少点错误 2024年08月22日
Just because an LLM said it doesn't mean it's true: an illustrative example
index_new5.html
../../../zaker_core/zaker_tpl_static/wap/tpl_guoji1.html

 

文章通过询问Claude服装质量相关问题,发现其引用的研究存在诸多问题,进而探讨LLM言论作为证据的不恰当性。

🌐作者以Claude为例,询问其关于现成服装质量是否下降的问题,Claude称‘许多研究’表明质量下降,但后续要求其引用具体研究时,发现部分研究是虚构的,部分存在错误引用。如‘A New Textiles Economy: Redesigning Fashion's Future’报告确实存在,Claude的部分引用内容属实,但存在理解偏差;‘The State of Fashion 2016’报告真实,但Claude关于质量和耐久性下降的说法缺乏依据;‘Valuing Our Clothes: The Cost of UK Fashion’报告存在,但Claude关于服装穿着次数下降的说法不真实。

🚫作者还指出,Claude列出的一些证明服装质量实际上有所提高的研究也是存在问题的。如剑桥大学的研究称服装平均寿命增加,但真实性存疑;美国服装与鞋类协会的研究称因质量问题退货的服装数量减少,但该研究可能并不存在;麦肯锡公司的消费者调查称消费者对服装质量的满意度提高,但调查的可靠性有待商榷;英国服装可持续发展研究中心的分析称高街时尚品牌的平均质量评级提高,但依据不明;国际纺织工程师联合会的报告称全球主要品牌的服装缺陷减少,但真实性难以确定。

💡文章最后强调,LLM会几乎接受任何给定的立场,无论其是否真实,因此不应该将LLM的言论当作支持自己立场的证据。

Published on August 21, 2024 9:05 PM GMT

This was originally posted in the comments of You don't know how bad most things are nor precisely how they're bad.; I've broken it out into a post because I think it might be a useful corrective more generally for people inclined to cite LLM remarks as fact.

I asked Claude, as an illustrative example, whether ready-made clothing had declined in quality over time, and it claimed that "many studies" indicated this was so. I subsequently asked for it to cite the studies in question; half of them were confabulated and AFAICT only one of them contained what Claude said it did. (Note that this is Claude 3 Haiku, which isn't the best; I'd expect frontier models to name four or maybe even five real papers in a list of six (and, with lower confidence, to have a more accurate understanding of the papers' contents)).

Here are the "studies" it cited—note that even some of the ones which actually exist do not contain the things Claude said they did.

    "A New Textiles Economy: Redesigning Fashion's Future" - Report by the Ellen MacArthur Foundation (2017) - This is a think-tank report (available here) rather than a study. However, it does actually exist. Claude says it "noted that the number of times clothes are worn before disposal has decreased by 36% compared to 15 years ago." It really does say this! Page 19, citing as a source the "Circular Fibres Initiative analysis based on Euromonitor International Apparel & Footwear 2016 Edition (volume sales trends 2005–2015)." A-. (Teeeeeechnically the report claims the clothing is still wearable and therefore throwing it away is a waste, which isn't the same thing as poor durability causing decreased wear time; humans cite studies that support something a little to the left of their point all the time, though, so in the interests of fairness I won't mark it down for that)."The State of Fashion 2016" - Report by the McKinsey Global Fashion Index (2016) - This one seems to be real! (Technically,  it's "The State of Fashion 2017"—their first report, so Claude can't have meant an earlier one—but it was released in 2016 and half of it is an overview of same, so IMO it's close enough.) It is not a research study, but rather a think-tank report (which is actually even worse than it seems, IMO, because on the few occasions I've checked sources on think-tank reports I've sometimes found that the results cited didn't seem to actually exist anywhere). Claude says it "concluded that the quality and durability of clothing has declined as the industry has shifted towards faster production cycles and lower prices." The report does say the industry has moved toward faster production cycles and lower prices (though it indicates that production costs have actually risen, leaving the authors quite worried about their profit margins), but does not, as far as I can tell, claim that quality and durability have declined."Valuing Our Clothes: The Cost of UK Fashion" - Study by the Waste & Resources Action Programme (WRAP) in the UK (2012) - This one actually exists! (Full report here). It's again not technically a study, but rather a "summary of the key findings of a major technical report (possibly this one?) published by WRAP"; however if they did any research themselves that's practically hard science compared to the other two, so whatever. However, Claude claims that it "found that the average number of times a garment is worn before being discarded has fallen by 36% compared to 15 years earlier," and as far as I can tell this is not true; I couldn't find a place in the report where it even mentioned the number of wearings before discarding."Apparel and Footwear Benchmarking Analysis" - Report by the American Apparel & Footwear Association - Confabulated (there are reports with similar titles from organizations which are not this one, but this specific report does not exist)"The Apparel Sourcing Caravan's New Route to the Market" - Report by the Boston Consulting Group - Confabulated"Clothing Durability and the Creation of Value" - Study by the University of Cambridge Institute for Manufacturing (2018) - Confabulated

Posting this list of sources without annotations would have been of negative value, leading unwary readers to change their minds based on studies that don't exist and forcing somewhat cannier ones to spend what could be several hours checking the bullshit I produced in a few minutes.

Moreover, LLMs will take almost any position you ask of them, regardless of whether or not it's true. Just for fun, here's Claude's list of five studies proving clothing's actually gotten better over time (veracity left as an exercise for the reader):

    A 2020 study by the University of Cambridge called "The Changing Lifespan of Garments: Evidence from Cambridge Consumer Survey Data" found that the average lifespan of mass-produced garments has increased by 30% over the past 20 years, due to advancements in manufacturing techniques and materials.Research by the American Apparel & Footwear Association called "Quality Trends in the US Apparel Industry" showed that the number of clothing items returned due to quality issues has decreased by 25% since 2015, suggesting clothing is better made.A 2019 consumer survey by McKinsey & Company called "The State of Fashion 2019" found that 66% of respondents reported being satisfied with the quality of their ready-to-wear purchases, up from 58% in 2014.A 2021 analysis by the UK's Clothing Sustainability Research Centre called "UK Clothing Sustainability Report 2021" concluded that the average quality rating for high-street fashion brands increased by 18% over the past decade, as they invested in improving manufacturing processes.A 2018 report by the International Federation of Textile Engineers called "Global Textile Industry Quality Trends" noted a 23% reduction in clothing defects across major global brands, attributed to enhanced quality control measures.

For these reasons, I do not think it is appropriate to cite the statements of LLMs as though they are evidence in favor of your position.



Discuss

Fish AI Reader

Fish AI Reader

AI辅助创作,多种专业模板,深度分析,高质量内容生成。从观点提取到深度思考,FishAI为您提供全方位的创作支持。新版本引入自定义参数,让您的创作更加个性化和精准。

FishAI

FishAI

鱼阅,AI 时代的下一个智能信息助手,助你摆脱信息焦虑

联系邮箱 441953276@qq.com

相关标签

LLM言论 服装质量 研究问题 证据恰当性
相关文章