Brands, accurate facial recognition and why transparency is critical

Introducing accurate facial recognition into the mix potentially removes the need for you to tell Facebook (or a future Facebook connected site or app) who you are before your data is shared and your experience modified. All you will need to do now is show up and let a camera see you long enough to capture a reasonably clear image of your face. From there you will be identified, placed into a particular context and things will happen. As a brand, there are some interesting opportunities. Imagine your guests arrive at your event and, instead of relying on guests to manually check in, a webcam at the door connected to your Facebook Page recognises the guests as they arrive and posts an update in your stream sharing their arrival. This isn’t happening yet but it is very possible. 

Facebook’s new artificial intelligence group recently published a research paper titled “DeepFace: Closing the Gap to Human-Level Performance in Face Verification” which describes its advances in facial recognition technology. The abstract is pretty technical so I highlighted the big takeaway that may interest you:

In modern face recognition, the conventional pipeline consists of four stages: detect => align => represent => classify. We revisit both the alignment step and the representation step by employing explicit 3D face modeling in order to apply a piecewise affine transformation, and derive a face representation from a nine-layer deep neural network. This deep network involves more than 120 million parameters using several locally connected layers without weight sharing, rather than the standard convolutional layers. Thus we trained it on the largest facial dataset to-date, an identity labeled dataset of four million facial images belonging to more than 4,000 identities, where each identity has an average of over a thousand samples. The learned representations coupling the accurate model-based alignment with the large facial database generalize remarkably well to faces in unconstrained environments, even with a simple classifier. Our method reaches an accuracy of 97.25% on the Labeled Faces in the Wild (LFW) dataset, reducing the error of the current state of the art by more than 25%, closely approaching human-level performance.

According to the MIT Technology Review’s article titled “Facebook Creates Software That Matches Faces Almost as Well as You Do”, human beings recognise faces correctly 97.53% of the time which makes DeepFace just about as accurate as humans when it comes to identifying your face. What does this mean for brands? Quite a lot although probably not right away.

One of the service features that will continue to distinguish brands and their service offerings is a brand’s ability to present its customers with a deeply personal and meaningful service. Brands have been working on ways to personalise their services for quite some time and have used demographics, location, culture and, more recently (and as we have increasingly seen on Facebook and Google properties), your interests. All of this information is being associated with your identity so when you connect to a site or an app with your Facebook profile, for example, you share your interests, connections and other signals from your profile with the site or the app which then customises your experience, tells you which of your friends are also using the site or the app (making it more likely that you will continue to use it) or do a number of other things to present a version of the site or the app that is more relevant to you.

Introducing accurate facial recognition into the mix potentially removes the need for you to tell Facebook (or a future Facebook connected site or app) who you are before your data is shared and your experience modified. All you will need to do now is show up and let a camera see you long enough to capture a reasonably clear image of your face. From there you will be identified, placed into a particular context and things will happen. As a brand, there are some interesting opportunities. Imagine your guests arrive at your event and, instead of relying on guests to manually check in, a webcam at the door connected to your Facebook Page recognises the guests as they arrive and posts an update in your stream sharing their arrival. This isn’t happening yet but it is very possible.

Of course whether users allow this will likely depend on Facebook’s (or the relevant service’s) data protection policy (with this sort of technology, the term “privacy policy” is totally inappropriate – privacy is a memory) and the controls Facebook will make available to users to permit the service to automatically identify and tag them more publicly than it does at the moment. The challenge is that most users don’t pay much attention to their privacy settings and don’t customise them to suit their preferences. That doesn’t prevent them from being outraged when brands use their profile data in otherwise permissible ways. This may not seem like a problem but, from a reputation perspective, it can be.

Even though this technology is not implemented particularly widely, accurate facial recognition associated with identities and personal information profiles is probably not far off. It is going to scare consumers who will become aware of the myriad cameras and opportunities for them to be identified and located in specific contexts. The remnants of their privacy (by obscurity) will be whittled down to almost nothing and they won’t expect it. As a brand, this technology offers a number of opportunities to engage with customers in a very meaningful and personal way but catching them by surprise is almost certainly going to backfire, largely because the backlash will be so much more intense, precisely because the possible applications of this technology are so personal.

Preparing customers for implementations of these sorts of technologies and reducing the risk of significant reputational harm requires transparency and a healthy dose of courage to be as transparent as you need to be about how you intend engaging with your customers. As I pointed out in my talk at the recent SA Privacy Management Summit, brands have little to gain by being opaque. Transparency is a critical risk management tool, it engenders trust and keeps brands accountable and honest. That is scary for brands not accustomed to being in the spotlight but if they want to engage more effectively with their customers and earn their loyalty, they can’t do it by being evasive and catching their customers by surprise.

Widespread facial recognition will have a fairly profound impact on data protection when businesses adopt it on a larger scale. The opportunities for brands are tremendous and could, literally, revolutionise how a customer perceives a brand. To paraphrase a worn adage, with this great power comes great responsibility and brands should think carefully about how to introduce these tools to their customers and obtain their buy-in. Even though facial recognition is still in fairly limited use, brands have been using various tools and techniques to leverage customers’ identifies and personal data to customise their experiences of a brand’s products and services for some time now. Transparency is more likely to win customers’ trust even though it scares many brands silly. That said –

Courage is not the absence of fear, but rather the judgement that something else is more important than fear.

— James Neil Hollingworth

Isparta Facebook defamation case highlights a fundamental legal question

The recent Isparta v Richter and Another case in the Pretoria High Court expands on an earlier Facebook defamation case in the Johannesburg High Court and addresses a question that most people assume is answered from the start: does the defamatory material relate to the person who claims to be wronged? This case also makes an important point about the kind of compensation successful litigants are likely to receive. It’s not as much as you might expect.

The North Gauteng High Court (in Pretoria) recently delivered a judgment of Isparta v Richter and Another which I found interesting for a few reasons:

The judgment also raises an important point about the kind of monetary compensation successful litigants can expect and is a sobering reminder that the cost of litigating can far exceed a compensation award.

Who Is Defamed?

Acting Judge Hiemstra described the background to the case in paragraphs 9 and 10 of the judgment:

[9] The plaintiff and the second defendant were married to each other, but were divorced after acrimonious litigation. The plaintiff and the second defendant are still engaged in consequent litigation. The plaintiff obtained an order for the committal of the second defendant for contempt of court. She also obtained an interim interdict against him, with a return date in September 2013. The ongoing litigation concerns the second defendant’s alleged failure to comply with a settlement agreement entered into between the plaintiff and the second defendant in their divorce proceedings.

[10] The plaintiff has remarried and the first and second defendants have married each other. The plaintiffs husband has a son aged 16, who lives with her and her husband. She also has two children from her marriage with the second defendant. They are a girl, P-A, then aged 6, and a boy, G, aged 4

The first defendant posted several comments on her Facebook wall which tagged her husband, the second defendant, and which the plaintiff contended were defamatory. You can read the comments in the judgment in paragraphs 13 to 16. The posts attracted some comments from people who seemed to know the parties. The defendants admitted that the first defendant posted the comments on Facebook and that the second defendant was tagged in the comments. The plaintiff’s contentions were set out in her particulars of claim (the document describing a claim and which is usually attached to a summons) and Acting Judge Hiemstra summarised them as follows:

The plaintiff alleges in her particulars of claim that both the comments, quoted in paragraphs [13] and [16], are defamatory of her. The first one she claims, is disparaging and belittles her. She claims that the second one is malicious and aimed at damaging her reputation by implying that the plaintiff allows inappropriate interaction between her teenaged stepson and her minor daughter and that she is a bad mother.

Acting Judge Hiemstra then asked a question which is often assumed to be answered:

Do the comments refer to the plaintiff?

The judgment then evaluated a few tests which other courts had established in order to answer this question. Before I deal with that, bear in mind what the test for defamation is. In the H v W case, Judge Willis reviewed established case law in South Africa including a Supreme Court of Appeal judgment in the 2004 Mthembi-Mahanyele v Mail & Guardian which, according to Judge Willis –

affirmed the principle that the test for determining whether the words in respect of which there is a complaint have a defamatory meaning is whether a reasonable person of ordinary intelligence might reasonably understand the words concerned to convey a meaning defamatory of the litigant concerned

The Court, in the Mthembi-Mahanyele case set out the test for defamation as follows (and cited a 1993 case in the then-Appellate Division of Argus Printing and Publishing Co Ltd v Esselen’s Estate) –

The test for determining whether words published are defamatory is to ask whether a ‘reasonable person of ordinary intelligence might reasonably understand the words … to convey a meaning defamatory of the plaintiff… . The test is an objective one. In the absence of an innuendo, the reasonable person of ordinary intelligence is taken to understand the words alleged to be defamatory in their natural and ordinary meaning. In determining this natural and ordinary meaning the Court must take account not only of what the words expressly say, but also of what they imply’

Judge Willis didn’t seem to deem it necessary to deal with the question Acting Judge Hiemstra asked in the Isparta case but you can see why Acting Judge Hiemstra’s question is pretty fundamental to this case and every defamation case. The challenge facing a plaintiff in a defamation case is as follows (authorities in footnote 4 of the judgment):

A plaintiff in a defamation action must prove that the impugned statements are directed at him or her. If a plaintiff is not directly referred to in the defamatory statement, the plaintiff must plead the circumstances which would have identified him or her to the addressees.

The first test which Acting Judge Hiemstra referred to came from the 1944 English case of Knupffer v London Express Newspapers Ltd and is a two part test:

There are two questions involved in the attempt to identify the appellant as the person defamed. The first question is a question of law – can the article, having regard to its language be regarded as capable of referring to the appellant? The second question is a question of fact namely does the article in fact lead reasonable people, who know the appellant, to the conclusion that it does refer to him"? Unless the first question can be answered in favour of the appellant, the second question does not arise …

Put another way –

(a) can the words be regarded as capable of referring to the plaintiff?; and

(b) did the words in fact lead reasonable readers who know the plaintiff to the conclusion that they do refer to her?

South African courts seemed to approve this test although, in a 2009 case in the Johannesburg High Court, Acting Judge Gautchi proposed streamlining the test and formulating a single question:

The only relevant rule is that in order to be actionable, the defamatory words must be understood to be published of and concerning the plaintiff.

Acting Judge Hiemstra found the logic of this revised approach to be “unassailable” and, in applying this test to the facts, found as follows that both comments were understood to relate to the plaintiff.

Framing Defamatory Comments and Whether They Are Individually or Collectively Defamatory?

After dealing with two further issues the defendants’ lawyers raised (whether each offending post should be interpreted in isolation or together – given the timing, a “reasonable Facebook user would have understood all the postings related to issues between the plaintiff and the defendants” – and whether posing one of the comments as a question changed its apparently defamatory nature – it was essentially a statement, so no), Acting Judge Hiemstra considered another interesting question:

Are the comments individually or individually and collectively, defamatory?

Here Acting Judge Hiemstra found the comments are both individually and collectively defamatory:

[32] The first comment is to the effect that the plaintiff is meddlesome and interfering. It is a personal message addressed to the plaintiff. If the first defendant had an issue with the plaintiff, she could have addressed it with her personally However, she chose to publish it on Facebook where all her friends and friends of the plaintiff would read it. Although the first message does not constitute serious defamation, publication thereof on her Facebook wall was gratuitous and with the intention to place the plaintiff in a bad light.

[33] The second impugned posting is scandalous to the extreme It suggests that the plaintiff encourages and tolerates sexual deviation, even paedophilia Some of the defendants’ friends lapped it up with relish and added their own snide comments, compounding the damage to the plaintiff’s reputation.

[34] I therefore find that both statements are defamatory, individually and collectively.

Keeping Up to Date on Technological Developments

In the post about the H v W case, I raised a concern about judges relying too heavily on dated research as authority for how social networks work. In that case, Judge Willis relied on an article by South African Professor of Private Law at the University of South Africa, Anneliese Roos (her article is titled “Privacy in the Facebook Era: A South African Legal Perspective” and was published in the 2012 South African Law Journal) as his basis for explaining how Facebook works.

As I read Roos’ article, it seemed to me that she didn’t have a clear and current understanding of how Facebook worked at the time (and some of its mechanics have changed at least once since her article was published last year). In my post about the H v W case, I wrote the following:

Unfortunately Roos’ understanding of Facebook’s privacy controls and how it works, generally, is limited as expressed by her article and Judge Willis has accepted her explanation as a complete and accurate exposition of privacy on Facebook. This will likely, and unfortunately, colour future Court judgments regarding Facebook privacy issues.

This has started to happen. Acting Judge Hiemstra quoted portions of Judge Willis’ judgment and Roos’ article in order to explain how Facebook works and I am certain we will see this happen more and more going forward. The challenge, of course, is that Facebook (and other services) change their mechanics and settings on an ongoing basis and if court decisions are not going to be based on updated models, we are going to see a strong disconnect emerging between how judges think these services work and how they actually work. The law that emerges from our courts will become problematic because it will be based on inaccurate factual and contractual models.

Hopefully judges will become more familiar with how social networks are changing and update their conceptual models to better fit changing circumstances.

The Finding

Ultimately, Acting Judge Hiemstra found the first and second defendants to have defamed the plaintiff and ordered them, collectively and individually, to pay the plaintiff R40 000. Many people think that a defamation action could earn them substantial monetary compensation if they win and that generally is not the case in South Africa. Acting Judge Hiemstra briefly discussed monetary compensation (or “damages”) in the judgment and, referencing a Supreme Court of Appeal judgment of Mogale and Others v Seima, pointed out that –

awards in defamation cases do not serve a punitive function and are, generally, not generous.

How much a court awards depends on the nature of the defamatory material, the harm suffered and the parties’ standing. Defamation awards are rarely high and are likely exceeded by the costs of going to court in the first place. Another thing to bear in mind is that even though the successful party usually also wins a portion of his or her legal costs too, a successful litigant almost never recovers all his or her legal costs and, if the damages are relatively low, a costs recovery can be even less because it would be measured on a lower court scale which allows for lower recoveries.


p>In this case, because the compensation awarded was only R40 000 and that amount is well within the Magistrates Court’s R100 000 limit, costs were ordered on the “appropriate magistrates court scale” which may not have amounted to a significant portion of the plaintiff’s actual legal costs of her attorney and advocate (very possibly in excess of R100 000, likely even more).