How we check a claim

2026-10-12

What this gives us: a practice. Five questions for any new claim, the word-swaps, the quantum red flags, two bad conclusions about reports, the warning that a failed test does not show absence, and a symmetry check. Each item links back to the page it came from.

This page is a checklist made from parts already on the site. It does not decide whether any AI system is, or is not, conscious.

Five questions for any new claim

These questions used to sit on the plants brief. They are the reusable part. The plants page now keeps only what the plant cases add.

1. Is the claim about behavior, learning, or feeling?
2. Is a figure of speech, like "memory" or "talk," posing as a finding?
3. Could the system have learned to do this with no experience behind it? The design rules for that problem are on Hardest to Fake.
4. Has an outside team repeated it, and do the papers match the retellings? Two plant cases that did not hold up are on the plants brief.
5. Which theory of consciousness does the argument need, and does it survive another? The map of those theories is the reference paper.

Common bad arguments, with replies

Lifted from the plants brief. The replies are short on purpose.

• "It learns, so it feels." Learning alone does not show experience. A sensitive plant can stop reacting to a harmless repeated touch, and that is still not a finding about feeling.
• "It says it feels, so it does." Words about feelings are evidence to weigh, not proof. The tests, and what they do not show, are on Can AI self-reports be tested?.
• "No neurons, so no feeling." That follows only from a theory that requires neurons, which is the point in dispute.
• "It's just a program," or "it's just chemistry." Brains are chemistry too. "Just" argues without a reason.
• "The company says it isn't conscious." A company has its own interests, and no agreed test for consciousness exists yet, so weigh its evidence, not its say-so. A denial settles no more than a claim. That warning's home is the self-report page.
• "We're nice to plants, so..." Kindness is a choice about how we act, not evidence about what anything experiences.
• "The wood wide web shows trees talk and care." The claim that old "mother trees" feed their own young through fungi has "no peer-reviewed, published evidence" (Karst et al. 2023). The warning is on the plants brief.

Word-swaps

The ten swaps, including "no evidence" treated as "evidence of absence," and "honestly" used as a trust signal, stay in the glossary. This page does not copy them. Argue about facts, not definitions.

Red flags for "quantum" claims

These five flags come from the quantum brief, which is where the studies behind them are checked.

• Watch for "quantum" used to mean mysterious. Ask which effect is meant.
• Watch for a jump from "a quantum effect exists in biology" to "quantum explains consciousness." Enzyme tunneling says nothing about experience.
• Watch for coherence treated as entanglement, and entanglement treated as communication. All three differ.
• Watch for "proven" said about one study, a lab-protein result, or a model. Say "one study found" instead.
• Watch for "quantum" dismissed as woo across the board. Some quantum effects in biology, like enzyme tunneling, are well supported. Judge each claim on its own evidence.

Two bad conclusions about reports

People sometimes sincerely explain choices they never made. That fact gets used in two bad ways. Both are spelled out, with the split-brain and choice-blindness examples, on the self-report page.

• "Humans do it too, so AI reports are fine." Human failures do not make a model's report more trustworthy.
• "Reports are worthless, so nothing is there." An unreliable report is not evidence of absence.

A report counts only when you check it against something independent.

A failed test does not show absence

In one large patient study, 69 of 112 people who could follow commands showed nothing on the brain scans. A negative test does not show that nothing is there. Tests of what goes on inside can also fool us through bad statistics. Any AI inside-test needs a check like Goldfine's reanalysis. The numbers and the dispute are on the hidden-awareness brief.

A symmetry check

Hardest to Fake warns against self-sealing discounting: doubting only the positive signs. For each claim, ask whether we would judge a denial by the same standard.

• If we discount a positive sign because it might have been gamed, do we discount a denial, or a missing sign, for the same reason?
• If we ask an outside team to repeat a positive result, do we ask the same of a negative one?
• If we refuse to treat a model's claim as proof, do we also refuse to treat a company's denial as proof?
• If a failed test does not show absence, do we also refuse to treat a passed test as proof that someone is there?

The fourth question is the same standard pointed the other way. It is not a verdict.