Arguments
How machine learning holds a key to combating misinformation
Posted on 13 December 2021 by John Cook
“A lie can travel halfway around the world before the truth can get its boots on.”
This quote appears in many forms. In some variants, the quote involves footwear. In other cases, the truth is struggling to get its pants on.
Regardless of the details, the sentiment encapsulates a key challenge of misinformation. By the time the meticulous task of fact-checking is complete and the correction has been disseminated, the misinformation has already spread widely and achieved all sorts of mischief.
Consequently, misinformation researchers speak wistfully of the “holy grail of fact-checking” – automatically detecting and debunking misinformation in one fell swoop. Machine learning offers the potential of both speed and scale – the ability to identify misinformation the instant it appears online, and the technical capacity to distribute solutions at the scale required to match the size of the problem.
But the holy grail quest faces a seemingly insurmountable hurdle. Misinformation evolves and sprouts new forms. How can you detect a myth before you even know what it is or what form it will take?
Misinformation and climate change
When it comes to misinformation about climate change, you often hear the terms “whack-a-mole” or “climate zombies” – typically expressed through clenched teeth. These refer to the fact that climate myths never seem to die, persistently rearing up to be debunked over and over. Indeed, the misleading arguments found in climate misinformation in the early 1990s are the same myths we now hear in 2021.
While this can be annoying, climate zombies present a research opportunity. The fact that climate misinformation shows so much stability makes it possible to train a machine to detect misinformation claims.
A number of years ago, myself and my colleagues Travis Coan and Mirjam Nanko from Exeter University, as well as Constantine Boussalis from Trinity College Dublin, began our quest for the fact-checking holy grail – specifically focused on misinformation about climate change.
The first step in this process was building a taxonomy of contrarian claims. As we developed and refined the many claims we were seeing in climate misinformation, five main categories became clear – it’s not happening; it’s not us; it’s not bad; solutions won’t work; and experts are unreliable.
These five categories of climate misinformation are noteworthy because they directly mirror the five key climate beliefs developed from survey data by Ed Maibach – it’s happening; it’s us; it’s bad; there’s hope; and experts agree. Consequently, we called our five categories of climate misinformation the five key climate disbeliefs.
Once we had our taxonomy, it was time to roll up our sleeves and start training the machine.
The principle of supervised machine learning is straightforward – take a paragraph of text from known sources of climate misinformation, and match it to a contrarian claim in our taxonomy (if there is a match). Then repeat that same process tens of thousands of times, until our machine is sufficiently trained to detect each misinformation claim. (Easy, right?).
Fortunately, we were able to draw upon the help of the climate-literate Skeptical Science team (which had form on crowd-sourcing content analysis of large climate datasets).
Once we had trained our machine to detect and categorise different misinformation claims, we fed our model 20 years’ worth of climate misinformation – more than 250,000 articles from 20 prominent conservative think-tank websites and 33 blogs. It’s the largest content analysis to date on climate misinformation, making it possible to construct a two-decade history of climate misinformation.
The results weren’t what I expected at all.
The erosion of public trust in climate scientists
During the past 15 years, I’ve been debunking scientific climate misinformation – the type of myths that fell under the categories “it’s not happening”, “it’s not us”, or “it’s not bad”.
It turns out these were the least common forms of climate misinformation. Instead, the largest category of climate misinformation was attacks on scientists and on climate science itself.
Climate misinformation isn’t about providing its own alternative explanation of what’s happening to our climate. Instead, it’s focused on casting doubt on the integrity of climate science, and eroding public trust in climate scientists.
This has significant consequences for scientists, educators, and fact-checkers. The majority of our efforts have focused on debunking scientific myths such as “global warming isn’t happening” or “climate change is caused by the sun”.
But that’s not where misinformation is focused – the focus is on attacking scientists and science itself. There’s a dearth of research into understanding and countering this type of misinformation, let alone public engagement and education campaigns to counter their damage.
Another strong trend was a growing prevalence of misinformation targeting climate solutions – claims that climate policies were harmful, attacking renewables, or spruiking fossil fuels. This category is becoming an increasingly dominant proportion of climate misinformation. This is particularly the case with conservative think-tanks, which tend to focus more on climate policy than science denial.
The overall pattern in our data is clear – solutions denial is the future of climate misinformation.
Our research was recently published in the Nature journal Scientific Reports. This was an important first step on our quest for the fact-checking holy grail. The next step is to synthesise our machine learning research with critical thinking research into deconstructing and analysing climate misinformation.
This task requires bringing together the vastly different disciplines of computer science and critical thinking philosophy. This is challenging, but interdisciplinary solutions are essential when dealing with complex, interconnected issues like misinformation.
We still have a long way to go, but for now it’s important to recognise the lessons already learnt while pursuing this quest. Not to mention the friends made along the way.
This article was first published on Monash Lens. Read the original article
Machine learning is only as good as the annotation of the examples. I doubt that the team of "climate literate" volunteers can do that objectively. Do they really understand climate policy? The lackluster recall for the "climate solutions won't work" suggests another problem: drift. ML performance is also determined by the sample set and using the 1998 to 2020 corpus, because that's the data you happen to have gathered, causes a problem.
Examples of contrarian climate solution claims from 1998 are not similar to examples from 2020 (and vice versa) because climate solutions have changed too much. A related problem shows up in the further analysis of contrarian funding. Cato, as just one example, has changed a lot from the days of climate contrarian Pat Michaels to the current climate policy writings which appear to be headed by a lawyer writing about policy. The bulk of the paper appears to be a rehash of those old grievances, and not a precise critique of what is actually wrong about so-called contrarian policy.
John - you wrote "It turns out these were the least common forms of climate misinformation. Instead, the largest category of climate misinformation was attacks on scientists and on climate science itself."
I agree that smearing the science and scientists has indeed been the predominant form of denial/pathological scepticism for a long time - it's what I've found from my own experience tackling the toughest exponents, however I think the mechanisms they use to achieve the 'smear' are still the old tried and true 'Skepsci' favourites - Soon's 'it's the Sun', Climategate, Briffa's Yamal tree rings, Curry's 'uncertainty monster', Mann's hockey stick PCA's, Svensmark's cosmic rays, Morner on sea level etc. etc., although the originators are not nowadays mentioned by name so often these days - they don't need to be - their 'sceptical' objections have become established as canon in the denialosphere.
There seem to be two important points in the posted statement. The first favors the capability of machine learning. The second illustrates a belief in the value of more efficient fact-checking, and that such will better expose wrongful information, thus making the world a better place. I would suggest two books for review, one for each topic of false hope in the science that brought us climate change.
1. Shockwave Rider, John Brunner, 1975: somewhat of a bible for hackers that used its terminology for their purposes. Brunner makes an important distinction between artificial (machine-based) intelligence and natural (nature-based) intelligence. With this he outlines an eternal weakness always found in machine-based anything as it is presented in science fiction, thus carrying over into presentation of science fact.
2. On Bullshit - Harry Frankfurt, Princeton U Press. 2005: A philosophical doorway into why fact-checkers are to become irrelevant to general public discourse. Speakers/writers that intend to lie once cared about lies and how best to cover their basis up, as in science articles with weakness but purpose, yet need to be withdrawn. That has now changed. Such personalities with immortality complexes now work to create systems of lies. These are tightly bound into bundles where concern with the true and false are replaced by emphasis on anything it takes to persuade the listener/reader. You will see why this became one of Princeton U Press's best selling books. It explains why politics has changed, and why we should avoid problem solvers that act like German leadershp of the 1930s. They came to ban discussion of "politics, off topic comments and ad hominem" statements, or anything else that questioned their assumed basis for their statements of the central problem of the society supporting them. (My keynote lecture to the annual meeting of the Leibnitz Society in 2007 was on this use of a qualifier in England that year, and no longer in Germany. Now its widely used in the US.)
Yes, discourse on the when, why and where of climate change has evolvled since the work of Eunice Foot in 1856, but those are noise factors, not a basis for the essential change that humans over 20 years old are not prepared to make.
This comes from my 1979 book at the University of Pennsylvania, then its 2019 reprint "Too Early, Too Late, Now what?. That book illustrated why politics are key, where a political point of view defines your relations with nature, each other and self. The two year study it was based on, with 20 companies and 6 govenments, illustrated why analaytic politics would make regulation of climate change irrelevant to the problematique; the system of problems, not a narrow problem from whatever you decided to analyze that day.
[BL] Repeated posting about the same book from the 1970s deleted. You have been warned about this before.
This is an excellent development within the scope of SkS (to identify and counter climate science misinformation and disinformation). Hopefully others outside of the SkS scope will be able to extend this learning to help identify, restrict and penalize individuals and organizations who are the prime harmful motivators of the development and dissemination of misinformation and disinformation. The ‘whack-a-mole’ challenge of addressing harmful actions, like misleading appeals to easily impressed people, after the fact has a limited ability to limit the harm done.
‘After the fact’ of the development of harmful beliefs and actions it is hard to completely amend and end the harm being done. Effectively and pro-actively limiting the harm done requires the high status people associated with, and benefiting from, harmful actions to be identified and be effectively corrected and limited by:
Note: The current legal need for an identifiable person with ‘standing - recognition’ in a legal system to provide substantial proof that they were personally physically or monetarily harmed by specific provable ‘actions of another identifiable person or corporation that the legal system applies to’ has failed to protect humanity from many harmful developments, not just the harm of climate change impacts. Many people are working to develop improved legal systems that address inter-generational and inter-national harm. But they lack popular support among global leadership. And they are attacked just like climate scientists are attacked, typically as promoters of Evil Socialist Global Government. Also note that Government can be understood to mean ‘To act collectively to Guide and Limit everyone’ and that really irritates people who are determined to have more freedom to believe and do as they please.
That is my perspective as a Professional engineer in Canada who obtained an MBA in the 1980s and has tried to pay attention to what is going. I have learned that engineers sustainably succeed by self-governing their pursuit of learning about ways to limit harm done. They seek increased awareness and constantly improved understanding of what is potentially harmful. Limiting harm done requires understanding the root causes. Repairs to an identified harmful result can look like the harm has been dealt with. But if the cause of the harm has not been properly identified, what appears to be a repair has not solved the problem. The harm will re-occur like whack-a-mole, or climate-zombie beliefs.
Responsible engineers can’t just try to create the appearance that a harmful result, or risk of harm, has been addressed. They need to pursue understanding of what caused the harmful result. Then they pursue ways to properly rebuild things to avoid future problems. They will also extend their new learning to everything that has been built. And they will take things out of service until the harmful problem is able to be corrected. To a responsible engineer, nothing harmful is so important that it must be allowed or continue in service (compromised bridges and buildings can stay in service with reduced use limits).
Responsible engineers limit freedoms of others for the benefit of everyone. Their actions even benefit the people they place limits on in spite of some of those people being so determined to personally benefit that they angrily resist understanding that their actions should be limited, or be more expensive (Note: it clearly can be understood that being able to afford to be harmful is unacceptable. An engineer should not accept a higher payment as the ‘marketplace based justification’ for providing a more harmful or less safe service).
A key understanding is that “pursuits of personal interest in competition for perceptions of superiority relative to Others” compromise that fundamental engineering understanding of “Do not allow harm to be done”.
The engineering pursuit of the root of the problem leads me to consider 4.3.2 “Low Public Support” to be the key objective of the climate science denial system. And closely related points are 4.1.1 “Policy increases cost” and 4.3.5 “Limits Freedoms”. Increased public awareness and understanding will lead to support for policy that makes it more expensive or more difficult to continue to benefit from the harmful systemic developments. The required action is to limit the ability of people to believe whatever they want as the excuse for doing as they please. And “Limiting Freedom” is an insidious argument. More awareness and increased evidence limits the freedom to believe things to the subset of beliefs that are not contradicted by the evidence. And the “Limiting Freedom” complaint is easily liked by anyone who wants to be freer to do as they please in defiance of being able to learn that what they want to do is harmful or risks causing harm. The “Limits Freedom” and “makes things more expensive” arguments are powerful ways to make something “Less Popular”. The many other categories of made-up claims are also ways to get a diversity of people to have a stronger harmful selfish attitude.
A master stroke is the use of nonsense claims about 4.1.4 “Rich future generations”. That fairy tale is based on the fatally flawed holy grail belief in the Constantly Richer Future because GDP per capita has continued to increase so far. Increased awareness and improved understanding has amply exposed how destructive the continued Growth of GDP has been because harmful activity counts and, as a result, harmful GDP contributions won’t be shut down unless a cheaper and easier alternative is developed that maintains the fatally flawed belief that growing GDP is required to develop lasting improvements.
The evidence harmfully contradicts developed popular beliefs that unjustly excuse or defend harmful activities that some people benefit(ed) from to the detriment of others, especially to the detriment of poor people, people in other nations, or future generations (people with little or no legal or marketing power). So the people who pursue increased awareness and improved understanding of the harm of developed beliefs and actions need to be attacked in order to delay the correction of harmful unjust developments. The result is the powerful need for the harmfully selfish who are learning resistant to harmfully unjustly impugn with impunity.
Significant system changes are required. Popularity and profit have proven repeatedly to be a poor basis for deciding winners in socioeconomic competition. Perceptions of good results from a system do not excuse or make up for harmful results produced by the developed system. Even perceptions of poverty reduction due to fossil fuel use are a harmful unsustainable developed perception.
The current systems are the developed results of people who have (had) status allowing them to form and transform the systems they are part of. And the evidence indicates that there is a long history of the problem being “people who develop a willingness to benefit from harmful actions winning higher status”. Wealth and power enables the harmful to make the system more harmfully suit, and defend, their interests.
An example would be the case of people who have developed or obtained and shared “what wealthy powerful people wanted to keep hidden”, like Chelsea Manning, Edward Snowden and Julian Assange, being tracked down for punishment. People who pursued information they could twist into attacks on increased awareness and improved understanding of climate science and attacks on people who try to increase other’s awareness and understanding have not been “as powerfully pursued for punishment”. There are laws that were broken by Manning, Snowden and Assange that powerful wealthy people applied, along with considerable resources they control for their interests, to the maximum capability. Comparatively, very little has been done by powerful wealthy people about Climategate, other than the actions by many of them to maximize the personal benefit they obtain from the unjust impugning with impunity they could get away with.
Returning to engineering, doing something helpful should not absolve a person of any harm they continue to cause. Ending harmfulness is required to get any credit for helpfulness. Imagine a temptation for personal benefit leading to a lack of diligence to properly review all aspects of the design and construction of a building, with the result being a building where the majority of the structure is perfectly sound, but small parts are fatally flawed (like some: balconies or railings failing, windows not keeping the rain out or falling off). Do the parties who were less diligent than they could and should have been deserve “massively net-positive evaluations” because the vast majority of the building was well designed and well built?
Learning to be less harmful and more helpful to Others (especially the future generations), and understanding that the non-human non-technological environment needs to be protected and improved, is anathema to harmful individualists (Libertarian Freedom demanding types – the types demanding that all opinions are equally valid). Helpful collectivists need to also be diligent to protect against authoritarian rule. They need to try to help Righteous Minded people avoid becoming harmfully righteous. The Righteous Mind, by Jonathan Haidt, explains some human predispositions that are potentially serious impediments to efforts to get everyone to learn to be less harmful and more helpful, especially concerning when those traits are connected to the evidence-based concerns presented by Timothy Snyder in On Tyranny.
The roots of a problem need to be understood in order to develop a sustainable solution. A root of this climate science denial problem appears to be the developed harmful cheaper and easier ways of enjoying life that are an undeniable legacy, and continuing to grow, problem of humans who develop harmful attitudes and pursue “improvements that are restricted to what they benefit from”. Humans who have developed that bias, rather than a bias for learning to be less harmful and more helpful, can be expected to fight against increased awareness and improved understanding that what they benefit from is harmful. From the perspective of the harmfully selfish, that type of learning is “not an improvement”. They see it as a threat.