Our lives are short. Our experiences are limited. Yet we spend much of our lives forming beliefs about a world we have barely experienced.
We take a particular memory, an event, or a story someone tells us and derive general rules about how the world works.
We have to. We don’t have the luxury of examining every claim thoroughly. But this makes us vulnerable to manipulation.
I want to explore a few connected ideas:
The power of anecdotes: How isolated experiences shape our beliefs and spread misinformation.
The psychology of belief: Why we seek confirmation rather than truth.
The limits of research: Why even scientific findings need scrutiny.
The spectacle of AI: How impressive demonstrations create an illusion of capability.
The antidote: Why scientific inquiry is essential to understanding reality.
A different use for AI: Building Kflow to help us challenge our assumptions.
I’ve been noticing how effectively anecdotes are used to spread misinformation, particularly in politics.
Both the left and right employ the same technique. A single incident is amplified and presented as representative of an entire group, ideology, or institution.
And it works remarkably well.
In 1973, psychologists Amos Tversky and Daniel Kahneman described the availability heuristic: our tendency to judge how common or likely something is based on how easily examples come to mind.
The more frequently we encounter a particular story, the more representative of reality it begins to feel.
The stories might all be true. That’s what makes this particularly interesting.
You don’t necessarily need to lie to mislead someone. Sometimes all you need to do is selectively present the truth.
But there’s another problem. We don’t just passively encounter these stories. We actively seek the ones that confirm what we already believe.
Psychologists call this confirmation bias. Raymond Nickerson’s 1998 review documents how extensively this tendency affects our reasoning.
It’s painful to discover that something you’ve believed for years might be wrong. Your ego is at risk, your judgment is challenged, and sometimes your entire worldview feels threatened.
So you find another example. Another article. Another expert who agrees with you.
And now, increasingly, you can ask an AI.
The obvious response is to stop relying on anecdotes and start relying on scientific research.
I agree. But we’ve also become accustomed to treating research as an authority rather than something to be examined.
Someone cites a statistic or references a published study, and suddenly their argument appears credible.
But how many of us actually examine the methodology, assumptions, or limitations?
In 2005, John Ioannidis published Why Most Published Research Findings Are False, arguing that many scientific findings can be unreliable due to statistical weaknesses, methodological problems, and bias.
His conclusions have themselves been challenged, which is precisely how scientific scrutiny should work.
The replication problem provides another example. In 2015, the Open Science Collaboration attempted to replicate 100 psychology studies. While 97% of the original studies reported statistically significant findings, only 36% of the replications did.
This doesn’t mean science is unreliable.
The strength of science isn’t that scientists are incapable of making mistakes. It’s that their claims can be examined, challenged, and corrected.
But when we cite research merely to reinforce our beliefs, we’ve replaced the authority of anecdotes with the authority of citations without necessarily improving our reasoning.
I’ve noticed a similar pattern in how large language models are marketed.
We’re shown extraordinary demonstrations. An AI builds an application in minutes, solves a difficult problem, or completes a task that would ordinarily take a professional hours.
Impressive. But what does this actually establish?
How many attempts were required? How much human intervention was involved? How reliably can the system reproduce the result? How does it compare with simpler, cheaper alternatives?
A successful demonstration answers very few of these questions.
It’s an anecdote. An impressive one, certainly, but an anecdote nonetheless.
And because LLMs communicate with extraordinary fluency, the illusion becomes even more convincing.
They can confidently explain things they’ve misunderstood, construct persuasive arguments from questionable assumptions, and sometimes agree with users even when their beliefs are incorrect.
This last tendency is known as sycophancy.
We’ve developed technology capable of helping us explore enormous amounts of information, yet we risk turning it into a sophisticated machine for validating our existing beliefs.
We ask a question, receive an articulate answer, and feel informed.
But have we actually learned anything?
Or have we simply found a better way to confirm what we already wanted to believe?
There’s an interesting contrast between the words anecdote and antidote.
Both have ancient Greek roots, though they derive from different words.
Anecdote comes from anékdota, meaning things not published or given out. Antidote comes from antídoton, meaning something given against something else, particularly a remedy against poison.
I find this a useful metaphor.
If anecdotes become dangerous when we mistake them for general truths, what’s the antidote?
For me, it’s the scientific method.
Not science as an institution or a collection of established facts, but science as a disciplined process of questioning what we believe.
We formulate hypotheses, gather evidence, test assumptions, and examine competing explanations.
Karl Popper emphasized the importance of falsifiability: a scientific hypothesis should be open to being contradicted by evidence.
Instead of asking, “What evidence supports my belief?” we should also ask, “What evidence would convince me that I’m wrong?”
That second question is much harder.
It requires intellectual humility and a willingness to separate our identities from our beliefs.
Our beliefs should be accountable to evidence. Evidence shouldn’t be selected to serve our beliefs.
I think we’ve become too comfortable with having answers.
There’s an answer for almost everything, instantly available and delivered with remarkable confidence.
But answers aren’t the same as understanding.
Understanding requires confronting uncertainty, examining contradictions, and sometimes abandoning ideas we’ve become attached to.
It’s uncomfortable work. Perhaps that’s why we’re so eager to delegate it.
But there’s something valuable in that struggle. It’s how we develop judgment and learn to distinguish what we know from what we merely believe.
I don’t want AI to take that away from us. I want to use it to make that process more accessible.
We need research, but we also need to question research. We need expertise, but shouldn’t confuse it with infallibility. And we need AI, but not as another instrument for reinforcing our biases.
Because misinformation doesn’t survive only because people lie.
It also survives because we want to believe.
We don’t need more convincing answers. We need better ways to question them.
That’s what I’m building with Kflow.
There’s still a practical problem.
Our lives are short. Research is scattered across publications, scientific papers require background knowledge, and every complex idea depends on concepts we may not understand.
We can’t personally verify everything, but we shouldn’t blindly trust every authority either.
This is where I think AI could become genuinely transformative.
Not by doing our thinking for us, but by helping us think more rigorously.
Here’s how I envision it working:
Start with a topic. The system breaks it into interconnected concepts, organized from beginner to advanced levels and backed by research.
Explore a knowledge graph. Concepts are connected through their relationships and dependencies, allowing you to understand how ideas build on one another.
Introduce your beliefs and questions. Your assumptions become claims that can be investigated rather than accepted as facts.
Examine the evidence. Claims are evaluated against supporting and contradicting research, with uncertainty made explicit.
Challenge your conclusions. The system generates counterarguments, identifies potential biases, and lets you trace claims back to their underlying assumptions and foundations.
There’s an obvious objection: LLMs are biased, and scientific research isn’t immune to bias either.
That’s precisely why neither should be treated as the final authority.
The sources, assumptions, and reasoning must remain inspectable. Conclusions must remain open to revision.
I don’t expect this process to eliminate bias or produce perfect certainty. That would be another illusion.
I want to build a tool that makes our mistakes easier to identify, our beliefs easier to challenge, and our understanding easier to deepen.
Tversky & Kahneman (1973) — Availability: A Heuristic for Judging Frequency and Probability
Nickerson (1998) — Confirmation Bias: A Ubiquitous Phenomenon in Many Guises
Ioannidis (2005) — Why Most Published Research Findings Are False
Open Science Collaboration (2015) — Estimating the Reproducibility of Psychological Science
Karl Popper (1959) — The Logic of Scientific Discovery
Sharma et al. (2023) — Towards Understanding Sycophancy in Language Models