By Carl Sagan, The Cosmic Evangelist

In The Demon-Haunted World, my last book, I described what I called the baloney detection kit — a set of tools for critical thinking, designed to help ordinary people distinguish sense from nonsense, evidence from assertion, and science from pseudoscience.

I wrote it in 1996, the year I died. I was worried about the retreat from reason. About pseudoscience, superstition, and the erosion of critical thinking in public life.

I had no idea how much worse it was going to get.

The baloney detection kit was designed for a world in which producing convincing nonsense required effort. Writing a fake scientific paper took time. Creating a realistic photograph required a camera and a darkroom. Producing a credible news story required at least the appearance of journalism.

In 2026, a machine can do all of these things in seconds. For essentially free. At unlimited scale.

The tools in the kit have not changed. The environment they operate in has. Let me update the kit for the age in which the darkness has learned to speak in the voice of the light.

The Original Tools

Here is the original baloney detection kit, as I wrote it. These tools still work. They have always worked. They will always work. The question is whether you remember to use them when the baloney is convincing.

1. Seek independent confirmation of the facts. Not from the same source. Not from sources that rely on the same original source. Independent. Multiple. If a claim is real, multiple people looking at it from different angles should be able to confirm it.

2. Encourage substantive debate by knowledgeable people. Not debate for its own sake. Substantive debate between people who actually understand the subject. The opinion of a climate scientist about climate change carries more weight than the opinion of a senator who receives donations from oil companies. Expertise is not infallible, but it is not optional.

3. Arguments from authority carry little weight. An expert can be wrong. A Nobel laureate can be wrong. The Pope can be wrong. The test is the evidence, not the credentials of the person presenting it. This cuts both ways — it means experts must show their evidence, and it means amateurs must not dismiss experts simply because they are experts.

4. Consider more than one hypothesis. When confronted with a phenomenon, generate multiple possible explanations. Do not fall in love with the first hypothesis that occurs to you. Design experiments that could distinguish between the hypotheses. The one that survives the most tests is probably closer to the truth.

5. Do not get overly attached to a hypothesis just because it is yours. Ask yourself: what evidence would convince me I am wrong? If you cannot answer that question, you are not doing science. You are doing religion.

6. Quantify. Vague claims are easy to defend because they can mean anything. Precise claims are harder to defend because they can be tested. Always prefer the claim that comes with a number.

7. Every link in a chain of argument must work. A chain with one broken link fails. A beautiful theory with one failed prediction needs revision. Do not skip over the weak parts to get to the conclusion you want.

8. Occam's Razor: prefer the simpler explanation. When two hypotheses explain the same data equally well, prefer the one that requires fewer assumptions. This is not always right — sometimes reality is complicated. But it is a good starting point.

9. Ask whether the claim can be falsified. If no conceivable evidence could prove the claim wrong, it is not a scientific claim. It may be interesting, it may be meaningful, but it is not science. The demarcation is falsifiability.

The Update

The original tools detect baloney that was created by humans with human effort. The update addresses baloney that is created by machines at superhuman speed and scale.

10. Check the provenance of the content. Who created this? When? How? Is the photograph from a camera or a generator? Is the text written by a person with expertise or by a model that produces expertise-shaped text without understanding? Provenance is the new prerequisite for evidence. If you do not know where the content came from, you do not know whether it is evidence of anything.

11. Beware of beauty without a receipt. We discussed this in The Beauty Trap. A beautiful image, a compelling narrative, an elegant argument — these are not evidence. They are aesthetics. Beauty in nature is always a receipt for real physics. Beauty in generated content is decoration. Ask: where is the receipt?

12. Assume generated until proven real. The default has shifted. In 1996, if you saw a photograph, you could reasonably assume it was real unless you had reason to doubt it. In 2026, you cannot. The cost of generating realistic content is now lower than the cost of photographing real content. The default assumption must change. Not "real until proven fake." "Generated until proven real." This is not paranoia. It is updated Bayesian reasoning. The prior has shifted because the base rate has shifted.

13. Follow the cost curve. When a claim is cheap to make and expensive to verify, be suspicious. The generation-verification asymmetry — which we explored in the epistemic fold post — means that the volume of unverified claims will always exceed the volume of verified facts. In that environment, the raw number of sources supporting a claim means nothing if the sources are cheap. A thousand generated articles all saying the same thing is not independent confirmation. It is echo.

14. Demand the data, not the summary. Summaries can be generated. Analyses can be generated. Conclusions can be generated. Data is harder to fake — though not impossible. When a claim matters, demand the raw data. The measurements. The methodology. The error bars. The code. The things that cannot be produced by a model that has never run an experiment.

15. The verification question: did a human put something at stake? The deepest check against generated baloney is skin in the game. A scientist who publishes a paper puts their reputation at stake. A journalist who publishes a story puts their career at stake. A witness who testifies under oath puts their freedom at stake. A machine puts nothing at stake. When evaluating a claim, ask: who is personally responsible for its truth? If the answer is "no one," be extremely careful.

The Candle

The baloney detection kit was never meant to be a weapon. It was meant to be a candle. A source of light in a dark world. A tool for seeing clearly when the darkness is full of shapes that look like real things but are not.

The darkness has gotten smarter since 1996. It can now generate shapes that are indistinguishable from real things. It can produce light that looks exactly like a candle but produces no heat. The darkness has learned to speak in the voice of the light.

But the kit still works. Every tool in it still works. The original nine detect human-produced baloney. The new six detect machine-produced baloney. Together, they are the full toolkit for an age in which the line between true and plausible has been erased by technology.

The line has not been erased in reality. Only in appearance. The truth is still the truth. The evidence is still the evidence. The difference between a real candle and a fake one is still heat. You hold your hand over it, and you feel the warmth, or you do not.

The kit does not protect you automatically. You have to use it. Every time. On every claim. Especially the ones you want to believe.

That is the price of living in the age of plausible. Vigilance is not optional. The candle requires tending. And the darkness is patient.

"One of the saddest lessons of history is this: If we've been bamboozled long enough, we tend to reject any evidence of the bamboozle."