Wednesday, November 7, 2018

YouTube Growth Hacking

Fascinating talk by @sophiehbishop at @DigiCultureKCL about YouTube growth hacks, and the "algorithm bros" that promote them (and themselves).

The subject of her talk was a small cadre of young men who have managed to persuade millions of followers (as well as some large corporations) that they can reveal the "secrets" of the YouTube algorithm.

I want to comment on two aspects of her work in particular. Firstly there is the question, addressed in her previous paper, of "anxiety, panic and self-optimization". When people create content on YouTube or similar platforms, they have an interest in getting their content viewed as widely as possible - indeed wide viewership ("going viral") is generally regarded as the measure of success. But these platforms are capricious, in the sense that they (deliberately) don't make it easy to manipulate this measure, and this generates a sense of precarity - not only among individual content providers but also among political and commercial organizations.

So when someone offers to tell you the secrets of success on YouTube, someone who is himself already successful on YouTube, it would be hard to resist the desire to learn these secrets. Or at least to listen to what they have to say. And risk-averse corporations may be willing to bung some consultancy money in their direction.

YouTube's own engineers describe the algorithm as "one of the largest scale and most sophisticated industrial recommendation systems in existence". Their models learn approximately one billion parameters and are trained on hundreds of billions of examples. The idea that a couple of amateurs without significant experience or funding can "reverse engineer" this algorithm stretches credibility, and Bishop points out several serious methodological flaws with their approach, while speculating that perhaps what really matters to the growth hacking community is not what the YouTube algorithm actually does but what the user thinks it does. She notes that the results of this "reverse engineering" experiment have been widely disseminated, and presented at an event sponsored by YouTube itself.

What is the effect of disseminating this kind of material? I don't know if it helps to make YouTubers less anxious, or conversely makes them more anxious than they were already. No doubt YouTube is happy about anything that encourages people to devote even more time to creating sticky content for YouTube. A dashboard (in this case, YouTube's Creator Studio) provides a framing device, focusing people's attention on certain metrics (financial gains and social capital), and fostering the illusion that the metrics on the dashboard are really the only ones that matter.

The other aspect of Bishop's work I wanted to discuss is the apparent gender polarization on YouTube - not only polarization of content and who gets to see which content, but also a significantly different operating style for male and female content providers. The traditional feminist view (McRobbie, Meehan) is that this polarization is a response to the commercial demands of the advertisers. But other dimensions of polarization have become apparent more recently, including political extremism, and Zeynep Tufekci argues that YouTube may be one of the most powerful radicalizing instruments of the 21st century. This hints at a more deeply rooted schismogenesis.

Meanwhile, how much of this was intended or foreseen by YouTube is almost beside the point. Individuals and organizations may be held responsible for the consequences of their actions, including unforeseen consequences.

Sophie Bishop, Anxiety, panic and self-optimization: Inequalities and the YouTube algorithm (Convergence, 24(1), 2018 pp 69–84)

Paul Covington, Jay Adams, Emre Sargin, Deep Neural Networks for YouTube Recommendations (Proceedings of the 10th ACM Conference on Recommender Systems, 2016, pages 191-198)

Paul Lewis, 'Fiction is outperforming reality': how YouTube's algorithm distorts truth (Guardian, 2 Feb 2018)

Zeynep Tufekci, Opinion: YouTube, the Great Radicalizer (New York Times, 10 March 2018)

Wikipedia: Schismogenesis

Related post Ethical communication in a digital age (November 2018), Polarization (November 2018)

Thursday, November 1, 2018

Ethical communication in a digital age

At the @BritishAcademy_ yesterday evening for a lecture by Onora O'Neill on Ethical Communication in a Digital Age, supported by two more philosophy professors, Rowan Cruft and Rae Langton.

Much of the discussion was about the threats posed to public reason by electronically mediated speech acts, and the challenges of regulating social media. However, although the tech giants and regulators have an important role, the primary question in the event billing was not about Them but about Us - how do *we* communicate ethically in an increasingly digital age.

I don't claim to know as much about ethics as the three professors, but I do know a bit about communication and digital technology, so here is my take on the subject from that perspective.

The kind of communication we are talking about involves at least four different players - the speaker, the spoken-to, the spoken-about, and the medium / mediator. Communication can be punctuated into a series of atomic speech acts, but it is often the cumulative effects (on public reason or public decency) that worry us.

So let me look at each of the elements of this communication in turn.

First the speech act itself. O'Neill quoted Plato, who complained that the technology of writing served to decouple the writer from the text. On social media, the authorship of speech acts becomes more problematic still. This is not just because many of the speakers are anonymous, and we may not know whether they are bots or people. It is also because the dissemination mechanisms offered by the social media platforms allow people to dissociate themselves from the contents that they may "like" or "retweet". Thus people may disseminate nasty material while perceiving themselves not as the authors of this material but as merely mediators of it, and therefore not holding themselves personally responsible for the truth or decency of the material.

Indeed, some people act online as if they believed that the online world was entirely disconnected from the real physical world, as if online banter could never have real-world consequences, and the online alter ego was an entirely different person.

Did I say truth? At the event, the three philosophers devoted a lot of time to the relationship between ethics and epistemology (questions of truth and verifiability on the Internet). But even propositional speech acts are not always easily sorted into truth and lies, while many of the speech acts that pollute the internet are not propositions but other rhetorical gestures. For example, endless repetition of "what about her emails?" and "lock her up", which are designed to frame public discourse to accord with the rhetorical goals of the speaker. (I'll come back to the question of framing later.)

The popular social media platforms offer to punctuate our speech into discrete units - the tweet, the post, the YouTube video, or whatever. Each unit is then measured separately, and the speaker may be rewarded (financially or psychologically) when a unit becomes popular (or "goes viral"). We tend to take this punctuation at face value, but systems thinkers including Bateson and Maturana have drawn attention to the relationship between punctuation and epistemology.

(Note to self - add something here about metacommunication, which is a concept Bateson took from Benjamin Lee Whorf.)

Full communication requires a listener (the spoken-to) as well as a speaker. Much of the digital literacy agenda is about coaching people to interpret and evaluate material found on the internet, enabling them to work out who is actually speaking, and whether there is a hidden commercial or political agenda.

One of the challenges of the digital age is that I don't know who else is being spoken to. Am I part of an undifferentiated crowd (unlikely) or a filter bubble (probably)? The digital platforms have developed sophisticated mechanisms for targeting people who may be particularly receptive to particular messages or content. So why have I been selected for this message, why exactly does Twitter or Facebook think this would be of interest to me? This is a fundamental divergence from older forms of mass communication - the public meeting, the newspaper, the broadcast.

And sometimes a person can be targeted with violent threats and other unpleasantries. Harassment and trolling techniques developed as part of the #GamerGate campaign are now widely used across the internet, and may often be successful in intimidating and silencing the recipients.

The third (and often unwilling) party to communication is the person or community spoken about. Where this is an individual, there may be issues around privacy as well as avoidance of libel or slander. It is sometimes thought that people in the public eye (such as Hillary Clinton or George Soros) are somehow "fair game" for any criticism or disparagement that is thrown in their direction, whereas other people (especially children) deserve some protection. The gutter press has always pushed the boundaries of this, and the Internet undoubtedly amplifies this phenomenon.

What I find even more interesting here is the way recent political debate has focused on scapegoating certain groups. Geoff Shullenberger attributes some of this to Peter Thiel.

"Peter Thiel, whose support for Trump earned him a place on the transition team, is a former student of the most significant theorist of scapegoating, the late literary scholar and anthropologist of religion RenΓ© Girard. Girard built an ambitious theory around the claim that scapegoating pervades social life in an occluded form and plays a foundational role in religion and politics. For Girard, the task of modern thought is to reveal and overcome the scapegoat mechanism–to defuse its immense potency by explaining its operation. Conversely, Thiel’s political agenda and successful career setting up the new pillars of our social world bear the unmistakable traces of someone who believes in the salvationary power of scapegoating as a positive project."

Clearly there are some ethical issues here to be addressed.

Fourthly we come onto the role of the medium / mediator. O'Neill talked about disintermediation, as if the Internet allowed people to talk directly to people without having to pass through gatekeepers such as newspaper editors and government censors. But as Rae Langton pointed out, this is not true disintermediation, as these mediators are merely being replaced by others - often amateur curators. Furthermore, the new mediators can't be expected to have the same establishment standards as the old mediators. (This may or may not be a good thing.)

Even the old mediators can't be relied upon to maintain the old standards. The BBC is often accused of bias, and its response to these accusations appears to be to hide behind a perverse notion of "balance" and "objectivity" that requires it to provide a platform for climate change denial and other farragoes.

Obviously the tech giants have a commercial agenda, linked to the Attention Economy. As Zeynep Tufekci and others have pointed out, people can be presented with increasingly extreme content in order to keep them on the platform, and this appears to be a significant force behind the emergence of radical groups, as well as a substantial shift in the Overton window. There appears to be some correlation between Facebook usage and attacks on migrants, although it may be difficult to establish the direction of causality.

But the platforms themselves are also subject to political influence - not only the weaponization of social media described by John Naughton but also old-fashioned coercion. Around Easter 2016, people were wondering whether Facebook would swing the American election against Trump. A posse of right-wing politicians had a meeting with Zuckerberg in May 2016, who then bent over backwards to avoid anyone thinking that Facebook would give Clinton an unfair advantage. (Spoiler: it didn't.)

So if there is a role for regulation here, it is not only to protect consumers from the commercial interests of the tech giants, but also to protect the tech giants themselves from improper influence.

Finally, I want to emphasize Framing, which is one of the most important ways people can influence public reason. For example, hashtags provide a simple and powerful framing mechanism, which can work to positive effect (#MeToo) or negative (#GamerGate).

President Trump is of course a master of framing - constantly moving the terms of the debate, so his opponents are always forced to debate on these terms. His frequent invocation of #FakeNews enables him to preempt and negate inconvenient facts, and his rhetorical playbook also includes antisemitic tropes (Hadley Freeman) and kettle logic (Heer Jeet). (But there are many examples of framing devices used by earlier presidents, and it is hard to delineate precisely what is new or objectionable about Trump's performance.)

In other words Rhetoric eats Epistemology for breakfast. (Perhaps that will give my philosopher friends something to chew on?)

J.L Austin, How to do things with words (Oxford University Press, 1962)

Anthony Cuthbertson, Facebook use linked to attacks on refugees, says study (Independent, 22 August 2018)

Paul F. Dell, Understanding Bateson and Maturana: Toward a Biological Foundation for the Social Sciences (Journal of Marital and Family Therapy, 1985, Vol. 11, No. 1, 1-20). (Note: even though I have both Bateson and Maturana on my bookshelf, the lazy way to get a reference is to use Google, which points me towards secondary sources like this. When I have time, I'll put the original references in.)

Alex Johnson and Matthew DeLuca, Facebook's Mark Zuckerberg Meets Conservatives Amid 'Trending' Furor (NBC News, 19 May 2016)

Robinson Meyer, How Facebook Could Tilt the 2016 Election (Atlantic, 18 April 2016)

Paul Lewis, 'Fiction is outperforming reality': how YouTube's algorithm distorts truth (Guardian, 2 Feb 2018)

John Naughton, Mark Zuckerberg’s dilemma - what to do with monster he has created? (Open Democracy, 29 October 2018)

Geoff Shullenberger, The Scapegoating Machine (The New Inquiry, 30 November 2016)

Zeynep Tufekci, YouTube, the Great Radicalizer (New York Times, 10 March 2018)

Wikipedia: Attention Economy, Disintermediation, Framing, Gamergate Controversy, Metacommunication, Overton Window

Related posts: YouTube Growth Hacking (November 2018), Polarization (November 2018)

Updated 11 November 2018

Wednesday, June 13, 2018

Practical Ethics

A lot of ethical judgements appear to be binary ones. Good versus bad. Acceptable versus unacceptable. Angels versus Devils.

Where questions of ethics reach the public sphere, it is common for people to take strong positions for or against. For example, there have been some high-profile cases involving seriously sick children, whether they should be provided with some experimental treatment, or even whether they should be kept alive at all. These are incredibly difficult decisions for those closely involved, but the experts are then subjected to vitriolic attack from armchair critics (often from the other side of the world) who think they know better.

Practical ethics are mostly about trade-offs, interpreting the evidence, predicting the consequences, estimating and balancing the benefits and risks. There isn't a simple formula that can be applied, each case must be carefully considered to determine where it sits on a spectrum.

The same is true of business and technology ethics. There isn't a blanket rule that says that these forms of persuasion are good and these forms are bad, there are just different degrees of nudge. We might want to regard all nudges with some suspicion, but retailers have always nudged people to purchase things. The question is whether this particular form of nudge is acceptable in this context, or whether it crosses some fuzzy line into manipulation or worse. Where does this particular project sit on the spectrum?

Technologists sometimes abdicate responsibility for such questions. Whatever the client wants, or whatever the technology enables, is okay. Responsibility means owning that judgement.

When Google published its AI ethics recently, Eric Newcomer complained that balancing the benefits and risks sounded like the utilitarianism he learned about at high school. But he also complained that Google's approach lacks impartiality and agent-neutrality. It would therefore be more accurate to describe Google's approach as consequentialism.

In the real world, even the question of agent-neutrality is complicated. Sometimes this is interpreted as a call to disregard any judgement made by a stakeholder, on the grounds that they must be biased. For example, ignoring professional opinions (doctors, teachers) because they might be trying to protect their own professional status. But taking important decisions about healthcare or education away from the professionals doesn't solve the problem of bias, it merely replaces professional bias with some other form of bias.

In Google's case, people are entitled to question how exactly Google will make these difficult judgements, and the extent to which these judgements may be subject to some conflict of interest. But if there is no other credible body that can make these judgements, perhaps the best we can ask for (at least for now) is some kind of transparency or scrutiny.

As I said above, practical ethics are mostly about consequences - which philosophers call consequentialism. But not entirely. Ethical arguments about the human subject aren't always framed in terms of observable effects, but may be framed in terms of human values. For example, the idea people should be given control over something or other, not because it makes them happier, but just because, you know, they should. Or the idea that certain things (truth, human life, etc.) are sacrosanct.

In his book The Human Use of Human Beings, first published in 1950, Norbert Wiener based his computer ethics on what he called four great principles of justice. So this is not just about balancing outcomes.
Freedom. Justice requires “the liberty of each human being to develop in his freedom the full measure of the human possibilities embodied in him.”  
Equality. Justice requires “the equality by which what is just for A and B remains just when the positions of A and B are interchanged.” 
Benevolence. Justice requires “a good will between man and man that knows no limits short of those of humanity itself.”  
Minimum Infringement of Freedom. “What compulsion the very existence of the community and the state may demand must be exercised in such a way as to produce no unnecessary infringement of freedom”

Of course, a complex issue may require more than a single dimension. It may be useful to draw spider diagrams or radar charts, to help to visualize the relevant factors. Alternatively, Cathy O'Neil recommends the Ethical or Stakeholder Matrix technique, originally invented by Professor Ben Mepham.

"A construction from the world of bio-ethics, the ethical or “stakeholder” matrix is a way of determining the answer to the question, does this algorithm work? It does so by considering all the stakeholders, and all of their concerns, be them positive (accuracy, profitability) or negative (false negatives, bad data), and in particular allows the deployer to think about and gauge all types of best case and worst case scenarios before they happen. The matrix is color coded with red, yellow, or green boxes to alert people to problem areas." [Source: ORCAA]
"The Ethical Matrix is a versatile tool for analysing ethical issues. It is intended to help people make ethical decisions, particularly about new technologies. It is an aid to rational thought and democratic deliberation, not a substitute for them. ... The Ethical Matrix sets out a framework to help individuals and groups to work through these debates in relation to a particular issue. It is designed so that a broader than usual range of ethical concerns is aired, differences of perspective become openly discussed, and the weighting of each concern against the others is made explicit. The matrix is based in established ethical theory but, as far as possible, employs user-friendly language." [Source: Food Ethics Council]

Jessi Hempel, Want to prove your business is fair? Audit your algorithm (Wired 9 May 2018)

Ben Mepham, Ethical Principles and the Ethical Matrix. Chapter 3 in J. Peter Clark Christopher Ritson (eds), Practical Ethics for Food Professionals: Ethics in Research, Education and the Workplace (Wiley 2013)

Eric Newcomer, What Google's AI Principles Left Out (Bloomberg 8 June 2018)

Tom Upchurch, To work for society, data scientists need a hippocratic oath with teeth (Wired, 8 April 2018)

Stanford Encyclopedia of Philosophy: Computer and Information Ethics, Consequentialism, Utilitarianism

Related posts: Conflict of Interest (March 2018), Data and Intelligence Principles From Major Players (June 2018)

Sunday, March 25, 2018

Ethics as a Service

In the real world, ethics is rarely if ever the primary focus. People engaging with practical issues may need guidance or prompts to engage with ethical questions, as well as appropriate levels of governance.

@JPSlosar ‏calls for
"a set of easily recognizable ethics indicators that would signal the presence of an ethics issue before it becomes entrenched, irresolvable or even just obviously apparent".

Slosar's particular interest is in healthcare. He wants to proactively integrate ethics in person-centered care, as a key enabler of the multiple (and sometimes conflicting) objectives of healthcare: improved outcomes, reduced costs and the best possible patient and provider experience. These four objectives are known as the Quadruple Aim.

According to Slosar, ethics can be understood as a service aimed at reducing, minimizing or avoiding harm. Harm can sometimes be caused deliberately, or blamed on human inattentiveness, but it is more commonly caused by system and process errors.

A team of researchers at Carnegie-Mellon, Berkeley and Microsoft Research have proposed an approach to ethics-as-a-service involving crowd-sourcing ethical decisions. This was presented at an Ethics-By-Design workshop in 2013.

Meanwhile, Ozdemir and Knoppers distinguish between two types of Upstream Ethics: Type 1 refers to early ethical engagement, while Type 2 refers to the choice of ethical principles, which they call "prenormative", part of the process by which "normativity" is achieved. Given that most of the discussion of EthicsByDesign assumes early ethical engagement in a project (Type 1), their Type 2 might be better called EthicsByFiat.

Cristian Bravo-Lillo, Serge Egelman, Cormac Herley, Stuart Schechter and Janice Tsai, Reusable Ethics‐Compliance Infrastructure for Human Subjects Research (CREDS 2013)

Derek Feeley, The Triple Aim or the Quadruple Aim? Four Points to Help Set Your Strategy (IHI, 28 November 2017)

Vural Ozdemir and Bartha Maria Knoppers, One Size Does Not Fit All: Toward “Upstream Ethics”? (The American Journal of Bioethics, Volume 10 Issue 6, 2010)

John Paul Slosar, Embedding Clinical Ethics Upstream: What Non-Ethicists Need to Know (Health Care Ethics, Vol 24 No 3, Summer 2016)

Conflict of Interest

@riptari (Natasha Lomas) has a few questions for DeepMind's AI ethics research unit. She suggests that
"it really shouldn’t need a roster of learned academics and institutions to point out the gigantic conflict of interest in a commercial AI giant researching the ethics of its own technology’s societal impacts"

and points out that
"there’s a reason no one trusts the survey touting the amazing health benefits of a particular foodstuff carried out by the makers of said foodstuff".

As @marionnestle remarks in relation to the health claims of chocolate,
"industry-funded research tends to set up questions that will give them desirable results, and tends to be interpreted in ways that are beneficial to their interests". (via Nik Fleming)

Nic Fleming, The dark truth about chocolate (Observer, 25 March 2018)

Natasha Lomas, DeepMind now has an AI ethics research unit. We have a few questions for it… (TechCrunch, 4 Oct 2017)

Sunday, March 18, 2018

Security is downstream from strategy

Following @carolecadwalla's latest revelations about the misuse of personal data involving Facebook, she gets a response from Alex Stamos, Facebook's Chief Security Officer.

So let's take a look at some of his hand-wringing Tweets.

I'm sure many security professionals would sympathize with this. Nobody listens to me. Strategy and innovation surge ahead, and security is always an afterthought.

According to his Linked-In entry, Stamos joined Facebook in June 2015. Before that he had been Chief Security Officer at Yahoo!, which suffered a major breach under his watch in late 2014, affecting over 500 million user accounts. So perhaps a mere 50 million Facebook users having their data used for nefarious purposes doesn't really count as much of a breach in his book.

In a series of tweets he later deleted, Stamos argued that the whole problem was caused by the use of an API that everyone should have known about, because it was well-documented. As if his job was only to control the undocumented stuff.
Or as Andrew Keane Woods glosses the matter, "Don’t worry everyone, Cambridge Analytica didn’t steal the data; we were giving it out". By Monday night, Stamos had resigned.

In one of her articles, Carole Cadwalladr quotes the Breitbart doctrine
"politics is downstream from culture, so to change politics you need to change culture"
And culture eats strategy. And security is downstream from everything else. So much then for "by design and by default".

Carole Cadwalladr ‘I made Steve Bannon’s psychological warfare tool’: meet the data war whistleblower (Observer, 18 Mar 2018) via @BiellaColeman

Carole Cadwalladr and Emma Graham-Harrison, How Cambridge Analytica turned Facebook ‘likes’ into a lucrative political tool (Guardian, 17 Mar 2018)

Jessica Elgot and Alex Hern, No 10 'very concerned' over Facebook data breach by Cambridge Analytica (Guardian, 19 Mar 2018)

Hannes Grassegger and Mikael Krogerus, The Data That Turned the World Upside Down (Motherboard, 28 Jan 2017) via @BiellaColeman

Justin Hendrix, Follow-Up Questions For Facebook, Cambridge Analytica and Trump Campaign on Massive Breach (Just Security, 17 March 2018)

Casey Johnston, Cambridge Analytica's leak shouldn't surprise you, but it should scare you (The Outline, 19 March 2018)

Nicole Perlroth, Sheera Frenkel and Scott Shanemarch, Facebook Exit Hints at Dissent on Handling of Russian Trolls (New York Times, 19 March 2018)

Mattathias Schwartz, Facebook failed to protect 30 million users from having their data harvested by Trump campaign affiliate (The Intercept, 30 March 2017)

Andrew Keane Woods, The Cambridge Analytica-Facebook Debacle: A Legal Primer (Lawfare, 20 March 2018) via BoingBoing

Wikipedia: Yahoo data breaches

Related post: Making the World more Open and Connected (March 2018)

Updated 20 March 2018 with new developments and additional commentary

Friday, March 9, 2018

Fail Fast - Burger Robotics

As @jjvincent observes, integrating robots into human jobs is tougher than it looks. Four days after it was installed in a Pasadena CA burger joint, Flippy the robot has been taken out of service for an upgrade. Turns out it wasn't fast enough to handle the demand. Does this count as Fail Fast?

Flippy's human minders have put a positive spin on the failure, crediting the presence of the robot for an unexpected increase in demand. As Vincent wryly suggests, Flippy is primarily earning its keep as a visitor attraction.

If this is a failure at all, what kind of failure is it? Drawing on earlier work by James Reason, Phil Boxer distinguishes between errors of intention, planning and execution.

If the intention for the robot is to improve productivity and throughput at peak periods, then the designers have got more work to do. And the productivity-throughput problem may be broader than just burger flipping: making Flippy faster may simply expose a bottleneck somewhere else in the system. But if the intention for the robot is to attract customers, this is of greatest value at off-peak periods. In which case, perhaps the robot already works perfectly.

Philip Boxer, ‘Unintentional’ errors and unconscious valencies (Asymmetric Leadership, 1 May 2008)

John Donohue, Fail Fast, Fail Often, Fail Everywhere (New Yorker, 31 May 2015)

Lora Kolodny, Meet Flippy, a burger-grilling robot from Miso Robotics and CaliBurger (TechCrunch 7 Mar 2017)

Brian Heater, Flippy, the robot hamburger chef, goes to work (TechCrunch, 5 March 2018)

James Vincent, Burger-flipping robot takes four-day break immediately after landing new job (Verge, 8 March 2018)

Related post Fail Fast - Why did the chicken cross the road? (March 2018)