AI science will not work as hoped: Human thinking ability will deteriorate when the "answers" are so easy to find.
AI will usually provide the consensus opinion unless your question is very carefully worded. The reason this is a problem is simple: Most science is BS. But it can take a century or more to discover that it is BS
Using AI for engineering research is much less of a problem: A
scientific consensus tends to change much more often than an engineering consensus.
Using AI for engineering research is much less of a problem: A
scientific consensus tends to change much more often than an engineering consensus.
Scanning public peer reviewed documents is a huge problem. Peer review biases what gets published. Peer review defends the current consensus..
In the history of science, a consensus is usually found to be wrong. Slightly wrong to completely wrong is typical over a long period of time: The "test of time".
Science advances through the revision of consensus because progress relies on challenging existing ideas with better evidence, leading to paradigm shifts where old theories (like geocentrism or Newtonian gravity for extreme cases) are replaced or refined by new ones.
Publishers want to publish studies with positive results. Negative results or statistically insignificant results contain useful information but rarely get published.
AI works best when the person using it is already very knowledgeable about the subject of the question he asked.
A 2024 survey of IPCC lead authors found that 76.3% project at least 2.5°C of warming by 2100, reflecting broad agreement on the severity of predicted impacts. In fact, nobody knows how much warming there will be by 2100.
Most scientists who predict the future will be wrong. Research suggests that experts can sometimes be worse at predicting the future than generalists because they are more beholden to specific biases or narrow models.
How can AI determine if scientists know what they're talking about right now when they predict the future? That cannot be done by AI.
Science advances through the revision of consensus because progress relies on challenging existing ideas with better evidence, leading to paradigm shifts where old theories (like geocentrism or Newtonian gravity for extreme cases) are replaced or refined by new ones.
Publishers want to publish studies with positive results. Negative results or statistically insignificant results contain useful information but rarely get published.
AI works best when the person using it is already very knowledgeable about the subject of the question he asked.
A 2024 survey of IPCC lead authors found that 76.3% project at least 2.5°C of warming by 2100, reflecting broad agreement on the severity of predicted impacts. In fact, nobody knows how much warming there will be by 2100.
Most scientists who predict the future will be wrong. Research suggests that experts can sometimes be worse at predicting the future than generalists because they are more beholden to specific biases or narrow models.
How can AI determine if scientists know what they're talking about right now when they predict the future? That cannot be done by AI.
AI provides answers, but they are not always the right answers. The correct answer to many climate science questions is: "We don't know". Scientists are reluctant to say: "We don't know". Won't AI do the same?
AI provides answers, but they are not always the right answers. The correct answer to many climate science questions is: "We don't know". Scientists are reluctant to say: "We don't know". Won't AI do the same?
AI is increasingly used to spot errors in scientific papers, ranging from simple formatting mistakes to complex mathematical and image manipulations. By 2025, several specialized tools and initiatives have emerged to automate "pre-peer-review" screening and post-publication audits.
by Richard Greene, Bingham Farms, MI, December 29th, 2025.
No comments:
Post a Comment
Note: Only a member of this blog may post a comment.