Responsible application of artificial intelligence to surveillance: What prospects? 1

2.1 AI in the abstract

The term Artificial Intelligence was coined in the mid-20th century, based on “the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it” (McCarthy et al., 1955). The word ‘artificial’ implies ‘artefactual’ or ‘human-made’. Its conjunction with ‘intelligence’ leaves open the question as to whether the yardstick is ‘equivalent to human’, ‘different from human’ or ‘superior to human’. Conventionally (Albus, 1991; Russell & Norvig, 2003; McCarthy, 2007):

Artificial Intelligenceis exhibited by an artefact if it:

(1) evidences perception and cognition of relevant aspects of its environment; (2) has goals; and (3) formulates actions towards the achievement of those goals.

Histories of AI (e.g. Russell & Norvig, 2009, pp. 16–28; Boden, 2016, pp. 1–20) identify multiple strands and successive re-visits to much the same territory. Over-enthusiastic promotion has always been a feature of the AI arena. The revered Herbert Simon averred that “Within the very near future – much less than twenty-five years – we shall have the technical capability of substituting machines for any and all human functions in organisations. … it would be surprising if it were not accomplished within the next decade” (Simon, 1960). Unperturbed by ongoing failures, he repeated such predictions throughout the following decades. His mantle was inherited: “by the end of the 2020s [computers will have] intelligence indistinguishable to biological humans” (Kurzweil, 2005, p. 25). Such repeated exaggerations have resulted in under-delivery against expectations, a cyclical ‘boom and bust’ pattern of ‘AI winters’, and existential doubts.

The last decade has seen an(other) outbreak. Spurred on by the hype, and by the research funding that proponents’ promises have extracted, AI has excited activity in a variety of fields. Some of potential significance are natural language understanding, image processing and manipulation, artificial life, evolutionary computation aka genetic algorithms, and artificial emotional intelligence.

AI intersects with robotics, to the extent that the software installed in a robot is justifiably regarded as artificially intelligent. Robotics involves two key elements:

•

programmability, implying computational or symbol-manipulative capabilities that a designer can combine as desired (a robot is a computer); and

•

mechanical capability in the form of actuators, enabling it to act on its environment rather than merely function as a data processing or computational device (a robot is a machine).

Two further, frequently-mentioned elements of robotics are sensors, to enable the gathering of data about the devices’s environment; and flexibility, in that the device can both operate using a range of programs, and manipulate and transport materials in a variety of ways. Where robotics incorporates AI elements, additional benefits may be achieved, but the disbenefits and risks are considerably greater, because of the inherent capacity of a robot to act autonomously in the real world (Clarke, 2014a), and the temptation and tendency for the power of decision and action to be delegated to the artefact, whether intentionally or merely by accident.

2.2 Claims regarding AI applications to surveillance

Considerable scepticism is necessary when evaluating the claims of AI’s successes. This applies in all domains, but none moreso than surveillance. Civil society analyses of AI in surveillance depend very heavily on media reports that repeat the content of corporate and government media releases and that are highly superficial in their depiction of the underlying technologies. Typical of this vacuity is the assertion, unsupported by any evidence, that many workplace surveillance tools are “powered by artificial intelligence” (Cater & Heikkilä, 2021).

In a widely-read study, Feldstein (2019) identified relevant ‘AI surveillance techniques’ as being smart city/safe city platforms, facial recognition systems and smart policing (p. 16). The author was, however, unable to provide much detail about actual techniques used, beyond the ubiquitous example of facial recognition (e.g., Heikkilä, 2021). The limited technical information that is publicly available reflects the strong tendency of the operators of surveillance to obscure the nature of their activities, and in some cases even the fact of their conduct. The reasons underlying the obfuscation appear to include institutional cultures of secrecy, intellectual property considerations, weaknesses in the technologies such that transparency would enable effective countermeasures, and/or the technology’s ineffectiveness for the claimed purpose.

One of the few concrete examples that Feldstein was able to provide was “iBorderCtrl … an AI-based lie-detecting system … based on ‘affect recognition science’, which purports to read facial expressions and infer emotional states in order to render legal judgments or policy decisions” (p. 22. See iBorderCtrl, 2016). The claims for it were criticised as “pseudo-scientific security hocus pocus” (Campbell et al., 2020; Bacchi, 2021). The project is reported to have ended in 2019, in apparent failure. Evidence is not available, however, because, despite being funded as an EU research project, public access to meaningful information about it was denied on commercial grounds (Breyer, 2021).

A further feature of AI applications to surveillance is the vagueness with which the term ‘AI’ is used. From some descriptions, it could be inferred that the longstanding techniques of pattern recognition are being applied, e.g. to the extraction of useful data from images of vehicle registration-plates and faces. In other cases, the intention of product promoters may be to claim that machine learning is being employed, e.g. to perform trawls of data collections, and of mergers of disparate data collections, with the intention of detecting anomalies that create suspicion about people, objects or places. On the other hand, most forms of surveillance were developed independently of AI. They may be being enhanced using AI features; but they may merely be having gloss added through the unjustified appropriation of the AI tag to refer to an ‘advanced’ or merely ‘the latest’ version of a product.

A case in point is another of Feldstein’s few named instances: “The idea behind smart policing is to feed immense quantities of data into an algorithm – geographic location, historic arrest levels, types of committed crimes, biometric data, social media feeds – in order to prevent crime, respond to criminal acts, or even to make predictions about future criminal activity” (p. 20). The example provided was the PredPol predictive analytics program; but the description on the company’s websites (PredPol, 2021) makes clear that it is not an AI product.

2.3 Fields of AI potentially relevant to surveillance

Given that supplier obfuscation severely hampers the independent evaluation of existing products, an analysis is reported here that is based on the generic characteristics of AI technologies. This approach identifies several fields of AI that have apparent potential for application to surveillance.

Many AI fields involve ‘pattern recognition’, for which four major components are needed: “data acquisition and collection, feature extraction and representation, similarity detection and pattern classifier design, and performance evaluation” (Rosenfeld & Wechsler, 2000, p. 101).

Pattern recognition can be applied in a variety of contexts. Those relevant to surveillance include:

•

spatial patterns and image analysis (e.g. Gose et al., 1996), relevant to the monitoring of spaces and objects;

•

human biometrics images, which has surveillance applications to fingerprints (e.g., Cole, 2004), iris-scans (Daugman, 1998) and facial recognition (Ryan et al., 2009);

•

speech (e.g., O’Shaughnessy, 2008), relevant to audio-surveillance;

•

written language (Indurkhya & Damerau, 2010), relevant to the monitoring of communications recorded, made available, and read in textual form;

•

data, in the forms of data mining or data analytics (e.g. Pal & Mitra, 2004), relevant to dataveillance.

The last of these requires closer attention. Common features of the classical approaches to pattern-recognition in data have been that:

(1)

data is posited to be a sufficiently close representation of some real world phenomenon;

(2)

that data is processed by software;

(3)

inferences are drawn from that data processing;

(4)

the inferences are claimed to have relevance to the understanding or management of the phenomenon.

The software that is used may be developed in a number of different ways. An algorithm is a procedure, or set of steps. The steps may be serial and invariant. Alternatively, and more commonly, the steps may also include repetition constructs (e.g. ‘perform the following steps until a condition is fulfilled’) and selection constructs (e.‘perform one of the following sets of steps depending on some test’). From about 1965 until some time after 2000, the preparation of computing software was dominated by languages designed to enable the convenient expression of algorithms, sometimes referred to as procedural or imperative languages. Software developed in this manner represents a humanly-understandable solution to a problem, and hence the rationale underlying an inference drawn using it can be readily expressed. Techniques that express algorithms do not qualify as AI, unless the resulting artefact fulfils the conventional criteria identified in the previous section: evidence of perception, cognition and goal-seeking behaviour.

Other approaches to developing software exist (Clarke, 1991). Two that are represented as being AI techniques are highly relevant to the issues addressed in the present analysis. The approach adopted in rule-based ‘expert systems’ is to express a set of rules that apply within a problem-domain. A classic rule-example is:

If <Person> was born within the UK or any of <list of UK Colonies> between <date = 1 Jan 1949> and <date = 31 Dec 1982>, they qualify as <a Citizen of the United Kingdom and Colonies (CUKC)> and hence qualify for a UK passport

When software is developed at this level of abstraction, a model of the problem-domain exists; but there is no explicit statement of a particular problem or a solution to it. In a simple case, the reasoning underlying an inference that is drawn in a particular circumstance may be easy to provide, whether to an executive, an aggrieved person upset about a decision made based on that inference, or a judge. However, this may not be feasible where data is missing, the rulebase is large or complex, the rulebase involves intertwining among rules, the rulebase embodies some indeterminacies and/or decision-maker discretions exist. Inferences drawn from rule-based schemes in surveillance contexts, e.g. for suspicion-generation about individuals who may be involved in bomb-making, and in so-called ‘predictive policing’, are subject to these challenges.

A further important software development approach is (generically) machine learning (sometimes referred to as AI/ML), and (specifically) connectionist networks or artificial neural networks (ANNs). ANNs originated in the 1940s in the cognitive sciences, prior to the conception of AI (Medler, 1998). They have subsequently been co-opted by AI researchers and are treated as an AI technique. The essence of neural network approaches is that tools, which were possibly developed using a procedural or imperative language, are used to process examples taken from some problem-domain. Such examples might comprise the data relevant to 5% of all applicants for UK passports during some time-period who were born in, say, Jamaica, including the results of the applications.

The processing results in a set of weights on such factors as the tool treats as being involved in drawing the inference. Although the tool may have been developed using a procedural or imperative language implementing an algorithm, the resulting software that is used to process future cases is not algorithmic in nature. The industry misleadingly refers to it as being algorithmic, and critics have adopted that in terms such as ‘algorithmic bias’; but the processing involved is empirically-based, not algorithmic, and hence more appropriate terms are empirical bias and sample bias.

A critical feature of ANNs is a-rationality, that is to say that there is no reasoning underlying the inference that is drawn, and no means of generating an explanation any more meaningful than ‘based on the data, and the current weightings in the software, you don’t qualify’. The approach is referred to as ‘machine learning’ partly because the means whereby the weightings are generated depend on the collection of prior cases that are fed to the tool as its ‘learning set’. Hence, in the (to many people, somewhat strange) sense of the term used in this field, the software ‘learns’ the set of weightings. In addition, the system may be arranged so as to further adapt its weightings (i.e. ‘keep learning’), based on subsequent cases.

There are two different patterns whereby the factors and weightings can come about (DeLua, 2021). The description above was of supervised learning, in that the factors were fed to the tool by a supervisor (‘labelled’ or ‘tagged’), and in each case the ‘right answer’ was provided within the data. In the case of unsupervised learning, on the other hand, there are no labels, and ‘right answers’ are not provided with the rest of the data. The tool uses clusterings and associations to create the equivalent of what a human thinker would call constructs, but without any contextual information, ‘experience’ or ‘common sense’ about the real world that the data purports to relate to. On the one hand, ‘unsupervised learning’ is touted as being capable of discovering patterns and relationships that were not previously known; but, on the other, this greatly exacerbates the enormous scope that already exists with ‘supervised learning’ for inferences to be drawn that bear little or no relationship to the real world.

The vagaries of ‘tagging’ and even moreso of automated construct creation, coupled with the a-rationality of all AI/ML and its inherently mysterious and inexplicable inferencing, leads people who are not AI enthusiasts to be perturbed and even revulsed by the use of ANNs to make decisions that materially affect people. The issues are all the more serious when ML-based inferencing is conducted in surveillance contexts, because of the severity of the consequences for the unjustly-accused, the absence of a rationale for the inference, the strong tendencies in the system towards reversal of the onus of proof, and the near-impossibility in such circumstances of prosecuting one’s innocence.

2.4 Disbenefits and risks of AI

A great many claims have been made about the potential benefits AI might offer. Many of these feature vague explanations of the process whereby the benefits would arise. A proportion of the claims have some empirical evidence to support them, but many are mere assertions in media releases, without the support of independent testing. The analysis reported here is concerned with the downsides: disbenefits, by which is meant impacts that are predictable and harmful to some party, and risks, that is to say harmful impacts that are contingent on particular conditions arising in particular circumstances.

Pattern-matching of all kinds is inherently probabilistic rather than precise. It results in inferences that include false-positives (wrongly asserting that a match exists) and false-negatives (wrongly asserting the absence of a match). When used carefully, with inbuilt and effective safeguards against misinterpretation, benefits may arise and disbenefits and risks may be manageable. Where safeguards are missing or inadequate, the likelihood that disbenefits and risks will arise, and even dominate, increases rapidly. For example, where facial recognition is used for identity authentication, low-quality pattern-matching may cause only limited harm when a device refuses its owner permission to use it, but some alternative authentication mechanism such as a password is readily available. On the other hand, there are many other circumstances in which no alternative is available, the scope for error is high, and serious harm can arise. This is common with uses for identification of individuals presumed to be within populations for which biometric measures have already been recorded, such as at border-crossings.

As regards AI generally, the disbenefits and risks have been presented in many different ways (e.g., Scherer, 2016, esp. pp. 362–373; Yampolskiy & Spellchecker, 2016; Duursma, 2018; Crawford, 2021). Clarke (2019b) identifies five factors underlying concerns about AI:

•

Artefact Autonomy, arising from software making decisions on the basis of automated inferences, and even taking action by means of actuators under the artefact’s direct control. This can be applied as an enhancement to CCTV, for example, where directional control is programmed, e.g. to follow and zoom in on rapid movement within the device’s field of view. Many kinds of action are possible that are more directly harmful to individuals’ interests.

•

Unjustified Assumptions about Data, including its quality and its correspondence with the real-world phenomena it is assumed to represent. The risk of inappropriate outcomes is compounded when the data is drawn from different sources that had different original purposes, different attitudes to quality, and different attributed meanings for data-items.

•

Unjustified Assumptions about the Inferencing Process, due to the unsuitability of data as input to the particular inferencing process, failure to demonstrate both theoretically and empirically the applicability of the process to the particular problem-category or problem-domain, and/or assertions that empirical correlation unguided by theory is enough, and that rational explanation is an unnecessary luxury (e.g., Anderson, 2008; LaValle et al., 2011; Mayer-Schoenberger & Cukier, 2013).

•

Opaqueness of the Inferencing Process. In many circumstances, as with AI/ML, this may be empirically-based a-rationality.

•

Irresponsibility, in that none of the organisations in the AI supply-chain are subject to effective legal constraints and obligations commensurate with the roles that they play. In the currently-typical context of a long chain of technology providers and a network of inter-operating service providers, there is ample opportunity for plausible deniability, finger-pointing and hence universal liability avoidance.

The fourth of these, the lack of access to reasoning underlying inferences, has particularly serious implications (Clarke, 2019b, pp. 428–429). Where no rationale for the outcome exists and none can be convincingly constructed, no humanly-understandable explanation can be provided. The process may also be impossible to replicate, because parameters affecting it may have since changed and the prior state may not be able to be reconstructed. This means that the process may not be able to be checked by an independent party such as an auditor, judge or coroner, because records of the initial state, intermediate states and triggers for transitions between states, may not exist and may not be able to be re-constructed, such that the auditability criterion is failed.

Where an outcome appears to be in error, the factors that gave rise to it may not be discoverable, and undesired actions may not be correctable. These factors combine to provide entities that have nominal responsibility for a decision or action with an escape clause, in a manner similar to force majeure: AI’s opaqueness may be claimed to be a force that is beyond the capacity of a human entity or organisation to cope with, thereby absolving it of responsibility. In short, every test of due process and procedural fairness may be incapable of being satisfied, and accountability destroyed. Surveillance, with its inherent tendency to generate suspicion and to justify the exercise of power, is a particularly dangerous application for AI tools that, by their nature, absolve the operator from accountability.

In summary, “AI gives rise to errors of inference, of decision and of action, which arise from the more or less independent operation of artefacts, for which no rational explanations are available, and which may be incapable of investigation, correction and reparation” (Clarke, 2019b, p. 426).

The second of the five factors relates to problematic aspects of the data. In respect of AI-based data analytics, the quality of outcomes is dependent on many features of data that need to reach a threshold of quality before they can be reliably used to draw inferences (Wang & Strong, 1996; Shanks & Darke, 1998; Piprani & Ernst, 2008; summarised in Clarke 2016 into 13 factors).

As regards the third of the five factors, process quality, all data analytics techniques embody assumptions about the form that the data takes (such as the scale against which it is measured), and its quality, and the reliability of the assumptions made about the associations between the data and some part of the real world. Text-books on data analytics teach almost nothing about the need for, and the techniques that need to be applied to deliver, assurance of inferencing quality. This gives rise to challenges in relation to the use of the inferences drawn by data-analytical processes from data-sets. For inferences to be reliable, and decisions and actions taken based on those inferences equitable, there is a need for:

•

reality testing, to gain insight into the reliability of the data as a representation of relevant real-world entities and their attributes;

•

safeguards against mis-match between the abstract data-world and the real world in which impacts arise;

•

mechanisms to ensure the reasonableness and proportionality of decisions made and actions taken based on the inferences; and

•

processes whereby decisions can be contested.

Yet, despite the substantial catalogue of problems with data meaning, data quality, and inconsistencies among data-sets, data analytics teaching and practice invest a remarkably small amount of effort into quality assurance. That is the case even with long-established forms of data analytics. The reason such cavalier behaviour is possible is discussed in the following section.

AI/ML-based data analytics, on the other hand, is inherently incapable of addressing any of these issues. Further, the opacity issue overlays all of the other problems. Pre-AI, genuinely ‘algorithmic’ inferencing is capable of delivering explanations, enabling the various elements of accountability to function. Rule-based ‘expert systems’ dilute explainability. AI/ML inferencing, on the other hand, comprehensively fails the explainability test, and undermines accountability.

Procedural fairness has long been a requirement in the hitherto conventional environment of human-made or at least human-mediated decisions, for which courts demand a rational explanation. In the new world of AI, and particularly AI/ML, decisions are being imposed and actions taken that are incapable of being explained and justified before a court of law. The need for effective regulatory mechanisms is clear. What is far less clear is how protective mechanisms can be structured, and whether they are in place, or at least emergent.

3. Regulatory alternatives

AI may have very substantial impacts, both good and ill, both intended and accidental, and both anticipated and unforeseen. Building on the above review, this section presents an analysis of the regulatory spectrum, to support assessment of whether the threats inherent in AI applied to surveillance can be dealt with appropriately. The regulatory framework proposed in Clarke (2021a) is applied, in particular the Regulatory Layers in s.2.2, presented in graphical form in Fig. 1.

OPEN IN VIEWER

The foundational layer, (1) Natural Regulation, is a correlate of the natural control processes that occur in biological systems. It comprises natural influences intrinsic to the relevant socio-economic system, such as countervailing power by those affected by an initiative, activities by competitors, reputational effects, and cost/benefit trade-offs. It is incumbent on any party that argues for regulatory intervention to demonstrate that such natural influences as exist are inadequate to prevent harms arising. In the case of AI, marketing energy and unbounded adopter enthusiasm exist, and surveillance is an idea in good standing. These appear to go close to entirely negating the effects of natural regulatory processes.

The second-lowest layer, (2) Infrastructural Regulation, is exemplified by artefacts like the mechanical steam governor. It comprises particular features of the infrastructure that reinforce positive aspects and inhibit negative aspects of the relevant socio-economic system. A popular expression for infrastructural regulation in the context of IT is ‘[US] West Coast Code’ (Lessig, 1999; Hosein et al., 2003), which posits that software, and more generally architecture, have regulatory impact. Another such notion is security-by-design (Anderson, 2020). If privacy-by-design (Cavoukian, 2009) is ever articulated, and graduates beyond aspirational status, it would also represent a Layer (2) intervention. However, it appears very challenging to embed safeguards within AI-based software and artefacts (Clarke, 1993).

At the uppermost layer of the regulatory hierarchy, (7) Formal Regulation exercises the power of a parliament through statutes and delegated legislation such as Regulations. Laws demand compliance with requirements that are expressed in more or less specific terms, and are complemented by sanctions, enforcement powers and resources, and actual enforcement. Lessig (1999) refers to formal regulation as ‘[US] East Coast Code’.

Formal regulation appears to be the most logical approach when confronted by a threat of the magnitude that AI may prove to be. However, IT providers and insufficiently critical user organisations clamour for the avoidance of constraints on innovation. Corporate power has been instrumental over many decades in greatly reducing regulatory commitment in many jurisdictions and in many contexts. De-regulation and ‘better regulation’ movements have achieved ratcheting back of existing controls, commonly followed by unacceptable levels of harm, stimulating clumsy re-regulation (Braithwaite & Drahos, 1999). Safeguards have also been avoided through the outsourcing of both activities and responsibilities, including the use of low-regulation havens, and jurisdictional arbitrage. In the public sector, key factors include the drift from subcontracting, via comprehensive outsourcing, to public-private partnerships, and on towards the corporatised state (Schmidt & Cohen, 2014). A particular factor that appears to have largely ‘flown under the radar’ to date is the conversion of locally-installed software products to remotely-provided services (AI as a Service – AIaaS), of which IBM’s Watson was an early exemplar.

Several intermediate forms lie between the informal and formal ends of the regulatory hierarchy. Examples of (3) Organisational Self-Regulation include internal codes of conduct and ‘customer charters’, and self-restraint associated with expressions such as ‘business ethics’ and ‘corporate social responsibility’ (Parker, 2002). Layer (4) Industry Sector Self-Regulation involves schemes that express technical or process standards, codes of conduct or of practice or of ethics, and industry Memoranda of Understanding (MoUs). These commonly lack much impact, because organisations use them primarily as means to create an appearance of safeguards and thereby avoid formal regulatory activity. Braithwaite (2017) notes that “self-regulation has a formidable history of industry abuse of privilege” (p. 124). The conclusion of Gunningham & Sinclair (2017) is that ‘voluntarism’ is generally an effective regulatory element only when it exists in combination with ‘command-and-control’ components.

Other, intermediate forms have emerged that have been claimed to offer greater prospects of achieving regulatory objectives. These are clustered into layer (6) Meta- and Co-Regulation. In many areas, convincing arguments can reasonably be made by regulatees to the effect that government is poorly placed to cope with the detailed workings of complex industry sectors and/or the rate of change in industries’ structures, technologies and practices. Hence, the argument proceeds, parliaments should legislate the framework, objectives and enforcement mechanisms, but delegate the articulation of the detailed requirements (Ayres & Braithwaite, 1992; Parker, 2007). In practice, examples of effective layer (6) designs are uncommon, because the interests of regulatees dominate, advocates for the nominal beneficiaries lack influence and in many cases are not even at the table, and the powers of regulators are so weak that the resulting ‘enforceable Codes’ are almost entirely ineffective. For this reason, the framework in Fig. 1 also identifies the commonly-experienced layer (5) Pseudo Meta- and Co-Regulation

Despite the sceptical tone of the above analysis, several techniques in the mid-layers (3) to (6) of the hierarchy might make contributions, if they are elements within a complex of safeguards. Organisational risk assessment and management is one such technique. However, it considers risks only from the viewpoint of the organisation itself. For example, the focus of the relevant ISO Standards series (31000) is on ‘managing risks faced by organizations’. Harm to stakeholders is only within-scope where the stakeholder has sufficient power to undermine fulfilment of the organisation’s objectives. It is highly desirable that risk assessment and management processes also be conducted for those stakeholders that have legitimacy but lack power (Achterkamp & Vos, 2008). Although multi-stakeholder risk assessment is feasible (Clarke, 2019b), it remains highly unusual.

Another approach to identifying or anticipating potential harm, and devising appropriate safeguards, is impact assessment. This is a family of techniques that has matured in the environmental context, is understood in the privacy arena in theory but is to date very poorly applied (Clarke, 2009), and is emergent in broader areas of social concern (Becker & Vanclay, 2003). Impact assessment has also been described in the specific context of surveillance (Wright & Raab, 2012). However, there is no impetus for any such process to be undertaken, and little chance is evident of such approaches assisting powerless stakeholders harmed by AI.

A further possible source of protection might be application of ‘the precautionary principle’ (Wingspread, 1998). Its strong form exists in some jurisdictions’ environmental laws, along the lines of: “When human activities may lead to morally unacceptable harm that is scientifically plausible but uncertain, actions shall be taken to avoid or diminish that potential harm” (TvH, 2006). In the context of AI, on the other hand, the ‘principle’ has achieved no higher status than an ethical norm to the effect that: ‘If an action or policy is suspected of causing harm, and scientific consensus that it is not harmful is lacking, then the burden of proof falls on those taking the action’.

Finally, the notion of ‘ethically-based principles’ has been popular during the latter part of the decade to 2020, with a wave of ‘Principles for Responsible AI’ published. The documents range from trivial public relations documents from ICT corporations to serious-minded proposals from policy agencies and public interest advocates. Several catalogues have been developed, and analyses undertaken, e.g. Zeng et al., (2019), Clarke, (2019c) and Jobin et al., (2019). The second of those analyses developed a consolidated super-set of 50 Principles. This was then used as a basis for assessing the coverage of multiple further sets. The conclusion from that assessment was that “The main impression is of sparseness, with remarkably limited consensus [even on quite fundamental requirements]” (Clarke, 2019c, p. 415).

The likelihood of any combination of Layer (1)–(5) elements providing effective protection for public interests against the ravages of AI appears very low. What, then, are the prospects of effective interventions at Layers (6) and (7), Formal, Meta- and Co-Regulation?

A new phase was ushered in by a proposal for statutory intervention published by the European Commission (EC) in April 2021. This is sufficiently significant that the Proposal is evaluated here as a proxy for formal regulation generally.

5. Conclusions

There is strong evidence that data analytics practices in general are not subject to adequate safeguards for public interests, even before AI’s incursions into the field. One prominent example is the RoboDebt scandal in Australia, in which a new AUD 1 billion system both incorrectly and illegally imposed automatically-generated debts on welfare-recipients, resulting in serious impacts on half a million individuals, AUD 2 billion of repayments, and withdrawal of the scheme (Clarke, 2020). In another case, 20,000 false accusations by the taxation authority of fraudulent drawing of child benefits resulted in the Dutch government resigning (Erdbrink, 2021).

The present article has summarised the many signs of alarm about the damage AI can do, both generally, and specifically when applied to surveillance. To the formal evidence can be added the implicit recognition by AI proponents that the public has much to fear, in that they have undertaken a ‘charm offensive’ involving good news stories about AI applications, and utterances of ‘principles’ further glossed by the word ‘ethical’.

A review of the many forms that regulation can take found nothing outside the uppermost layers of formal regulation that would appear at all likely to deliver meaningful safeguards for the public against AI. Until the second quarter of 2021, there was very little evidence of formal regulation being emergent. The first such proposal, from the EC, when reviewed against a consolidated set of ‘principles for responsible AI’, has been found to be extremely poor.

Given these inadequacies, and the power of the government agencies and corporations that apply surveillance, the current prospects of effective control being achieved over AI applications to surveillance are extremely low. The history of deregulatory/regulatory cycles suggests that, unless very prompt action is taken to elevate both the urgency and the quality of proposals, regulatory protections will come long after the damage has commenced, and in the form of ill-considered, kneejerk reactions to the damage arising during the early, ‘wild west’ phase of deployment.