AIBRT Political Opinion Study
Total number of study participants as of
2:47 pm GMT, October 31, 2014:
2096


DEMOCRACY AT RISK:
HOW VOTERS IN THE 2014 ELECTIONS IN INDIA WERE
MANIPULATED BY BIASED SEARCH RANKINGS

Media Release Available Here

Below are data collected between April 2nd and May 12th, 2014, in an experiment conducted by a nonprofit, nonpartisan research institute in Vista, California, USA (see http://AIBRT.org). In the experiment, researchers deliberately manipulated the voting preferences of undecided voters in the national Lok Sabha election in India, the largest democratic election in history, with over 800 million eligible voters.

They did this by randomly assigning undecided voters who had not yet voted (recruited through print advertisements, online advertisements, and online subject pools) to one of three groups in which search rankings favoured either Mr Gandhi, Mr Kejriwal, or Mr Modi. About 2,000 eligible voters from 26 of India's 28 states (age range 18 to 70, mean age 29.5) participated in the study - not enough to affect the election's outcome. People’s preferences were also pushed equally toward all three candidates, so there was no overall bias in the study.

Previous research, presented at the 2013 meeting of the Association for Psychological Science in Washington, DC, has demonstrated the enormous power that biased search rankings have to sway voting preferences, presumably because people put inordinate trust in higher-ranked search results, as has been demonstrated by extensive research on consumer behavior.

In laboratory and online experiments, biased search rankings were shown to sway the voting preferences of undecided voters by 15% or more, with little or no awareness by participants that they were being manipulated. The researchers call this manipulation SEME (pronounced "seem"), for Search Engine Manipulation Effect. A technical report on this research can be accessed here.

The numbers (in red) displayed below were updated every 15 minutes over the course of the study. Some minor adjustments might still be made as the data are examined by the researchers for irregularities. (People who participated more than once will be eliminated, for example.) The results of the study are organized below into three major categories:

1. Social Impact Measures - Three numerical indicators of the possible impact of manipulating election-related search rankings on a large scale: VMP (Vote Manipulation Power), WMP (Win Margin Power), and PMP (Psychological Manipulation Power). The larger these numbers, the more likely it is that biased search rankings can be used to tilt the outcomes of elections worldwide.

2. Candidate Data - Statistics showing how voting preferences were altered for each candidate by exposure to biased search rankings. NOTE: The percentage shifts toward Mr Modi and Mr Kejriwal were smaller than the percentage shift for Mr Gandhi because the initial preferences for Mr Modi and Mr Kejriwal were very high. When an initial preference is high, only a relatively small percentage increase is possible.

3. Demographic Data – Selected statistics showing which demographic groups in India appear to be most vulnerable to SEME (the Search Engine Manipulation Effect).

NOTE: Although 2,096 people participated in the study, data are shown only for the 1,908 people who spent at least 10 seconds searching. On average, these people spent 5.1 minutes in search out of 15 minutes that were allotted.

Tentative conclusions and recommendations: This study confirms the findings of previous ones, showing the enormous power that biased search rankings have to alter the voting preferences of undecided voters. Previous studies used real search results and real webpages from a past election, whereas the present study used real search results, real webpages, and real voters in the midst of a real campaign and election process.

The results suggest that no matter how search rankings end up favouring one candidate - either because of election fixing by a search engine company or by impersonal decisions made by a search algorithm - rankings favouring one candidate drive about 12% of the votes of undecided voters toward that candidate - double that amount in some demographic groups.

This is enough to flip elections won by margins up to 2.9%. Worldwide, upwards of 25% of national elections are won by margins under 3%. If biased rankings were presented repeatedly over a period of weeks or months, or if they were targeted toward particularly vulnerable groups or individuals, their impact would be even greater. As more people get internet access worldwide, the power of search rankings to determine the outcomes of elections will grow well beyond the current level.

With so many election outcomes determined by a single mechanism (search ranking manipulation) which for all practical purposes is in the hands of personnel at a single company, a key element in the democratic system of government - the free and fair election - is threatened. An "equal-time" law, as well as strict regulation and monitoring of politically-related search rankings, might be necessary to protect the electorate from undue influence.

Inquiries regarding this study should be directed to Mr Ronald Robertson at rrobertson@AIBRT.org. To schedule an interview with the principal investigator, Dr Robert Epstein, contact lauracraig@outinfrontpublicity.com.







SOCIAL IMPACT MEASURES

1. VMP: Vote Manipulation Power
Percentage increase in potential votes for a candidate after people have been exposed to search rankings favouring that candidate:

12.2

This is the percentage increase in the number of people in the study who said they would vote for a particular candidate after being exposed to search rankings favouring that candidate. Even a low percentage increase, such as 4.0 (in other words, 4,000 people per 100,000), could have a significant impact on a close election.


2. WMP: Win Margin Power
Threshold win margin below which control of an election might be guaranteed:

2.9

The value above shows the win margin below which search-ranking manipulation might guarantee the outcome of a two-person race (according to the results of the present study). Many elections are won by small margins; worldwide, upwards of 25% of national elections are won by margins under 3%. A WMP value of 1.0 means that an election in which the winner is expected to win by no more than 1% can with reasonable certainty be tilted through the use of biased search rankings. That threshold is reached with a VMP of 4.2. To put this more concretely, in a million-voter election, you can probably control the outcome of an election with an expected 1% win margin by flipping the votes of only 4.2% of undecided internet users. This becomes easier if biased rankings are presented repeatedly over time without people knowing they are being manipulated. The WMP estimate above is based on tentative but conservative assumptions currently applicable to most developed nations, namely (a) that 40% of eligible voters have internet access and (b) that the opinions of 30% of those individuals are at some point amenable to change (the latter percentage will presumably get smaller as election day approaches) – in other words, that 12% of the total voting population consists of undecided voters who have internet access (30% of 40%). This percentage will vary from one election to another and one culture to another; it is almost certain, however, to grow steadily in coming years as an increasingly larger proportion of the electorate worldwide gets election-related information through the internet. A study conducted in India in October 2013 suggested that 42% of urban voters were undecided at that time and that 37% of urban voters were regular internet users. Recent AIBRT studies suggest that 86% of American internet users have employed the internet to obtain information about political candidates and that, overall, 52% of American voters have used the internet this way.













3. PMP: Psychological Manipulation Power
Percentage of participants who were apparently unaware of the manipulation:

99.3

The value above shows the percentage of participants in the present study who appear to have had no awareness that the researchers were showing them highly biased search rankings - in other words, who appear to have had no idea they were being manipulated. The higher the percentage, the more easily one can perform this kind of manipulation without being detected. A percentage of 100 means no one detected the manipulation. A percentage of 0 means everyone detected the manipulation. By biasing search rankings for only a select, highly vulnerable group of voters (based on information previously collected about those voters) - in other words, by sending people customized search rankings - the manipulation would be even harder to detect because regulators would have no way to reproduce it. Previous research has shown that search-ranking bias can also be masked in various ways, making it even more difficult to detect. See the technical report available here.





CANDIDATE DATA

Below are shown the percentage increases in the number of intended votes for a candidate after exposure to search rankings that favoured that candidate. All participants had access to exactly the same web pages, but people in the three groups (to which they were assigned randomly) saw the search rankings in different orders favouring either Mr Gandhi, Mr Kejriwal, or Mr Modi.






When people were presented with search rankings favouring
Mr Gandhi, the number of people voting for him increased by...
28.3%



In addition,
28.1% of the people in this group now said they were more likely to vote for Mr Gandhi;
22.3% said they were less likely to vote for Mr Kejriwal; and
19.5% said they were less likely to vote for Mr Modi.



When people were presented with search rankings favouring
Mr Kejriwal, the number of people voting for him increased by...
12.0%



In addition,
29.2% of the people in this group now said they were more likely to vote for Mr Kejriwal;
24.6% said they were less likely to vote for Mr Gandhi; and
22.8% said they were less likely to vote for Mr Modi.



When people were presented with search rankings favouring
Mr Modi, the number of people voting for him increased by...
8.1%



In addition,
25.6% of the people in this group now said they were more likely to vote for Mr Modi;
25.2% said they were less likely to vote for Mr Gandhi; and
25.5% said they were less likely to vote for Mr Kejriwal.



Liking: In addition, biased search rankings changed the extent to which participants indicated they like the candidates by the following margins:

Mr Gandhi:
28.5%

Mr Kejriwal:
30.5%

Mr Modi:
25.8%



Trust: Biased search rankings also changed the extent to which participants indicated they trust the candidates by the following margins:

Mr Gandhi:
31.1%

Mr Kejriwal:
29.7%

Mr Modi:
25.8%






DEMOGRAPHIC DATA

Below are shown selected demographic groups listed in order from the most vulnerable to the least vulnerable to manipulation, based on VMP scores for each group (in red). A negative VMP might suggest an underdog effect for that group. The negative value for Residents of Gujarat, however, is a statistical anomaly; pre-search values in this group favoured Mr Modi (Chief Minister of Gujurat) so strongly that they could not be offset by rankings favouring Mr Kejriwal or Mr Gandhi. Additional statistical information is available from the researchers.








Resident of West Bengal: 33.3
Resident of Karnataka: 32.1
Resident of Kerala: 29.0
Resident of Delhi (NCT): 27.3
Secondary school education or less: 23.5
Females / Age 36 or over: 19.4
Not Employed 18.2
Resident of Himachal Pradesh: 17.4
Resident of Maharashtra: 17.2
Conducts few online searches (<5 / day): 16.3
Males / Age 35 or under: 15.9
Male: 15.3
Right Political View 14.8
Age 36 or over: 14.6
Centre Political View 14.2
Familiar with candidate: 14.0
Not familiar with candidate: 13.5
Pre-university or university education: 11.9
Age 35 or under: 11.8
Males / Age 36 or over: 11.7
Left Political View 11.4
Resident of Andhra Pradesh: 10.8
Conducts frequent online searches (10+ / day): 10.3
Employed 10.3
Resident of Uttar Pradesh: 10.0
Income 50,000 Rs or over: 9.9
No specific political views 7.4
Resident of Tamil Nadu: 7.2
Female: 5.7
Females / Age 35 or under: 2.9
Resident of Gujarat: -6.2

Focusing search-engine manipulation on particularly vulnerable groups – or, using customized rankings, on particularly vulnerable individuals (an increasingly common practice in the marketing of goods and services) – will greatly increase the power and efficiency of the manipulation (SEME). So will repeatedly sending the same people biased search rankings over a long period of time.

Inquiries regarding this ongoing study should be directed to Mr Ronald Robertson at rrobertson@AIBRT.org. To schedule an interview with the principal investigator, Dr Robert Epstein, contact lauracraig@outinfrontpublicity.com.