Trustworthiness of YouTube Recommendations: Artificial Intelligence and Bias

by Abhishek Ganguly

Artificial intelligence (“AI”) has developed to an extent where it is now seen as a viable alternative to human labour and further, as an alternative for profit maximisation by corporations. For AI to work effectively, personalisation of content becomes imperative. This requires companies to collect and analyse vast amounts of data from their consumers. YouTube’s algorithm, which uses consumer data to provide accurate recommendations, is a perfect example to explain the ethical implications of AI technology usage and the consequent need for regulation. This paper seeks to discuss the ethics and trustworthiness of the YouTube algorithm for user specific recommendations.

Even though AI technology is employed in various areas such as social media sites and Over-the-top (“OTT”) media platforms, one must never forget that AI cannot remain stagnant. It will only become smarter and quicker as a result of functional learning. AI will continue to develop and change as the availability of data improves. As a result, we may reach a situation where AI takes over tasks which were originally tasks of the programmer. The development of Google Brain, which employs unsupervised learning to provide recommendations which programmers could not have guessed themselves, signifies this shift towards smarter AI. Hence, YouTube’s decision to replace Sibyl, an algorithm limited by machine learning, with Google Brain appears to be a natural economic choice.

Smarter AI, though profit maximising, may not guarantee ethical behaviour. YouTube’s ability to provide recommendations based on initial search requests must be seen as both a positive as well as a negative development. While it allows consumers to view curated content, it also points towards erroneous judgements that may have been developed by AI autonomously.

Google Brain enhances the consumer’s browsing experience by allowing YouTube to process search requests and identifying relations between the consumer’s preferences in order to recommend similar videos. But these recommendations may be seemingly flawed as in certain instances, such recommendations may suggest videos contrary to viewers preferences. As stated by Guillaume Chaslot, a former Google software engineer, the AI algorithm is built not to assist consumers but to get them addicted to YouTube. For instance, even though the viewer may watch videos on climate change and global warming, YouTube may recommend conspiracy theories stating that the earth is flat. Such theories are not factual but speculative theories based on inconclusive evidence.

While it is at the consumer’s risk to choose to watch the recommended video or not, it is easy to see how the algorithm reels in the consumer to continue watching more videos – this fulfils the algorithms purpose to keep viewers addicted to the platform for longer durations but does not improve the viewers experience. Even Jim McFadden, the technical lead for YouTube recommendations claimed that Google Brain assisted YouTube to solely generalise. A mere generalisation, while simplifying the task of the algorithm, completely disregards the diversity of viewer choices, thus impacting personalised results. Should one trust algorithms that disrespect individual autonomy and promote third-party interests?

Further, YouTube employed Google Brain not just to improve personalisation, but also to gather data on the duration a viewer spends on such videos, clearly, YouTube would have longer viewership only if successive recommendations were actually accurate. A limitation of AI is that it only acts based on the information it receives – it may recommend videos that may be related to a consumer’s past searches; however, such recommendations, a result of improper processing of data, may have no logical link. For instance, on an initial search for videos on the African-American community in America, recommendations may include videos on crime rates in America – there may be a correlation, insofar as it limits itself to a particular geography, but in the absence of an established causation. In addition, such conclusions inadvertently made by the algorithm can prove disastrous for the community. Such rampant correlations, and the spread of such information owing to the wide reach of social media, have significantly reduced prospects for the African-American community. As discussed later, some jurisdictions in the United States employ AI to decide jail sentences for the accused – the above stated correlations have resulted in African-American convicts being sent to prisons for longer durations for petty crimes, simply because the algorithm informs the judge that the accused may be a repeat offender.

An ethical question YouTube recommendations’ raise is that of profiling, that is, the ability to identify psychological and behavioural characteristics of individuals based on their preferences on social media and platforms such as YouTube. Recommendations are based on the data points fed into it by the consumer based on the genre of videos he or she may view. A collection of such data, that is rarely infallible, allows AI to ascertain sufficient information about that individual. With the absence of any regulation and collection of data on thousands of individuals per minute, there exists a fear as to the extent of profiling that is possible. Unregulated collection of data raises problematic privacy concerns. One is unaware if such data-points on each of YouTube’s viewers are shared with other social media platforms – YouTube may even sell this information to third-parties to generate revenue through targeted advertisements.

While in the United States, AI is currently employed in the judiciary to assist judges in delivering sentences based on the likelihood of the accused to be a repeat offender, even the executive, through the police, may employ such technology to gather more information on individuals – on one hand, such information can be beneficial for improving the reach and impact of government services, however, the executive machinery may also target particular individuals or communities, thereby impinging on their privacy. YouTube’s algorithms, as well as algorithms of social media and OTT platforms may be ideal for such information.

YouTube’s algorithms have also been documented to have assisted in formulating public opinion in favour of a particular candidate in the 2016 Presidential elections in the United States. Guillame Chaslot’s software showed that a bias existed in the algorithm particularly for videos on Hillary Clinton. At least 88% of the videos on Hillary Clinton’s campaign had an ‘Up Next’ recommendation favourable to Donald Trump. This was higher than the recommendations to the Trump campaign post initial searches on Trump. An assessment of the content of these videos showed that they were overwhelmingly supportive of Donald Trump and further that the content available on Hillary Clinton was characterised by conspiracy theories and false information – a trend, as discussed above is only beneficial to the platform. This instance makes it clear that AI, if encouraged liberally and without strict ethical guidelines, it can be dangerous for a democratic society.

In conclusion, it must be noted that algorithms are invariably value laden. To avoid explicit biases in the use of technology, emphasis must be placed on the need to design data systems that promote fairness and safeguard against discrimination. It is clear that though AI improves productivity, it also has social costs. Since AI is employed owing to the lack of more reliable means and high costs, the conclusions suffer from several limitations. Results can only be as reliable as the data they are based on. Inherent bias in technology such as YouTube’s algorithms emphasise on the need to necessarily reject the understanding that all results produced by AI algorithms are value-neutral and also that they cannot be erroneous.


1. Brent Daniel Mittelstadt, Patrick Allo et al, The ethics of algorithms: Mapping the debate, Big Data & Society, (2016), pp.1-21, available at, (last accessed on 26 May,2020).

2. Big Data: A Report on Algorithmic Systems, Opportunity and Civil Rights, Executive Office of the President, (2016), available at, (last accessed on 26 May, 2020).

3. Julia Angwin, Jeff Larson et al, Machine Bias: There’s software used across the country to predict future criminals. And it’s biased against blacks, ProPublica, (2016), available at (last accessed on 26 May, 2020).

4. Casey Newton, How YouTube Perfected the Feed, The Verge, (2017), available at (last accessed on 27 May, 2020).

5. VICE News, How YouTube Algorithm Could Prioritize Conspiracy Theories (HBO), YouTube, 5 March, (2018), available at (last accessed on 11 May, 2020).

6. Micheal Spencer, Artificial Intelligence Regulation May be Impossible, Forbes, 2 March 2019, available at (last accessed on 27 May, 2020).

7. Paul Lewis, Eric McCormick, How an ex-YouTube insider investigated its secret algorithm, The Guardian, 2 February 2018, available at (last accessed on 28 May, 2020).

About the Author

Abhishek is a final year B.A., LL.B. (Hons.) student from the Jindal Global Law School. He is interested in pursuing corporate law and finds that AI cannot be excluded from the work undertaken by a corporate lawyer. Having attended a course seeking to understand the legal framework regulating AI, he is interested in analysing the core issues such as: the interrelation between AI and intellectual property; the debate surrounding personhood for autonomous technology; and the ethical concerns on the use of AI in everyday activities such as entering into contracts.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s