Assessing the risks of ‘infodemics’ in response to COVID-19 epidemics

May combine two claims.

When an epidemic reaches an area, reliable information spreads rapidly and attention shifts to certified sources.

“When the epidemics hit the same area, reliable information is quickly inoculated, like antibodies, and the system shifts focus towards certified informational sources.”

May combine two claims.

An estimated 40.4% of online messages in the period came from automated agents, twice the estimate from four years earlier.

“Specifically, we estimate that 40.4% of online messages during this period were due to such automated agents, doubling the activity with respect to estimates of only four years ago 35 .”

• cites [35] Watts, D. J. A simple model of global cascades on random networks. Proc. Natl Acad. Sci. USA 99, 5766–5771 (2002). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Average follower counts were 660 for unverified humans, 1,400 for unverified bots, 51,000 for verified humans, and 240,000 for verified bots.

“The average values are: 660 for Unverified Humans (circles), 1400 for Unverified Bots (squares, 51k for Verified Humans (stars) and 240k for Verified Bots (diamonds).”

May combine two claims.

The study found 4,988 domains, reduced to 4,417 after removing hard duplicates across databases.

“We have found a total of 4,988 domains, reduced to 4,417 after removing hard duplicates across databases.”

May combine two claims.

In South Korea, infodemic risk decreased as contagion spread, indicating growing public focus on reliable news.

“As the contagion spreads to significant levels, the infodemic risk further decreases, signalling an increasing focus of the public opinion toward reliable news sources.”

May combine two claims.

The probability distributions of follower counts for the four user classes all display a fat tail, but different categories have different outreach.

“All distributions display a fat-tail, but different categories of users have a different outreach.”

May combine two claims.

About 0.84% of collected posts were geo-tagged by users, giving highly accurate location information.

“About 0.84% of collected posts were geo-tagged by the user, providing highly accurate information about their geographic location.”

May combine two claims.

Canada had slightly higher infodemic risk than South Korea; risk rose as epidemics spread but remained low.

“Canada presents a slightly higher level of infodemic risk, and unlike South Korea, we see that the risk level increases as the epidemics spread, but stays at low levels.”

May combine two claims.

The geolocated data used in the paper slightly overestimated the fraction of unreliable news for unverified accounts and underestimated the fraction of fake news for verified profiles compared with the whole sample of Tweets.

“The difference, possibly caused by the correlation of this kind of behaviour with the choice of how customising the field”

May combine two claims.

In Venezuela, infodemics was fully active throughout observation, and verified sources contributed substantially to it.

“At the opposite, in a high-risk country such as Venezuela, the infodemics is in full swing throughout the period of observation, and in addition to the expected activity from unverified sources one notices that even verified ones contribute to a large extent to the infodemics.”

May combine two claims.

The relationship between declining infodemic risk and expanding epidemics appears general, with a nonlinearly negative association between confirmed cases and infodemic risk.

“Fig. 4 shows, however, that the relationship between reduction of infodemic risk and expansion of the epidemics seems to be a rather general trend, as the relationship between number of confirmed cases and infodemic risk is (nonlinearly) negative, confirming the result shown in Fig. 3.”

May combine two claims.

Iran, the United States, Germany, the Netherlands, Austria, and Norway maintained their relative infodemic risk levels, while Italy, South Korea, and Japan significantly reduced it as the epidemics progressed.

“with major roles played not only by countries such as Iran, but also United States, Germany, the Netherlands, Austria and Norway maintaining their relative levels, and other countries like Italy, South Korea and Japan significantly reducing it with the progression of the epidemics.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Processing reduced the overall database to 3,920 unique domains.

“This phase of processing reduced the overall database to unique 3,920 domains.”

May combine two claims.

Of the 3,892 entries, 1,434 were reliable, 2,264 were unreliable, and 194 were unknown.

“Type Count RELIABLE 1434 UNRELIABLE 2264 UNKNOWN 194 Total 3892”

May combine two claims.

In the USA, misinformation was mainly driven by verified sources and remained basically constant after the epidemic outbreak.

“For the USA, misinformation is mainly driven by verified sources, and it remains basically constant even after the epidemics outbreak.”

The quote only partly states this claim.

Reporting the first confirmed COVID-19 case in a country was associated with a drop in the infodemic risk index, followed by a return to the natural level at three confirmed cases.

“We observe that reporting of the first case is associated with a drop in the risk index, followed by an increase back to the”

May combine two claims.

Collective attention about COVID-19 first increased after the outbreak in Wuhan between 24 Jan and 02 Feb 2020, and rose strongly again after the epidemics spread in northern Italy from 20 Feb 2020 onward.

“We can observe a first increase in collective attention after the outbreak in Wuhan, China (between 24 Jan and 02 Feb 2020) and a second strong rise after the epidemics began to spread in northern Italy (20 Feb 2020 onwards).”

May combine two claims.

In Italy, infodemic risk was mostly from unverified sources, but misinformation collapsed and shifted suddenly to reliable sources after the epidemic outbreak.

“In the case of Italy the risk is mostly due to the activity of unverified sources, but we notice that with the outbreak of the epidemics, the production of misinformation literally collapses and there is a sudden shift to reliable sources.”

May combine two claims.

There are G9 countries with remarkable infodemic risk and developing countries with far lower risk levels.

“In fact, we find that there are G9 countries with remarkable infodemic risk and developing countries with far lower risk levels.”

May combine two claims.

In the COVID-19 crisis, complex infodemic effects are at work with significant cross-country variation, and socio-economic development is not the key discriminant between high and low infodemic risk.

“Here, we have shown that in the context of the COVID-19 crisis, complex infodemic effects are indeed at work, with significant variations across countries, where level of socio-economic development is not the key discriminant to separate countries with high vs. low infodemic risk.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The fraction of geolocated messages was constantly about 56% of the total volume recorded.

“The fraction of Geoolocated (messages with shared locations, or geonamed, indicated in green) is constantly about 56% of the total volume recorded”

May combine two claims.

In Russia, infodemic risk was erratic with sudden pronounced spikes, verified sources played a major role, and epidemic information was fragmented and mostly unreliable.

“Finally, in a relatively high-risk country such as Russia we notice that infodemic risk is erratic with sudden, very pronounced spikes, and again also verified sources play a major role.Here too, information about the epidemics is fragmented and mostly unreliable.”

Probably not a substantial claim: may not pass S3 (load-bearing).

After filtering poorly defined domains, the final database contained 3,892 entries.

“This action reduced the database to the final number of 3,892 entries”

May combine two claims.

About half of extracted URLs were fact-checked using nearly 4,000 expert-curated web domains; the rest pointed to vanished pages or content not automatically classifiable.

“About 50% of URLs have been fact-checked by screening almost 4,000 expert-curated web domains, whereas the remaining corpus was pointing to disappeared web pages or to content not classifiable automatically (eg, videos on YouTube) and unpopular sources.”

• AssumptionRests on an assumption: The reliability of a URL can be inherited from the reliability of its source domain.

Probably not a substantial claim: may not pass S4 (based).

Repeated exposure to a message from different sources acts as indirect validation of its reliability and relevance, leading users to spread it and become informationally infectious.

“As users are repeatedly hit by a given message from different sources, this works as an indirect validation of its reliability and relevance, leading the user to spread it in turn, and to become an informationally infectious agent.”

May combine two claims.

Human response to falsehood shows early-warning signals that adequate communication strategies might mitigate.

“Contrary to mainstream beliefs, we show that human response to falsehood exhibits early-warning signals that might be mitigated with adequate communication strategies.”

Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing).

It has been proposed that at population level, dynamics of multiple interacting contagions are indistinguishable from social reinforcement.

“Recently, it has been suggested that this challenging phenomenon might be due to the fact that, at population level, the dynamics of multiple interacting contagions are indistinguishable from social reinforcement 38 .”

• cites [38] Stella, M., Cristoforetti, M. & De Domenico, M. Influence of augmented humans in online interactions during voting events. PLoS ONE 14, e0214210 (2019). sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Information dynamics intended to change perceptions can direct collective attention to false or inflammatory content, a phenomenon called infodemics that resembles epidemics.

“In this study, we show that information dynamics tailored to alter individuals' perceptions and, consequently, their behavioral response, is able to drive collective attention 12 towards false 13,14 or inflammatory 15 content, a phenomenon named infodemics [16][17][18][19] , sharing similarities with more traditional epidemics and spreading phenomena [20][21][22] .”

• cites [12] Zarocostas, J. How to fight an infodemic. Lancet 395, 676 (2020). sourceNot checked yet

  Not checked in the free run.

• cites [13] Pastor-Satorras, R., Castellano, C., Van Mieghem, P. & Vespignani, A. Epidemic processes in complex networks. Rev. Mod. Phys. 87, 925–979 (2015). sourceNot checked yet

  Not checked in the free run.

• cites [14] De Domenico, M., Granell, C., Porter, M. A. & Arenas, A. The physics of spreading processes in multilayer networks. Nat. Phys. 12, 901–906 (2016). sourceNot checked yet

  Not checked in the free run.

• cites [15] Brockmann, D. & Helbing, D. The hidden geometry of complex, network-driven contagion phenomena. Science 342, 1337–1342 (2013). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based).

Infodemics arise from simultaneous activity of multiple human and non-human sources of fake or unreliable news.

“As shown in Fig. 1, an infodemics is the result of the simultaneous action of multiple human and non-human sources of fake or unreliable news.”

May combine two claims.

The danger posed by infodemics can be comparable to and additive with that of the epidemic itself.

“In fact, the dangerousness of infodemics can compare and sum up to a large extent to that of the epidemics itself 29 .”

• cites [29] Ferrara, E., Varol, O., Davis, C., Menczer, F. & Flammini, A. The rise of social bots. Commun. ACM 59, 96–104 (2016). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

Perceived and actual biological and social risks feed upon one another and may co-evolve in complex ways.

“Perceived and actual biological and social risks feed upon one another, and may co-evolve in complex ways.”

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Human and non-human accounts create unreliable COVID-19 content, and their followers are exposed to it and reshare it, becoming infectious agents.

“Human and non-human accounts forge unreliable content -such as fake or untrustworthy news -about the COVID19, a topic attracting the collective attention of the whole world.Their followers are exposed to such content, and reshare it, becoming infectious agents:”

May combine two claims.

The method has the highest accuracy (>90%) and precision (>95%) in identifying bots compared with state-of-the-art methods.

“More in detail, our method has the highest accuracy (>90%) and precision in identifying bots (>95%) when compared with state-of-the-art methods.”

May combine two claims.

The global decrease in infodemic risk suggests that contagion expansion led people to seek more reliable sources and that verified influencers began spreading more reliable news.

“suggesting that evidence of the expansion of the contagion leads people to look for relatively more reliable sources, and that verified influencers with many followers started inoculating the system with more reliable news (see Supplementary Figures 3 and 4), playing a role that presents interesting analogies to that of antibodies in the treatment of an infectious disease.”

Classified as probably not a substantial claim.

When effective therapies are unavailable, coordination of behaviors and diffusion of pro-social orientations driven by reliable information are key resources for mitigating adverse effects.

“Especially in situations where effective therapies to contrast the diffusion of the pandemic are not readily available, coordination of behaviors and diffusion of pro-social orientations driven by reliable information at all scales are the key resources for the mitigation of adverse effects.”

May combine two claims.

Until 24 Feb 2020, about 60,000 tweets were lost out of millions, capturing more than 99.5% of all posted messages.

“We estimate that until 24 Feb 2020 we lost about 60,000 tweets out of millions, capturing more than 99.5% of all messages posted (see Supplementary Fig. 1).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Countries with high infodemic risk might also report epidemic data less reliably, altering perceptions and misleading people seeking reliable information.

“It is also interesting to observe though that countries with high infodemic risk might also be more unreliable in terms of reporting of epidemic data, thus altering the perceptions of people and indirectly misleading them in their search for reliable information.”

May combine two claims.

Escalating epidemics lead people to pay more attention to reliable sources, potentially limiting infodemics, but the speed of adjustment may strongly affect social outcomes.

“The escalation of the epidemics leads people to progressively pay attention to more reliable sources thus potentially limiting the impact of the infodemics, but the actual speed of adjustment may make a major difference in determining the social outcome, and in particular between a controlled epidemics and a global pandemics.”

May combine two claims.

The Infodemic Risk Index is robust to user classification and is not sensitive to the performance of bot detection algorithms.

“IRI is robust to user classification, making it an indicator not sensitive to performances of bot detection algorithms.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The reliability classification used manually checked web domains collected from multiple publicly available databases.

“We have collected manually-checked web domains from multiple publicly available databases, including scientific and journalistic ones.”

May combine two claims.

The choice of Twitter API was the safest to date because biases affecting the Sample API are not found in REST and Filter APIs.

“Note that this choice is the safest as to date: in fact, it has been recently shown that biases affecting Sample API (which samples data based on rate limits), for instance, are not found in REST and Filter APIs 40 .”

• cites [40] Eysenbach, G. How to fight an infodemic: the four pillars of infodemic management. J. Med. Internet Res. 22, e21820 (2020). sourceNot checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

The findings cast new light on the social mechanics of infodemics-epidemics interaction and may help policy makers design a more integrated strategic approach.

“This casts new light on the social mechanics of the infodemics-epidemics interaction, and may be of help to policy makers to design a more integrated strategic approach, by suitably embedding communication and information management into a comprehensive, extended public health perspective.”

May combine two claims.

An estimated 161.2 million tweets were posted until 10 Mar 2020, and 112.6 million were collected.

“We have estimated a total of 161.2 millions tweets posted until 10 Mar 2020: we have successfully collected 112.6 millions of them, providing an unprecedented opportunity for infodemics analysis.”

May combine two claims.

Reliable and unreliable sources circulate in different ways and reach different targets.

“however, they circulate in different ways and reach different targets, a feature that is perfectly captured by the infodemic risk index introduced in this study.”

May combine two claims.

The decrease in infodemic risk was especially pronounced as epidemics escalated, suggesting mediation by perceived social alarm.

“The effect is particularly pronounced with the escalation of the epidemics, suggesting that this effect could be mediated by levels of perceived social alarm.”

Classified as probably not a substantial claim.

When two or more domains are soft duplicates, the classification with the highest Harm Score is kept as a conservative choice.

“When two or more domains are soft duplicates, we keep the classification with the highest Harm Score, as a conservative choice.”