Sunday, May 19, 2019

The Nudge as a Speech Act

As I said in my previous post, I don't think we can start to think about the ethics of technology nudges without recognizing the complexity of real-world nudges. So in this post, I shall look at how nudges are communicated in the real world, before considering what their artificial analogues might look like.


Once upon a time, nudges were physical rather than verbal - a push on the shoulder perhaps, or a dig in the ribs with an elbow. The meaning was elliptical and depended almost entirely on context. "Nudge nudge, wink wink", as Monty Python used to say.

Even technologically mediated nudges can sometimes be physical, or what we should probably call haptic. For example, the fitness band that vibrates when it thinks you have been sitting for too long.

But many of the acts we now think of as nudges are delivered verbally, as some kind of speech act. But which kind?

The most obvious kind of nudge is a direct suggestion, which may take the form of a weak command. ("Try and eat a little now.") But nudges can also take other illocutionary forms, including questions ("Don't you think the sun is very hot here?") and statements / predictions ("You will find that new nose of yours very useful to spank people with.").

(Readers familiar with Kipling may recognize my examples as the nudges given by the Bi-Coloured-Python-Rock-Snake to the Elephant's Child.)

The force of a suggestion may depend on context and tone of voice. (A more systematic analysis of what philosophers call illocutionary force can be found in the Stanford Encyclopedia of Philosophy, based on Searle and Vanderveken 1985.)

@tonyjoyce suggests that tone of voice seems incompatible with electronic means, and this may be particularly true for the ultra-short forms. But I think we already get some differentation from electronic communications. For example I should expect an electronic announcement to "LEAVE THE BUILDING IMMEDIATELY" to have a tone of voice that conveys urgency, and we might think it is inappropriate or even unethical to use the same tone of voice for selling candy. We might put this together with other attention-seeking devices, such as flashing red text. The people who design clickbait clearly understand illocutionary force (even if they aren't familiar with the term).

A speech act can also gain force by being associated with action. If I promise to donate money to a given charity, this may nudge other people to do the same; but if they see me actually putting the money in the tin, the nudge might be much stronger. But then the nudge might be just as strong if I just put the money in the tin without saying anything, as long as everyone sees me do it. The important point is that some communication takes place, whether verbal or non-verbal, and this returns us to something closer to the original concept of nudge.

From an ethical point of view, there are particular concerns about unobtrusive or subliminal nudges. Yeung has introduced the concept of the Hypernudge, which combines three qualities: nimble, unobtrusive and highly potent. I share her concerns about this combination, but I think it is helpful to deal with these three qualities separately, before looking at the additional problems that may arise when they are combined.

Proponents of the nudge sometimes try to distinguish between unobtrusive (acceptable) and subliminal (unacceptable), but this distinction may be hard to sustain, and many people quote Luc Bovens' observation that nudges "typically work better in the dark". See also Baldwin.


I'm sure there's more to say on this topic, so I may update this post later. Relevant comments always welcome.




Robert Baldwin, From regulation to behaviour change: giving nudge the third degree (The Modern Law Review 77/6, 2014) pp 831-857

Luc Bovens, The Ethics of Nudge. In Mats J. Hansson and Till Grüne-Yanoff (eds.), Preference Change: Approaches from Philosophy, Economics and Psychology. (Berlin: Springer, 2008) pp. 207-20

John Danaher, Algocracy as Hypernudging: A New Way to Understand the Threat of Algocracy (Institute for Ethics and Emerging Technologies, 17 January 2017)

J. Searle and D. Vanderveken, Foundations of Illocutionary Logic (Cambridge: Cambridge University Press, 1985)

Karen Yeung, ‘Hypernudge’: Big Data as a Mode of Regulation by Design (Information, Communication and Society (2016) 1,19; TLI Think! Paper 28/2016)


Stanford Encyclopedia of Philosophy: Speech Acts

Related post: On the Ethics of Technologically Mediated Nudge (May 2019)


Updated 21 May 2019. Many thanks to @tonyjoyce

Friday, May 17, 2019

On the Ethics of Technologically Mediated Nudge

Before we can discuss the ethics of technologically mediated nudge, we need to recognize that many of the ethical issues are the same whether the nudge is delivered by a human or a robot. So let me start by trying to identify different categories of nudge.

In its simplest form, the nudge can involve gentle persuasions and hints between one human being and another. Parents trying to influence their children (and vice versa), teachers hoping to inspire their pupils, various forms of encouragement and consensus building and leadership. In fiction, such interventions often have evil intent and harmful consequences, but in real life let's hope that these interventions are mostly well-meaning and benign.

In contrast, there are more large-scale forms of nudge, where a team of social engineers (such as the notorious "Nudge Unit") design ways of influencing the behaviour of lots of people, but don't have any direct contact with the people whose behaviour is to be influenced. A new discipline has grown up, known as Behavioural Economics.

I shall call these two types unmediated and mediated respectively.

Mediated nudges may be delivered in various ways. For example, someone in Central Government may design a nudge to encourage job-seekers to find work. Meanwhile, YouTube can nudge us to watch a TED talk about nudging. Some nudges can be distributed via the Internet, or even the Internet of Things. In general, this involves both people and technology - in other words, a sociotechnical system.

To assess the outcome of the nudge, we can look at the personal effect on the nudgee or at the wider socio-economic impact, either short-term or longer-term. In terms of outcome, it may not make much difference whether the nudge is delivered by a human being or by a machine, given that human beings delivering the nudge might be given a standard script or procedure to follow, except in so far as the nudgee may feel differently about it, and may therefore respond differently. It is an empirical question whether a given person would respond more positively to a given nudge from a human bureaucrat or from a smartphone app, and the ethical difference between the two will be largely driven by this.

The second distinction involves the beneficiary of the nudge. Some nudges are designed to benefit the nudgee (Cass Sunstein calls these "paternalistic"), while others are designed to benefit the community as a whole (for example, correcting some market failure such as the Tragedy of the Commons). On the one hand, nudges that encourage people to exercise more; on the other hand, nudges that remind people to take their litter home. And of course there are also nudges whose intended beneficiary is the person or organization doing the nudging. We might think here of dark patterns, shades of manipulation, various ways for commercial organizations to get the individual to spend more time or money. Clearly there are some ethical issues here.

A slightly more complicated case from an ethical perspective is where the intended outcome of the nudge is to get the nudgee to behave more ethically or responsibly towards someone else.

Sunstein sees the "paternalistic" nudges as more controversial than nudges to address potential market failures, and states two further preferences. Firstly, he prefers nudges that educate people, that serve over time to increase rather than decrease their powers of agency. And secondly, he prefers nudges that operate at a slow deliberative tempo ("System 2") rather than at a fast intuitive tempo ("System 1"), since the latter can seem more manipulative.

Meanwhile, there is a significant category of self-nudging. There are now countless apps and other devices that will nudge you according to a set of rules or parameters that you provide yourself, implementing the kind of self-binding or precommitment that Jon Elster described in Ulysses and the Sirens (1979). Examples include the Tomato system for time management, fitness trackers that will count your steps and vibrate when you have been sitting for too long, money management apps that allocate your spare change to your chosen charity. Several years ago, Microsoft developed an experimental Smart Bra that would detect changes in the skin to predict when a women was about to reach for the cookie jar, and give her a friendly warning. Even if there is no problem with the nudge itself (because you have consented/chosen to be nudged) there may be some ethical issues with the surveillance and machine learning systems that enable the nudge. Especially when the nudging device is kindly made available to you by your employer or insurance company.

And even if the immediate outcome of the nudge is benefical to the nudgee, in some situations there may be concerns that the nudgee becomes over-dependent on being nudged, and thereby loses some element of self-control or delayed gratification.


The final distinction I want to introduce here concerns the direction of the nudge. The most straightforward nudges are those that push an individual in the desired direction. Suggestions to eat more healthy food, suggestions to direct spare cash to charity or savings. But some forms of therapy are based on paradoxical interventions, where the individual is pushed in the opposite directly, and they react by moving in the direction you want them to go. For example, if you want someone to give up some activity that is harming them, you might suggest they carry out this activity more systematically or energetically. This is sometimes known as reverse psychology or prescribing the symptom. For example, faced with a girl who was biting her nails, the therapist Milton Erickson advised her how she could get more enjoyment from biting her nails. Astonished by this advice, which was of course in direct opposition to all the persuasion and coercion she had received from other people up to that point, she found she was now able to give up biting her nails altogether.

(Richard Bordenave attributes paradoxical intervention to Paul Watzlawick, who worked with Gregory Bateson. It can also be found in some versions of Neuro-Linguistic Programming (NLP), which was strongly influenced by both Bateson and Erickson.)

Of course, this technique can also be practised in an ethically unacceptable direction as well. Imagine a gambling company whose official message to gamblers is that they should invest their money in a sensible savings account instead of gambling it away. This might seem like an ethically noble gesture, until we discover that the actual effect on people with a serious gambling problem is that this causes them to gamble even more. (In the same way that smoking warnings can cause some people to smoke more. Possibly cigarette companies are aware of this.)

Paradoxical interventions make perfect sense in terms of systems theory, which teaches us that the links from cause to effect are often complex and non-linear. Sometimes an accumulation of positive nudges can tip a system into chaos or catastrophe, as Donella Meadows notes in her classic essay on Leverage Points.

The Leverage Point framework may also be useful in comparing the effects of nudging at different points in a system. Robert Steele notes the use of a nudge based on restructuring information flows; in contrast, a nudge that was designed to alter the nudgee's preferences or goals or political opinions could be much more dangerously powerful, as @zeynep has demonstrated in relation to YouTube.

One of the things that complicates the ethics of Nudge is that the alternative to nudging may either be greater forms of coercion or worse outcomes for the individual. In his article on the Ethics of Nudging, Cass Sunstein argues that all human interaction and activity takes place inside some kind of Choice Architecture, thus some form of nudging is probably inevitable, whether deliberate or inadvertent. He also argues that nudges may be required on ethical grounds to the extent that they promote our core human values. (This might imply that it is sometimes irresponsible to miss an opportunity to provide a helpful nudge.) So the ethical question is not whether to nudge or not, but how to design nudges in such a way as to maximize these core human values, which he identifies as welfare, autonomy and human dignity.

While we can argue with some of the detail of Sunstein's position, I think his two main conclusions make reasonable sense. Firstly, that we are always surrounded by what Sunstein calls Choice Architectures, so we can't get away from the nudge. And secondly, that many nudges are not just preferable to whatever the alternative might be but may also be valuable in their own right.

So what happens when we introduce advanced technology into the mix? For example, what if we have a robot that is programmed to nudge people, perhaps using some kind of artificial intelligence or machine learning to adapt the nudge to each individual in a specific context at a specific point in time?

Within technology ethics, transparency is a major topic. If the robot is programmed to include a predictive model of human psychology that enables it to anticipate the human response in certain situations, this model should be open to scrutiny. Although such models can easily be wrong or misguided, especially if the training data set reflects an existing bias, with reasonable levels of transparency (at least for the appropriate stakeholders) it will usually be easier to detect and correct these errors than to fix human misconceptions and prejudices.

In science fiction, robots have sufficient intelligence and understanding of human psychology to invent appropriate nudges for a given situation. If we start to see more of this in real life, we could start to think of these as unmediated robotic nudges, instead of the robot merely being the delivery mechanism for a mediated nudge. But does this introduce any additional ethical issues, or merely amplify the importance of the ethical issues we are already looking at?

Finally, some people think that the ethical rules should be more stringent for robotic nudges than for other kinds of nudges. For example, I've heard people talking about parental consent before permitting children to be nudged by a robot. But other people might think it was safer for a child to be nudged (for whatever purpose) by a robot than by an adult human. And if you think it is a good thing for a child to work hard at school, eat her broccoli, and be kind to those less fortunate than herself, and if robotic persuasion turns out to be the most effective and child-friendly way of achieving these goals, do we really want heavier regulation on robotic child-minders than human ones?




Richard Bordenave, Comment les paradoxes permettent de réinventer les nudges (Harvard Business Review France, 30 January 2019). Adapted English version: When paradoxes inspire Nudges (6 April 2019)

Jon Elster, Ulysses and the Sirens (1979)

Jochim Hansen, Susanne Winzeler and Sascha Topolinski, When the Death Makes You Smoke: A Terror Management Perspective on the Effectiveness of Cigarette On-Pack Warnings (Journal of Experimental Social Psychology 46(1):226-228, January 2010) HT @ABMarkman

Donella Meadows, Leverage Points: Places to Intervene in a System (Whole Earth Review, Winter 1997)

Robert Steele, Implementing an integrated and transformative agenda at the regional and national levels (AtKisson, 2014)

Cass Sunstein, The Ethics of Nudging (Yale J. on Reg, 32, 2015)

Iain Thomson, Microsoft researchers build 'smart bra' to stop women's stress eating (The Register, 6 Dec 2013)

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

Wikipedia: Behavioural Insights Team ("Nudge Unit"), Reverse Psychology,

Stanford Encyclopedia of Philosophy: The Ethics of Manipulation

Related posts: Have you got big data in your underwear? (December 2014), Ethical Communication in a Digital Age (November 2018), The Nudge as a Speech Act (May 2019)

Tuesday, May 14, 2019

Leadership versus Governance

@j2bryson has commented on her blog about the fate of Google's Advanced Technology External Advisory Council (ATEAC), to which she had been appointed.

She argues that the people who were appointed to the ATEAC were selected because they were "prominent" in the field. She notes that "although being prominent doesn't mean you're the best, it probably does mean you're at least pretty good, at least at something".

Ignoring the complexities of university politics, academics generally achieve prominence because they are pretty good at having interesting and original ideas, publishing papers and books, coordinating research, and supervising postgraduate work, as well as representing the field in wider social and intellectual forums (e.g. TED talks). Clearly that can be regarded as an important type of leadership.

Bryson argues that leading is about problem-solving. And clearly there are some aspects of problem-solving in what has brought her to prominence, although that's certainly not the whole story.

But that argument completely misses the point. The purpose of the ATEAC was not problem-solving. Google does not need help with problem-solving, it employs thousands of extremely clever people who spend all day solving problems (although it may sometimes need a bit of help in the diversity stakes).

The stated purpose of the ATEAC was to help Google implement its AI principles. In other words, governance.

When Google published its AI principles last year, the question everyone was asking was about governance:
  • @mer__edith (Twitter 8 June 2018, tweet no longer available) called for "strong governance, independent external oversight and clarity"
  • @katecrawford (Twitter 8 June 2018) asked "How are they implemented? Who decides? There's no mention of process, or people, or how they'll evaluate if a tool is 'beneficial'. Are they... autonomous ethics?" 
  • and @EricNewcomer (Bloomberg 8 June 2018) asked "who decides if Google has fulfilled its commitments".

Google's appointment of an "advisory" council was clearly a half-hearted attempt to answer this question.

Bryson points out that Kay Coles James (the most controversial appointee) had some experience writing technology policy. But what a truly independent governance body needs is experience monitoring and enforcing policy, which is not the same thing at all.

People talk a lot about transparency in relation to technology ethics. Typically this refers to being able to "look inside" an advanced technological product, such as an algorithm or robot. But transparency is also about process and organization - ability to scrutinize the risk assessment and the design and the potential conflicts of interest. There are many people performing this kind of scrutiny on a full-time basis within large organizations or ecosystems, with far more experience of extremely large and complex development programmes than your average professor.

Had Google really wanted a genuinely independent governance body to scrutinize them properly, could they have appointed a different set of experts? Can people appointed and paid by Google ever be regarded as genuinely independent? And doesn't the word "advisory" give the game away? As Brustein and Bergen point out, the actual decisions are made by an internal body, the Advanced Technology Review Council, and external critics doubt that this body will ever seriously challenge Google's commercial or strategic interests.

Veena Dubal suggests that the most effective governance over Google is currently coming from Google's own workforce. It seems that their protests were significant in getting Google to disband the ATEAC, while earlier protests (re Project Maven) had led to the production of the AI principles in the first place. Clearly the kind of courageous leadership demonstrated by people like Meredith Whittaker isn't just about problem-solving.




Joshua Brustein and Mark Bergen, The Google AI Ethics Board With Actual Power Is Still Around (Bloomberg, 6 April 2019)

Joanna Bryson, What we lost when we lost Google ATEAC (7 April 2019), What leaders are actually for (13 May 2019)

Veena Dubal, Who stands between you and AI dystopia? These Google activists (The Guardian, 3 May 2019)

Bobbie Johnson and Gideon Lichfield, Hey Google, sorry you lost your ethics council, so we made one for you (MIT Technology Review 6 April 2019

Abner Li, Google details formal review process for enforcing AI Principles, plans external advisory group (9to5 Google, 18 December 2018

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

Kent Walker, An external advisory council to help advance the responsible development of AI (Google, 26 March 2019, updated 4 April 2019)


Related post: Data and Intelligence Principles From Major Players (June 2018)

Updated 15 May 2019

Sunday, April 28, 2019

Responsible Transparency

It is difficult to see how we can achieve an ethical technology without some kind of transparency, although we are still trying to work out how this could be achieved in an effective yet responsible manner. There are several concerns that are thought to conflict with transparency, including commercial advantage, security, privacy, and the risk of the device being misused or "gamed" by adversaries. There is a good summary of these issues in Mittelstadt et al (2016).

An important area where demands for transparency conflict with demands for confidentiality is with embedded software that serves the interests of the manufacturer rather than the consumer or the public. For example, a few years ago we learned about a "defeat device" that VW had built in order to cheat the emissions regulations; similar devices have been discovered in televisions to falsify energy consumption ratings.

Even when the manufacturers aren't actually breaking the law, they have a strong commercial interest in concealing the purpose and design of these systems, and they use Digital Rights Management (DRM) and the US Digital Millenium Copyright Act (DMCA) to prevent independent scrutiny. In what appears to be an example of regulatory capture, car manufacturers were abetted by the US EPA, which was persuaded to inhibit transparency of engine software, on the grounds that this would enable drivers to cheat the emissions regulations.

Defending the EPA, David Golumbia sees a choice between two trust models, which he calls democratic and cyberlibertarian. For him, the democratic model "puts trust in bodies specifically chartered and licensed to enforce regulations and laws", such as the EPA, whereas in the cyberlibertarian model, it is the users themselves who get the transparency and can scrutinize how something works. In other words, trusting the wisdom of crowds, or what he patronizingly calls "ordinary citizen security researchers".

(In their book on Trust and Mistrust, John Smith and Aidan Ward describe four types of trust. Golumbia's democratic model involves top-down trust, based on the central authority of the regulator, while the cyberlibertarian model involves decentralized network trust.)

Golumbia argues that the cyberlibertarian position is incoherent. 
"It says, on the one hand, we should not trust manufacturers like Volkswagen to follow the law. We shouldn’t trust them because people, when they have self-interest at heart, will pursue that self-interest even when the rules tell them not to. But then it says we should trust an even larger group of people, among whom many are no less self-interested, and who have fewer formal accountability obligations, to follow the law."
One problem with this argument is that it appears to confuse scrutiny with compliance. Cyberlibertarians may be strongly in favour of deregulation, but increasing transparency isn't only advocated by cyberlibertarians and doesn't necessarily imply deregulation. It could be based on a recognition that regulatory scrutiny and citizen scrutiny are complementary, given two important facts. Firstly, however powerful the tools at their disposal the regulators don't always spot everything; and secondly, regulators are sometimes subject to improper influence from the companies they are supposed to be regulating (so-called regulatory capture). Therefore having independent scrutiny as well as central regulation increases the likelihood that hazards will be discovered and dealt with. This could include the detection of algorithmic bias or previously unidentified hazards/vulnerabilities/malpractice.

Another small problem with his argument is that the defeat device had already hoodwinked the EPA and other regulators for many years.

Golumbia claims that "what the cyberlibertarians want, even demand, is for everyone to have the power to read and modify the emissions software in their cars" and complains that "the more we put law into the hands of those not specifically entrusted to follow it, the more unethical behavior we will have". It is certainly true that some of the advocates of open source are also advocating "right to repair" and customization rights. But there were two separate requests for exemptions to DMCA - one for testing and one for modification. And the researchers quoted by Kyle Wiens, who were disadvantaged by the failure of the EPA to mandate a specific exemption to DMCA to allow safety and security tests, were not casual libertarians or "ordinary citizens" but researchers at the International Council of Clean Transportation and West Virginia University.

It ought to be possible for regulators and academic researchers to collaborate productively in scrutinizing an industry, provided that clear rules, protocols and working practices are established for responsible scrutiny. Perhaps researchers might gain some protection from regulatory action or litigation by notifying a regulator in advance, or by prompt notification of any discovered issues. For example, the UK Data Protection Act 2018 (section 172) defines what it calls "effectiveness testing conditions", under which researchers can legitimately attempt to crack the anonymity of deidentified personal data. Among other things, a successful attempt must be notified to the Information Commissioner within 72 hours.

Meanwhile, in the cybersecurity world there are fairly well-established protocols for responsible disclosure of vulnerabilities, and in some cases rewards are paid to the researchers who find them, provided they are disclosed responsibly. Although not all of us have the expertise to understand the technical detail, the existence of this kind of independent scrutiny should make us all feel more confident about the safety, reliability and general trustworthiness of the products in question.




David Golumbia, The Volkswagen Scandal: The DMCA Is Not the Problem and Open Source Is Not the Solution (6 October 2015)

Brent Mittelstadt et al, The ethics of algorithms: Mapping the debate (Big Data and Society July–December 2016)

Jonathan Trull, Responsible Disclosure: Cyber Security Ethics (CSO Cyber Security Pulse, 26 February 2015)

Aidan Ward and John Smith, Trust and Mistrust (Wiley 2003)

Kyle Wiens, Opinion: The EPA shot itself in the foot by opposing rules that could've exposed VW (The Verge, 25 September 2015)


Related posts: Four Types of Trust (July 2004), Defeating the Device Paradigm (October 2015)

Tuesday, April 23, 2019

Decentred Regulation and Responsible Technology

In 2001-2, Julia Black published some papers discussing the concept of Decentred Regulation, with particular relevance to the challenges of globalization. In this post, I shall summarize her position as I understand it, and apply it to the topic of responsible technology.

Black identifies a number of potential failures in regulation, which are commonly attributed to command and control (CAC) regulation - regulation by the state through the use of legal rules backed by (often criminal) sanctions.

  • instrument failure - the instruments used (laws backed by sanctions) are inappropriate and unsophisticated
  • information and knowledge failure - governments or other authorities have insufficient knowledge to be able to identify the causes of problems, to design solutions that are appropriate, and to identify non-compliance
  • implementation failure - implementation of the regulation is inadequate
  • motivation failure and capture theory - those being regulated are insufficiently inclined to comply, and those doing the regulating are insufficiently motivated to regulated in the public interest

For Black, decentred regulation represents an alternative to CAC regulation, based on five key challenges. These challenges echo the ideas of Michel Foucault around governmentality, which Isabell Lorey (2005, p23) defines as "the structural entanglement between the government of a state and the techniques of self-government in modern Western societies".

  • complexity - emphasising both causal complexity and the complexity of interactions between actors in society (or systems), which are imperfectly understood and change over time
  • fragmentation - of knowledge, and of power and control. This is not just a question of information asymmetry; no single actor has sufficient knowledge, or sufficient control of the instruments of regulation.
  • interdependencies - including the co-production of problems and solutions by multiple actors across multiple jurisdictions (and amplified by globalization)
  • ungovernability - Black explains this in terms of autopoiesis, the self-regulation, self-production and self-organisation of systems. As a consequence of these (non-linear) system properties, it may be difficult or impossible to control things directly
  • the rejection of a clear distinction between public and private - leading to rethinking the role of formal authority in governance and regulation

In response to these challenges, Black describes a form of regulation with the following characteristics

  • hybrid - combining governmental and non-governmental actors
  • multifaceted - using a number of different strategies simultaneously or sequentially
  • indirect - this appears to link to what (following Teubner) she calls reflexive regulation - for example setting the decision-making procedures within organizations in such a way that
    the goals of public policy are achieved

And she asks if it counts as regulation at all, if we strip away much of what people commonly associate with regulation, and if it lacks some key characteristics, such as intentionality or effectiveness. Does regulation have to be what she calls "cybernetic", which she defines in terms of three functions: standard-setting, information gathering and behaviour modification? (Other definitions of "cybernetic" are available, such as Stafford Beer's Viable Systems Model.)

Meanwhile, how does any of this apply to responsible technology? Apart from the slogan, what I'm about to say would be true of any large technology company, but I'm going to talk about Google, for no other reason than its former use of the slogan "Don't Be Evil". (This is sometimes quoted as "Do No Evil", but for now I shall ignore the difference between being evil and doing evil.) What holds Google to this slogan is not primarily government regulation (mainly US and EU) but mostly an interconnected set of other forces, including investors, customers (much of its revenue coming from advertising), public opinion and its own workforce. Clearly these stakeholders don't all have the same view on what counts as Evil, or what would be an appropriate response to any specific ethical concern.

If we regard each of these stakeholder domains as a large-scale system, each displaying complex and sometimes apparently purposive behaviour, then the combination of all of them can be described as a system of systems. Mark Maier distinguished between three types of System of System (SoS), which he called Directed, Collaborative and Virtual; Philip Boxer identifies a fourth type, which he calls Acknowledged.

  • Directed - under the control of a single authority
  • Acknowledged - some aspects of regulation are delegated to semi-autonomous authorities, within a centrally planned regime 
  • Collaborative - under the control of multiple autonomous authorities, collaborating voluntarily to achieve an agreed purpose
  • Virtual - multiple authorities with no common purpose

Black's notion of "hybrid" clearly moves from the Directed type to one of the other types of SoS. But which one? Where technology companies are required to interpret and enforce some rules, under the oversight of a government regulator, this might belong to the Acknowledged type. For example, social media platforms being required to enforce some rules about copyright and intellectual property, or content providers being required to limit access to those users who can prove they are over 18. (Small organizations sometimes complain that this kind of regime tends to favour larger organizations, which can more easily absorb the cost of building and implementing the necessary mechanisms.) 

However, one consequence of globalization is that there is no single regulatory authority. In Data Protection, for example, the tech giants are faced with different regulations in different jurisdictions, and can choose whether to adopt a single approach worldwide, or to apply the stricter rules only where necessary. (So for example, Microsoft has announced it will apply GDPR rules worldwide, while other technology companies have apparently migrated personal data of non-EU citizens from Ireland to the US in order to avoid the need to apply GDPR rules to these data subjects.)

But although the detailed rules on privacy and other ethical issues vary significantly between countries and jurisdictions, there is a reasonably broad acceptance of the principle that some privacy is probably a Good Thing. Similarly, although dozens of organizations have published rival sets of ethical principles for AI or robotics or whatever, there appears to be a fair amount of common purpose between them, indicating that all these organizations are travelling (or pretending to travel) in more or less the same direction. Therefore it seems reasonable to regard this as the Collaborative type.


Decentred regulation raises important questions of agency and purpose. And if it is to be maintain relevance and effectiveness in a rapidly changing technological world, there needs to be some kind of emergent / collective intelligence conferring the ability to solve not only downstream problems (making judgements on particular cases) but also upstream problems (evolving governance principles and practices).





Julia Black, Decentring Regulation: Understanding the Role of Regulation and Self-Regulation in a ‘Post-Regulatory’ World (Current Legal Problems, Volume 54, Issue 1, 2001) pp 103–146

Julia Black, Decentred Regulation (LSE Centre for Analysis of Risk and Regulation, 2002)

Martin Innes, Bethan Davies and Morag McDermont, How Co-Production Regulates (Social and Legal Studies, 2008)

Mark W. Maier, Architecting Principles for Systems-of-Systems (Systems Engineering, Vol 1 No 4, 1998)

Isabell Lorey, State of Insecurity (Verso 2015)

Gunther Teubner, Substantive and Reflexive Elements in Modern Law (Law and Society Review, Vol. 17, 1983) pp 239-285

Wikipedia: Don't Be Evil,


Related posts: How Many Ethical Principles (April 2019)

Saturday, April 20, 2019

Ethics committee raises alarm

Dr @BenGoldacre was the keynote speaker at an IT conference I attended recently. In the context of the growing interest in technology ethics, especially AI ethics, I asked him what IT could learn from medical ethics. He responded by criticising the role of the ethics committee, and mentioned a recent case in which an ethics committee had blocked an initiative that could have collected useful data concerning the effectiveness of statins. This is an example of what Goldacre calls the ethical paradox. As he wrote in 2008,
"You can do something as part of a treatment program, entirely on a whim, and nobody will interfere, as long as it’s not potty (and even then you’ll probably be alright). But the moment you do the exact same thing as part of a research program, trying to see if it actually works or not, adding to the sum total of human knowledge, and helping to save the lives of people you’ll never meet, suddenly a whole bunch of people want to stuck their beaks in."


Within IT, there is considerable controversy about the role of the ethics committee, especially after Google appointed and then disbanded its Ethics Board. In a recent article for Slate, @internetdaniel complains about company ethics boards offering "advice" rather than meaningful oversight, and calls this ethics theatre. @ruchowdh prefers to call it ethics washing.

So I was particularly interested to find a practical example of an ethics committee in action in this morning's Guardian. While the outcome of this case is not yet clear, there seem to be some positive indicators in @sloumarsh's report.

Firstly, the topic (predictive policing) is clearly an important and difficult one. It is not just about applying a simplistic set of ethics principles, but balancing a conflicting set of interests and concerns. (As @oscwilliams reports, this topic has already got the attention of the Information Commissioner's Office.)

Secondly, the discussion is in the open, and the organization is making the right noises. “This is an important area of work, that is why it is right that it is properly scrutinised and those details are made public.” (This contrasts with some of the bad examples of medical ethics cited by Goldacre.)

Thirdly, the ethics committee is (informally) supported by a respected external body (Liberty), which adds weight to its concerns, and has helped bring the case to public attention. (Credit @Hannah_Couchman)

Fourthly, although the ethics committee mandate only applies to a single police force (West Midlands), its findings are likely to be relevant to other police forces across the UK. For those forces that do not have a properly established governance process of their own, the default path may be to follow the West Midlands example.


So it is possible (although not guaranteed) that this particular case may produce a reasonable outcome, with a valuable contribution from the ethics committee and its external supporters. But it is worrying if this is what it takes for governance to work, because this happy combination of positive indicators will not be present in most other cases.





Ben Goldacre, Where’s your ethics committee now, science boy? (Bad Science Blog,23 February 2008), When Ethics Committees Kill (Bad Science Blog, 26 March 2011), Taking transparency beyond results: ethics committees must work in the open (Bad Science Blog, 23 September 2016)

Sarah Marsh, Ethics committee raises alarm over 'predictive policing' tool (The Guardian, 20 April 2019)

Daniel Susser, Ethics Alone Can’t Fix Big Tech (Slate, 17 April 2019)

Jane Wakefield, Google's ethics board shut down (BBC News, 5 April 2019)

Oscar Williams, Some of the UK’s biggest police forces are using algorithms to predict crime (New Statesman, 4 February 2019)

Saturday, March 9, 2019

Upstream Ethics

We can roughly characterize two places where ethical judgements are called for, which I shall call upstream and downstream. There is some inconsistency about how these terms are used in the literature; here are my definitions.

I use the term upstream ethics to refer to
  • Establishing priorities and goals - for example, emphasising precaution and prevention
  • Establishing general principles, processes and practices
  • Embedding these in standards, policies and codes of practices
  • Enacting laws and regulations
  • Establishing governance - monitoring and enforcement
  • Training and awareness - enabling, encouraging and empowering people to pay due attention to ethical concerns
  • Approving and certifying technologies, products, services and supply chains. 
Some people call these (or some of them) "pre-normative" ethics.


I use the term downstream ethics to refer to
  • Making judgements about a specific instance
  • Eliciting values and concerns in a specific context as part of the requirements elicitation process
  • Detecting ethical warning signals
  • Applying, interpreting and extending upstream ethics to a specific case or challenge
  • Auditing compliance with upstream ethics

There is also a feedback and learning loop, where downstream issues and experiences are used to evaluate and improve the efficacy of upstream ethics.


Downstream ethics does not take place at a single point in time. I use the term early downstream to mean paying attention to ethical questions at an early stage of an initiative. Among other things, this may involve picking up early warning signals of potential ethical issues affecting a particular case. Early downstream means being ethically proactive - introducing responsibility by design - while late downstream means reacting to ethical issues only after they have been forced upon you by other stakeholders.

However, some writers regard what I'm calling early downstream as another type of upstream. Thus Ozdemir and Knoppers talk about Type 1 and Type 2 upstream. And John Paul Slosar writes

"Early identification of the ethical dimensions of person-centered care before the point at which one might recognize the presence of a more traditionally understood “ethics case” is vital for Proactive Ethics Integration or any effort to move ethics upstream. Ideally, there would be 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."

For his part, as a lawyer specializing in medical technology, Christopher White describes upstream ethics as a question of confidence and supply - in other words, having some level of assurance about responsible sourcing and supply of component technologies and materials. He mentions a range of sourcing issues, including conflict minerals, human slavery, and environmentally sustainable extraction.

Extending this point, advanced technology raises sourcing issues not only for physical resources and components, but also for intangible inputs like data and knowledge. For example, medical innovation may be dependent upon clinical trials, while machine learning may be dependent on large quantities of training data. So there are important questions of upstream ethics as to whether these data were collected properly and responsibly, which may affect the extent to which these data can be used responsibly, or at all. As Rumman Chowdhury asks, "How do we institute methods of ethical provenance?"

There is a trade-off between upstream effort and downstream effort. If you take more care upstream, you should hope to experience fewer difficulties downstream. Conversely, some people may wish to invest little or no time upstream, and face the consequences downstream. One way of thinking about responsibility is shifting the balance of effort and attention upstream. But obviously you can't work everything out upstream, so you will always have further stuff to do downstream.

So it's about getting the balance right, and joining the dots. Wherever we choose to draw the line between "upstream" and "downstream", with different institutional arrangements and mobilizing different modes of argumentation and evidence at different stages, "upstream" and "downstream" still need to be properly connected, as part of a single ethical system.




(In a separate post, Ethics - Soft and Hard, I discuss Luciano Floridi's use of the terms hard and soft ethics, which covers some of the same distinctions I'm making here but in a way I find more confusing.)

Os Keyes, Nikki Stevens, and Jacqueline Wernimont, The Government Is Using the Most Vulnerable People to Test Facial Recognition Software (Slate 17 March 2019) HT @ruchowdh

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) https://doi.org/10.1080/15265161.2010.482639

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

Christopher White, Looking the Other Way: What About Upstream Corporate Considerations? (MedTech, 29 Mar 2017)


Updated 18 March 2019