We just learned something quite distressing – that one in 10 videos out in the online jungle we call the internet – can contain something potentially ‘harmful’ and ‘damaging’.

What we mean by this is that some videos contain certain elements that may not be accurately reflected by the title or tags associated with it. This poses a problem for brands who may not want to be associated with adult themes or extreme violence.

To find out what this means for brands and what options they have, we spoke to Kartik Mehta, Chief Revenue Officer, SilverPush. With their new product Mirrors Safe, the brand offers an AI-Powered context-relevant brand suitability platform to help prevent unwanted associations.

In order to build this product and understand they extent of the issue, Silverpush reviewed 15 million videos across the largest video hosting and sharing platforms in the SEA region using Mirrors Safe. With one in 10 or 10% containing images or negative associations (according to Silverpush criteria), there is definitely a need for a solution.

Visual Breakdown of Unsafe Content. Image courtesy of Silverpush

Kartik was kind enough to spare some time to educate us on the matter.

Congrats on the launch of Mirrors Safe. How do you see this new product helping brands and their advertisements?

Brands today are faced with different types of risks – financial risks, legal risks, and I guess the most important part of it is the reputation risk which could have larger concerns and probably a long-lasting impact on the overall brand equity. The existing brand safety measures like blocklists and whitelists are primarily focused on the principle of exclusion, which does protect brands to an extent but can also lead to one of the most pressing brand safety related concerns of over-blocking. Which can lead to brands missing the opportunity of engaging with the audiences across the right kind of content. 

Whereas Mirrors Safe’s computer vision powered in-video context detection identifies faces, actions, scenes, emotions, on-screen-text in a streaming video to detect content that features violence, smoking, nudity, arms & guns and more. It detects these contexts only when they feature in a video, and not just by relying on keywords used to describe the video – which often times are misleading and can result both in unsafe placements as well as over blocking. 

For instance, a video featuring smoking or violence might not be described so in its title, description or meta tags. There is no way for keyword-based solutions to identify these damaging contexts to filter out this video. Which can lead to household brands advertising across content which is highly unsuitable for their brand image. On the other hand, keywords like shoot, kill, crash, and even gun (some of the most blocked keywords) can easily be used within perfectly safe contexts, (like movies and songs). 

Moreover, Mirrors Safe’s context detection makes it possible to offer brand suitability that can be customized for each brand or each category without following the blanket exclusion principles. 

According to your research, 1 in 10 videos are deemed to be associated with dangerous or damaging content. How were you able to analyze around 15 million videos to generate this data?

Silverpush churned approximately 15 million videos across the largest video hosting and sharing platforms in the SEA region using Mirrors Safe. We used a randomly chosen inventory across platforms from our existing database – previously used to run campaigns using our video advertising platform Mirrors.  

It was found that nearly 8-9% of analyzed content to be deemed brand unsafe. This means these videos featured one or more unsafe contexts like nudity, smoking, violence, arms and guns, and more. 

Out of the unsafe content, the split for top unsafe categories include:

Context Percentage Split % Sub contexts include 
Violence 54Crime, guns, shooting, killing, bloodshed, murder, death and more
18+35Pornography, nudity, sex, and more
Extremist11Terrorism, hate speech, ISIS, and more
Smoking15Cigarettes and other similar substances, Smoking as an action…
The context detection is not mutually exclusive, i.e. a single video can have multiple damaging contexts

However, a bigger discovery was the difference between the results found by exclusion through traditional methods like keyword lists vs. Mirror Safe’s in-video context detection technology.

For instance, when we used both methods to identify unsafe videos for one of the top brand unsafe categories – nudity and adult content, Mirrors Safe (through its frame-by-frame parsing) identified 300% more video content featuring unsafe context in this category, compared to exclusion through keyword lists.

A single damaging ad placement can harm brand perception in the consumer’s mind. This discovery highlights the potential harm that existing traditional measures are unable to detect. This has been witnessed time and time again, with some of the largest video advertising platforms being unable to keep brands safe from damaging content.

Have you been able to measure or estimate the negative impact of these associations for the brand? 

There is already a plethora of information available on how even a single ad placement across harmful content can irreparably damage brand perception for a long time in the consumers mind. A 2019 study Trustworthy Accountability Group & Brand Safety Institute found that 80% consumers will stop or reduce buying products advertised against extreme or violent content. And, 70% believe advertiser and the agency are most responsible for a brand’s ad placements.

Dashboard location filters. Image courtesy of Silverpush

With our platform Mirrors, we have been serving contextually targeted video advertising across platforms since 2018. And we identified the challenge posed by traditional brand safety measures while serving our clients. We realized that ensuring brand safety is even more of a challenge across video formats, as NLP based technologies that work for other formats are ineffective in gauging the right context featured in video content. 

Conversations and feedback from partners first led us to introduce a safety feature in Mirrors, where brands working with us did not report a single unsafe exposure since the launch of the feature. This further led us to launch Mirrors Safe, which can be deployed as a standalone context relevant suitability platform. 

How has this solution helped brands so far? Do you have any initial test cases or beta usage that you can share?

I will start with the most interesting use case, that is highly relevant today.  

Helped brands navigate the extreme over-blocking of COVID-19 related content 

As brands and platforms rapidly add terms associated with COVID-19 to their keyword block, Coronavirus has become one of the most blocked keywords today. We have found advertisers looking to avoid unsafe brand exposure around this sensitive topic are forced to exclude news entirely from their list of targeted channels and publishers. However, excluding news and related channels entirely from advertising strategies across platforms is killing reach for brands. 

One of the key factors behind extreme COVID-19 related over-blocking is the inability to detect if the COVID-19 related stories are informative Vs. stories that can harm brands – leading to blanket exclusions. Mirrors Safe identified what the video content features to differentiate the stories, in the following ways: 

  • On-screen text recognition: Mirrors Safe identifies and filters out videos that have related text written on the screen (e.g. Coronavirus or COVID-19). And can help differentiate between morbidity related stories Vs. more positive stories around for instance precautions. 
  • Object and action detection: the system can identify objects like masks, stretchers, and actions like coughing and sneezing and understand the concentration of this content within a single video through frame by frame parsing.
  • Faces: with this outbreak certain public figures are also on brands’ blocklists (yes, Trump). In-video context detection can accurately identify faces to filter out related content.
Image courtesy of Silverpush

Context Relevant Brand Suitability 

Our initial test campaigns solidified Mirrors Safe’s ability to offer brand suitability in its truest form. For instance, a household consumer goods brand was categorically looking to exclude racy content like a car chase or extreme sports. In its past campaigns, using keyword blocking, the brand had found it very difficult to exclude only videos that feature these contexts, which led to over blocking on certain types of content (like movie clips and songs). With Mirrors Safe, we curated a brand suitable inventory that only excluded content that featured the said context without applying blanket exclusions to any particular type of content. 

Conversely, a cola brand specifically sought this very same category of content that featured extreme sports, and we were able successfully customize the exclusion and inclusion lists strictly basis what the brand values are and what is the current need. 

Campaign Performance and Auditing 

Mirrors safe also helped brands measure performance of their past campaigns. Brands used the platform to identify unsafe exposure they might have encountered in the past. Mirrors Safe provides a comprehensive exclusion percentage that tells how much of the total placements, percentage wise, were unsafe. Not only this, the platform detects the exact content across which the unsafe placement occurred and even indexes the instance when the unsafe content features in the video. Offering full transparency or it’s post campaign analysis. No other platform can provide this analysis with such accuracy and detail. 

What are your plans for Mirrors Safe in Southeast Asia? How do you see brands adopt this solution?  

Compared to the western counterparts, where in-video context detection might not be still used for brand safety and suitability, however there are more advanced NLP driven solutions available. However, I see that brands and marketers in Southeast Asia are still vastly using the traditional way of keyword blocking and probably making use of some meta data to block content. 

We already have a strong presence across 7 markets in SEA (Vietnam, Thailand, Philippines, Malaysia, Singapore, Indonesia, and India), and we are known for bringing disruptive solutions that furthering our video intelligence story. We now want to extend the value of context relevance to brand safety – having this solution now available to marketeers in SEA could definitely expedite their journey in terms of taking a leap from traditional brand safety and moving on to what we call brand safety 2.0 or brand suitability.   

Where do you see the programmatic advertising industry evolve in Southeast Asia? What are the new trends that are likely to happen?

One area I strongly feel will be impacted by programmatic is advertising on connected TV (CTV). TV viewing has fundamentally changed with more and more consumers opting for streaming and VOD services over linear broadcast channels. However, this is not the end of advertising on TV, rather an opportunity for bringing in the benefits of programmatic to TV, which is now possible because of CTV.  

However, the shortcomings of programmatic that have been allowed on digital channels won’t necessarily fly in the CTV and VOD world. Consumers expect more from their TV viewing experience, and are likely to be disappointed by poor ad experiences. This is where I feel contextual technologies like Mirrors + programmatic will help advertisers capitalize on consumers’ current state of mind and in serving more meaningful experiences. 

Another area where programmatic can help in delivering contextual messages at scale is DOOH media. For instance, programmatic can help dynamically deliver a creative that is relevant to that time and location, among other contextual moments that can be served through programmatic. 

Advertisements