How to Think Clearly When Everyone Has an AI Opinion
I recently finished Rolf Dobelli’s The Art of Thinking Clearly, 99 short chapters on the ways our minds fool us. It was remarkably entertaining and uncomfortably familiar to understand that all those biases were very inherent to me. Reading this book made me inspired to work on examples that are relevant to the AI world today. Survivorship bias in which we see routine examples of survivor stories of how someone built their AI empire in months to self-selection bias in which we find ‘it works on my laptop’ but breaks in production kind of scenarios. Our world has been flooded with examples like this one it seems. Let us review some of these biases and check what questions we can ask ourselves to protect against falling into any of these.
1. Survivorship Bias: the evidence that made it home
Abraham Wald was a statistician who in 1943, was commissioned to put armor on bomber airplanes during the World War. The data that the Air Force had was of the airplanes that came back home from the war and the Air Force examined the mapped bullet holes. It seemed obvious to armor the part with the most holes.
What Wald said was something ludicrous. He asked to armor the parts with no bullet holes. Nobody realized that a plane shot in the wing survived fine, but the one shot on the engine did not and that was something that needed protection. The missing holes on the engine were proof of its survival.
MIT’s NANDA report in 2025 showed ~95% of enterprise GenAI pilots deliver little to no measurable P&L impact on a business, but check how many blogs and YouTube videos and tutorials have been written on “How we deployed AI Agents successfully” and you will realize most of the legitimate ones are representative of the 5%.
Posts like “1M ARR in six months” from indie hackers make it seem very plausible for anyone to get there. The numbers for the business are real but the sample is still broken. These kinds of feed surface on your social media because stories of survivors catch more eyes and propel you to click the like button and hence that is most probable to land on your feed at the top.
Winners write history. We know the numbers Cursor, Harvey and UIpath has shown us and the incredible numbers they highlight. How often do we talk about the losers? Jasper, Yupp, Builder.ai and so many more that could not survive the ChatGPT wrapper stories.
The next time you are thinking of dropping off from school to start a billion dollar startup, thin again. That is why it is important for us to ask the question -
What had to happen before this evidence reached you?
2. The Swimmer's Body: what was already true of who got in
Elite swimmers have broad shoulders and long torsos. There can be two explanations that could be viable - training built such a frame or people with such a frame were selected to be swimmers.
The problem is that both are true and one cannot conclusively comment about causality here. The visible result cannot tell you how much of the outcome was selection and how much was transformation, and nobody selling you the program has an incentive to separate them.
Consider MBA schools advertising their 300K graduates' salaries running the same trade. So is a startup accelerator, a selective employer, a demanding fitness program, and a bootcamp with a 4% acceptance rate. The comparison they show you is graduates versus the general population. The comparison that would matter is graduates versus the people they rejected.
One might be very tempted to think that going to YC might make a founder successful. But take a look at their 1% success rate of accepting candidates and their rigirous process of vetting individuals.Pedigree as a proxy, founder-market fit, relentless resourcefulness are traits individuals carry within themselves and its hard to say if its because of YC or if YC inherently has individuals with these traits. Visible success is not proof that it added the thing that mattered.
Before you pay for the transformation, check who got in for free.
Before I choose this to obtain a benefit, are those benefits an inherent property of the one who joins?
3. Self-Selection: the people who left took their answers with them
Imagine you have built a survey in your community about what to build next. Or imagine that you already have your application running and you get user feedback via a Thumbs up or Thumbs down button. Voila, you have built a feedback mechanism, your job for analytics is done! Looks simple right, it isn't. The people clicking on the button or responding to the survey are the ones who stayed back - who were patient enough to grade, invest time enough to bother.
The user who never got one confident wrong answer never came back and left no rating, what about those?
The people who might not participate in surveys might have preferences too. They expressed them by leaving, which is the most honest feedback there is and the hardest to collect. Your NPS measures satisfaction among the satisfied. Your churn interviews reach the people who cared enough to reply to an email.
It is important to know
Whose absence is this data made of?
4. Availability Bias: vivid is not frequent
In February 2024, Klarna announced that its AI assistant was doing the equivalent work of 700 full time agents. The number came with a $40 million profit projection for emphasis. Presumably, Medium, YouTube and X followed with “AI replaces Support” from fringe to obvious industry wide.
Fifteen months later, the CEO told Bloomberg they had cut human support too hard and were hiring the human tier back because the quality had suffered. Something similar happened with Meta where they recorded productivity with AI going up and layoff many people to finally call them back in the company again.
Now our estimate of what AI can replace was installed by these announcements that are engineered to be remembered, and it never gets updated by the corrections that happen later on. These are classic examples of the Availability bias.
Am I estimating from a sample, or from a memory?
5. Cherry-Picking: the demo runs on the working prompts
A third-party enterprise vendor’s comparison page has ten rows. Their column is bold on all ten. Try to look at the rows, not the bold. Every dimension included is a dimension they win - latency, storage, “native agent support” - and the dimension you actually care about is missing without a footnote. This is an example of a cherry picked bias. Same goes with earlier examples of LLM benchmarks. Run the same algorithm on a handful of random seeds and the curves can disagree wildly, as was found by Henderson and colleagues in 2018 when they were training RL networks. Which means a reported result can be a method - or one flattering draw from it.
Consider you are building an LLM application in the wild. The RAG demo is flawless. It retrieves the right clause, cites the right page, answers in seconds.
Later you learn the knowledge base was written by the same team that wrote the demo questions - same sprint, same lunch meeting. The corpus and the questions grew up together. Of course retrieval finds the answer: the answer was planted the same week the question was.
What would this look like on the prompts I would have chosen?
6. Feature-Positive Effect: absence never volunteers
Your eye is built to find things that are there. This is a problem, because what is missing never announces itself. The untested failure modes, the issues that silently fail in the pipeline - you cannot see them. LLMs work only with what you give them, so they are prone to this feature-positive effect too.
Alerts fire on events - errors, latency spikes, failures. There is no alert for an event that stopped happening. The ingestion pipeline died at 3am, so there have been no errors since. Every panel reads zero and stays calm, because the dashboard has no way to draw an absence.
"Zero errors" and "no data" look identical in every tool you own. Only one of them is good news.
In the experiments, people test hypotheses by hunting for confirming instances of a feature rather than for its absence - even when the absence carries all the information. Give people a rule and they check the positive cases. The negative cases tell you more, and nobody looks.
What is this description designed to make me not notice?
7. Information Bias: more data is not a decision
Gathering information feels like progress. It is legible, low-risk, hard to criticize, and it lets you avoid the exposed act of deciding.
The team spent six weeks building the internal eval harness. It has forty tasks, three judges, and a dashboard. It is genuinely good work.
Now the test: if the score came back 62 instead of 71, what would change? Nobody can say. The ship/no-ship call happens in a different meeting, on different grounds, every time. The number is admired, exported, and pasted into updates.
An eval wired to no decision isn't measurement. It is a comfort object with a scoreboard. (Illustrative.)
The analytics layer gets installed the same way: it costs a week and produces dashboards, when ten user conversations would cost two days and produce the answer.
The discipline is embarrassingly simple - name the decision before you make the data request. What will I do differently if this number comes back high? What if it comes back low? If the answer is "nothing" in both cases, you are not gathering information. You are gathering comfort.
Which decision would change if I gathered more?
8. Clustering Illusion: randomness doesn't send a press release
Us humans have the habit of seeing patterns in the world. Astrology and constellatoons are a great example of that. Usually a pattern with a story is more comfortable than a pattern without one.
During WWII, Londoners mapped where German rockets fell. They saw clusters—streets hit repeatedly, neighborhoods spared—and concluded the strikes were targeted. Theories multiplied. In 1946, R. D. Clarke tested the map against the Poisson distribution and found the impacts fit a random process. The clusters were exactly what randomness looks like on a map.
Your dashboard is that map. A cluster of errors. A dip. A run of bad outputs. The chart shows the cluster. It does not show you the base rate, the denominator, or the Poisson. A trend needs a window, a base rate, and a sample size before it is a trend. Without those, it is a screenshot.
The same illusion runs at the model level. A model sits at the top of a leaderboard, and users infer it is the best. But the top rank is a post-selection event on noisy estimates: one 2026 study found the reported winner is the true best only 0–58% of the time under prompt resampling, and 95% confidence sets contain 8 to 35 models. The top of the leaderboard is a cluster. It looks like a signal. Often it is noise plus selection.
What would this chart look like if nothing were happening?
9. Regression to the Mean: the most underrated force in business
We over-react to anomalies. When the number drifts far from the mean, we act. When it comes back, we assume we acted well.
A sales rep has their worst quarter in two years. The VP sends them to a training program. Next quarter, they're back at their usual number. The training gets credited in the QBR deck, and the program is rolled out to the whole team.
But a bad quarter is partly luck. A big deal slipped a week into the next quarter. A key account went quiet. A competitor undercut on one renewal. The rep's real level never moved. The training didn't fix anything. The quarter just came home.
This is regression to the mean. The extreme reading triggered the intervention, and the return got credited to the intervention. The rep is fine. The training is fine. It just isn't why the number came back.
The flight instructors in Kahneman and Tversky's study learned the same lesson backwards. Praise for an unusually smooth landing was followed by a rougher one. Criticism of an unusually rough landing was followed by a smoother one. They concluded that punishment worked and praise backfired. The bounce was regression to the mean, not the feedback.
So before you rewrite the prompt, check the noise floor. If the swing fits inside the band you get when nothing changes, you are looking at weather. Keep one variant untouched and re-run the eval - if it bounces too, your fix did nothing. And wait for repetition: two weeks in the same direction is a signal, one week is a screenshot.
The question is not "did it improve after I changed something?" The question is:
Would it have improved anyway?
Sources
- Rolf Dobelli, The Art of Thinking Clearly. The modern applications here are mine. The chauffeur-knowledge chapter names the fluency trap behind the close-the-tab test.
- Mangel and Samaniego, Abraham Wald's Work on Aircraft Survivability, JASA 1984. https://www.jstor.org/stable/2288257 - the standard guide to Wald's 1943 memoranda. On how much of the popular armor story is reconstruction: Bill Casselman, The Legend of Abraham Wald, AMS Feature Column 2016. https://www.ams.org/publicoutreach/feature-column/fc-2016-06
- MIT NANDA, The GenAI Divide: State of AI in Business 2025. https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf (Preliminary findings, research period January-June 2025. Methodology: systematic review of 300+ publicly disclosed AI initiatives, structured interviews with representatives of 52 organizations, survey responses from 153 senior leaders at four industry conferences. The report's own limitations section notes self-selection.)
- Klarna's February 2024 announcement (Fortune): https://fortune.com/europe/2024/02/28/klarna-ai-altered-hiring-chatbot-700-full-time-staff-openai/ and the May 2025 reversal (Bloomberg): https://www.bloomberg.com/news/articles/2025-05-08/klarna-turns-from-ai-to-real-person-customer-service
- METR, July 2025 RCT: https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ and the February 2026 design update on a newer experiment: https://metr.org/blog/2026-02-24-uplift-update/
- Pennycook, Cannon and Rand, 2018: https://pubmed.ncbi.nlm.nih.gov/30247057/
- OpenAI, April 2025: https://openai.com/index/sycophancy-in-gpt-4o/ ; Anthropic on sycophancy: https://www.anthropic.com/research/towards-understanding-sycophancy-in-language-models
- Zheng et al., Judging LLM-as-a-judge with MT-Bench and Chatbot Arena, 2023: https://dl.acm.org/doi/10.5555/3666122.3668142
- Liu et al., Lost in the Middle, TACL 2024: https://aclanthology.org/2024.tacl-1.9/