05-2025) and the results are promising.
This section is the latest addition to the IV design pattern chapter of Causal Book. The book itself aims to:
See the full table of contents here.
We’ll next dive into the regression discontinuity design pattern, which I hope will be even more fun with the newly added support in DoubleML.
In short, by exaggerating.
It seems so convenient: when you are short of time, asking ChatGPT or another chatbot to summarise a scientific paper to quickly get a gist of it. But in up to 73 per cent of the cases, these large language models (LLMs) produce inaccurate conclusions…
Over a year, we collected 4,900 summaries. When we analysed them, we found that six of ten models systematically exaggerated claims they found in the original texts.
This prompts the question: why? My initial thought was that since these models are built to generalize, and exaggeration is a form of faulty generalization, they might actually be working as intended. This aligns with the study’s finding that LLMs are nearly five times more likely to produce broad generalizations than humans.
More interestingly, attempting to prompt LLMs for accuracy often backfires. They tend to double down on generalizations, becoming twice as likely to produce overgeneralized answers when asked to stick to the facts.
On the bright side, LLMs can quickly replace the media for reporting of science.