The Online Safety Act: proactive illegality duties, safeguards and proportionality

Part 4 of a short series of reflections on Ofcom’s Illegal Harms consultation under the Online Safety Act 2023 (OSA). 

A significant proportion of the consultation’s discussion of Ofcom's proposed Code of Practice recommendations — especially those involving proactive monitoring and detection of illegal content — is taken up with enumerating and evaluating safeguards to accompany each recommended measure.

That is to be expected, for two reasons. First, the OSA itself provides in Schedule 4 that measures recommended in a Code of Practice must be designed in the light of the importance of protecting the privacy of users and the right of users to freedom of expression within the law, and (where appropriate) incorporate safeguards for the protection of those matters.

Second, the potential interference with users' fundamental rights (notably freedom of expression and privacy) brings into play the European Convention on Human Rights (ECHR) and the Human Rights Act (which, following the UK's recent general election, we can assume will be with us for the foreseeable future).

The first step in the ECHR analysis is to consider whether the interference is “prescribed by law”. This is a threshold condition: if the interference fails that test, it is the end of the matter. When considering whether an interference contained in a statute is prescribed by law, it is not enough that the law has been passed by Parliament and is publicly accessible. It also has to have the “quality of law”: it must be sufficiently clear and precise that someone potentially affected by it can foresee in advance, with reasonable certainty, how the law will apply to them.

Requirements (strictly speaking, in the case of the OSA, Ofcom recommendations) for automated proactive detection, filtering and removal of user content present a particularly high risk of arbitrary interference with, and over-removal of, legal content. They can also be seen as a species of prior restraint. The European Court of Human Rights observed in Yildirim that "the dangers inherent in prior restraints are such that they call for the most careful scrutiny on the part of the Court".

Compatibility with the ECHR operates at two levels: the legislative measure and individual decisions taken under it. A court will regard itself as well placed, since it has the facts to hand, to determine whether an individual decision is or is not a justified interference with a Convention right. It is far less willing to declare a legislative measure per se incompatible, unless it is clear that when applied in practice it will result in a breach of Convention rights in most or all cases. If the measure is capable of being operated in a way that does not breach the Convention, then it will not be per se incompatible.

However, there is an important rider: the UK courts have said that in order to protect against arbitrary interference there must be safeguards which have the effect of enabling the proportionality of the interference to be adequately examined.

In the case of legislation such as the OSA, where the Act frames the duties at a very high level and a regulator is authorised to flesh them out, the necessary safeguards have to be provided in Ofcom's Codes of Practice and its statutory guidance. If such safeguards are not provided, or if they are not sufficient, then the regime will fall at the first Convention hurdle of not being prescribed by law. The ECHR compatibility of the regime on this score is thus heavily dependent on Ofcom's work product.

Much judicial ink has been expended on explaining the precise underlying rationale for the “capable of being adequately examined" test. It is safest to regard it as an aspect of the prescribed by law (a.k.a. “legality”) principle: the reason why legislation must be reasonably clear and precise is in order to prevent arbitrariness and the abuse of imprecise rules or unfettered discretionary powers. If the impact of the scheme is foreseeable, then its proportionality is capable of being assessed. If its impact across the board is not discernible, then its impact will be arbitrary.

Lady Hale said in the Supreme Court case of Gallagher:

“The foundation of the principle of legality is the rule of law itself - that people are to be governed by laws not men. They must not be subjected to the arbitrary - that is, the unprincipled, whimsical or inconsistent - decisions of those in power. 

This means, first, that the law must be adequately accessible and ascertainable, so that people can know what it is; and second, that it must be sufficiently precise to enable a person - with legal advice if necessary - to regulate his conduct accordingly. The law will not be sufficiently predictable if it is too broad, too imprecise or confers an unfettered discretion on those in power. 

This is a separate question from whether the law in question constitutes a disproportionate interference with a Convention right -the law in question must contain safeguards which enable the proportionality of the interference to be adequately examined. 

This does not mean that the law in question has to contain a mechanism for the review of decisions in every individual case: it means only that it has to be possible to examine both the law itself and the decisions made under it, to see whether they pass the test of being necessary in a democratic society.”

In the final analysis it may be said that safeguards have to provide sufficient protection against arbitrariness.

The courts have stressed that challenging an entire regime ex ante on proportionality grounds presents a high hurdle and will rarely succeed, compared with a challenge by an individual who claims that their rights have been violated in a particular instance. Nevertheless, the safeguards proposed by Ofcom have to pass the prescribed by law test. If they do pass, then the actual proportionality of a given interference can be considered should a case arise.

The impact of the legality requirement and the nature of the required safeguards have to be considered in the light of the triangular structure of the Online Safety Act regime. We are not here dealing with a discretionary power vested in a state official to direct a user to take down their post. The OSA regime places legal obligations on intermediary service providers. The steps that they take to comply with those obligations have the potential to affect users' rights, particularly freedom of expression. 

Foreseeability requires that a user should be able to predict, with reasonable certainty, whether their contemplated online post is liable to be affected by actions taken by a service provider in discharging its obligations under the Act.

The safeguards stipulated by Ofcom should therefore provide the requisite degree of predictability for users in respect of blocking and removal actions to be taken by service providers when carrying out Ofcom's recommended measures.

As regards the consultation’s general approach to ECHR compliance, two points stand out. The first is that there is virtually no discussion of the “prescribed by law” requirement. Its existence is recited in many places, but the substantive discussion of ECHR compatibility proceeds directly to discussion of legitimate aim, necessity and proportionality of the recommended measures. Para 1.14 of the consultation may provide a clue as to why that is:

“In passing the Act, Parliament has set out in legislation the interferences prescribed by law and which it has judged to be necessary in our democratic society.”

Similarly in para 12.64:

“…our starting point is that Parliament has determined that services should take proportionate steps to protect UK users from illegal content. Of course there is some risk of error in them doing this, but that risk is inherent in the scheme of the Act.”

There is possibly a hint here of regarding the fact that Parliament has passed legislation as being sufficient of itself to satisfy the “prescribed by law” requirement. That may be the starting point, but it is not the end point.

The second point is that insofar as Ofcom has focused on the need for clarity and certainty, it has done so from the perspective of providing clarity to service providers. The Act requires this. Schedule 4 provides that the measures described in a Code of Practice must be:

“sufficiently clear, and at a sufficiently detailed level, that providers understand what those measures entail in practice;”

That, however, does not detract from the ECHR requirement that the potential for interference with users’ privacy and freedom of expression must also be reasonably clear and precise.

The two requirements do not necessarily go hand in hand. A provision may be clear as to the amount of discretion that it gives to a service provider, yet unforeseeable in its effect on the freedom of expression of users.

Several aspects of Ofcom's proposed safeguards in relation to automated detection and related takedowns give pause for thought on the question of capability to assess the proportionality of the interference. The recommendations (which would apply only to some services) are: 

• Perceptual hash matching against a database of known CSAM material (draft U2U Code of Practice, A4.23)
• URL matching against a list of known CSAM URLs (draft U2U Code of Practice, A4.37)
• Fuzzy keyword matching to detect articles for use in fraud (draft U2U Code of Practice, A4.45)

The concerns are most apparent with the fraud keyword proposal, albeit they are not entirely absent with CSAM hash and URL matching. 

URL matching presents the fewest challenges. Ofcom's proposed safeguards relate entirely to the process for establishing and securing the list of URLs. They provide that the service provider should source the list from: 

“a person with expertise in the identification of CSAM, and who has arrangements in place to [inter alia] secure (so far as possible) that URLs at which CSAM is present, and domains which are entirely or predominantly dedicated to CSAM, are correctly identified before they are added to the list; to review CSAM URLs on the list, and remove any which are no longer CSAM URLs” [draft Code of Practice, A4.40]

By way of further safeguards, both the person with expertise and the service provider should secure the list from unauthorised access, interference or exploitation (whether by persons who work for the provider or are providing a service to the provider, or any other person).

The reasonable assumption is that the technology is capable of accurately matching detected URLs with the list, such that no further safeguards are required on that score.

If there were any concern about adequacy of these safeguards, it would probably be whether "a person with expertise in the identification of CSAM" is sufficiently precisely articulated.

For CSAM hash matching the draft Code of Practice contains equivalent safeguards to URL matching for establishment and security of the hash database. However, further safeguards are required since the recommendation of perceptual hashing introduces an element of judgement into the matching process, with the concomitant risk of false positives and consequent blocking or removal of legal user content.

Here the adequacy of the proposed safeguards may be open to more serious debate. The draft Code of Practice states that the perceptual hashing technology should be configured so that its performance strikes "an appropriate balance between precision and recall".

Precision and recall refers to the incidence of false positives and missed hits. There is typically a trade-off: fewer missed hits means more false positives.  As to what is an appropriate balance between them, the draft Code of Practice stipulates that the provider should ensure that the following matters are taken into account: 

- The risk of harm relating to image-based CSAM, as identified in the risk assessment of the service, and including in particular information reasonably available to the provider about the prevalence of CSAM content on its service.

- The proportion of detected content that is a false positive; and

- The effectiveness of the systems and/or processes used to identify false positives.

Annex 15 to the Consultation suggests various further factors that could point towards striking the balance towards either precision or recall.

The draft Code of Practice stipulates that human moderators should review “an appropriate proportion” of material detected as CSAM, and sets out principles that the service provider should take into account in deciding what proportion of detected content it is appropriate to review - for instance that the resource dedicated to review of detected content should be proportionate to the degree of accuracy achieved by the perceptual hash matching technology. It also provides various periodic review and record-keeping recommendations.

Annex 15 sets out Ofcom’s reasons (related to differences between perceptual hash technologies) for not setting a threshold which should be used to determine whether an image is a match.

The substantive balancing and proportionality decisions are thus parked firmly on the desk of the service provider. However, neither the draft Code of Practice nor the Act itself contains any indication of what is to be regarded as a proportionate or disproportionate level of interference with legal user content.

The result is that two different service providers could readily apply the stipulated safeguards in equivalent factual situations, follow the prescribed process and reach significantly differing conclusions about what is an appropriate balance between precision and recall, or about what resource should be devoted to human review. Consequently it can be argued that the effect on user content cannot be predicted. That smacks of arbitrariness. 

The safeguards for fuzzy keyword detection of articles for use in fraud are more extensive, as would be expected for a technology that is inherently more likely to throw up false positives. The consultation document points out that the recommendation:

"...differs from our proposed measures regarding CSAM hashing and the detection of CSEA links which focus on the detection of positive matches with content (or URLs that provide access to content) that has already been determined to be illegal." [Annex 15, A15.121]

Unlike with CSAM URL and hash matching the draft Code of Practice envisages that the service provider may compile its own list of fraud keywords. It contains safeguards around establishment, testing, review and security of the list. It contains equivalent provisions to perceptual hash matching for configuration of the technology so as to strike “an appropriate balance between precision and recall”, stipulating equivalent matters to be taken into account. Ofcom envisages that the safeguards will mean that it will be ‘highly likely’ that a keyword hit will correspond to an offence:

“In light of the above, we would expect any content detected as a result of applying this [keyword technology] measure to be highly likely to amount to an offence concerning articles for use in frauds.” [Volume 4, para 14.249]

It goes on:

“We recognise however that the keyword detection measure we are considering will enable services to identify content about which no prior illegal content judgment or determination has been made and that it may result in false positives. It may identify legitimate content (such as news articles or academic articles) which discuss the supply of articles for use in fraud. It is for this reason that we are not recommending that services take down all content detected by the technology, and are instead recommending that it be considered by services in accordance with their internal content moderation policies.” [ibid]

As with perceptual hash matching the draft Code of Practice provides for after the event periodic human review of some detected content. Whereas for perceptual hash matching this has to be ‘an appropriate proportion’, for fraud detection it has to be ‘a reasonable sample’. Again, it sets out principles to be taken into account in deciding what is a reasonable sample. These bear some similarities to, but are not identical to, those for perceptual hash matching. For instance there is no stipulation that review resource should be proportionate to the degree of accuracy achieved by the technology.

Evaluating the adequacy of the fraud keyword safeguards is complicated by the latitude that the recommendations give service providers as to what kind of action to take following initial keyword detection, and possible statutory interpretation questions as to whether (and if so in what way) the illegality judgement provisions of S.179 and the swift takedown obligations of S.10(3)(b) apply. 

Ofcom's approach is summarised thus:

"... we do not consider it appropriate to recommend that services swiftly take down all content detected as a positive match by their keyword detection technology, instead we recommend (as discussed below) that the decision on whether or not the content should be taken down should be taken in accordance with their content moderation systems and processes." [Annex 15, A15.122]

This is consistent with Ofcom’s broader policy approach to content moderation:

“Given the diverse range of services in scope of the new regulations, a one-size-fits-all approach to content moderation would not be appropriate. Instead of making very specific and prescriptive proposals about content moderation, we are therefore consulting on a relatively high-level set of recommendations which would allow services considerable flexibility about how to set up their content moderation teams.” [Volume 4, p.18]

 Ofcom continues, in relation to its fraud keyword recommendations:

“Consistent with Chapter 12, we are not persuaded that it would be appropriate to specify in detail how services should configure their content moderation systems and processes to take account of content detected by the keyword detection technology (for example, that there be human moderation of all such content), or the outcomes that those systems and processes should achieve (for example, through detailed KPIs).” [Annex 15, A15.123]

It then says:

“We are proposing in that Chapter that all U2U service providers must have in place content moderation systems or processes designed to take down illegal content swiftly.” [Annex 15, A15.124]

The area in which the keyword recommendations depart most significantly from hash and URL matching is thus in the steps to be taken in respect of positive keyword matches: treating them in accordance with the service provider’s internal content moderation systems and processes.  Ofcom’s approach is not to be prescriptive but to give service providers broad latitude in what steps to take in respect of positive keyword matches.

There is, however, an underlying dilemma. There are significant costs and risks associated with being prescriptive: the interference with a platform’s own rights (e.g under ECHR Protocol 1, Article 1), the unlikelihood that a single size of straitjacket can fit all in-scope service providers, prejudicing existing services, the chilling or dampening effect on development of new services, and the greater likelihood that faced with a prescriptive requirement service providers will take an over-cautious approach to blocking and removals. 

Yet the less prescriptive the measures, the broader the range of permissible approaches, the less predictable the effect on users and the greater the likelihood of arbitrary interference with user rights. This dilemma is not of Ofcom’s making. It is hardwired into the Act, but it falls to Ofcom to resolve it. It is an unenviable task. It may be impossible.

Specifically in relation to the fraud keyword detection recommendation, Ofcom says:

"... Implementations that substantially impact on freedom of expression, including the automatic take down of detected content, could be in accordance with the measure in our Code of Practice.” [Chapter 14, para 14.283]

and:

"whether or not such content were, incorrectly, subject to takedown would depend on the approach, to content moderation adopted by the service, rather than the content's detection by the keyword detection technology in and of itself." [Chapter 14, paras 14.284, 14.302]

Ofcom acknowledges that:

“There could therefore be variation in the impact on users’ freedom of expression arising from services’ different implementations of the technology and different approaches to moderation and take down of any detected content.” [para 14.283]

Ofcom, does not, however, discuss the implications for the “capable of being adequately examined" requirement if those variations are insufficiently foreseeable.

The discussion in Annex 15 contemplates that a service provider might have “no systems and processes in place to identify false positives before content is taken down”. That, it is said, would be a factor leaning towards configuring the system to towards greater precision at the expense of recall.

Recommended safeguards for content moderation generally include setting of performance targets, as they relate to accuracy of decision-making; training and materials; and appeals. For performance targets, it is for the service provider to balance the desirability of taking illegal content down swiftly against the desirability of making accurate moderation decisions. As above, different service providers could apply that guidance yet reach significantly different conclusions.

In the context of proportionality Ofcom seeks to diminish the impact on users’ freedom of expression by exempting news publisher content from the fraud keyword matching recommendation (reflecting the Act's exclusion of such content from regulated U2U content). However, that prompts the question of how service providers are to distinguish between news publisher content and the rest, in the context of an automated system: something which raises its own safeguards issues.

Ofcom’s fraud keywords recommendation cross-refers to its Recommendation 4B for large or multi-risk services: that the provider should set and record (but need not necessarily publish) internal content policies setting out rules, standards and guidelines around what content is allowed and what is not, and how policies should be operationalised and enforced. The policies should be drafted such that illegal content (where identifiable as such) is not permitted.    

Recommendation 4A (which is stated not to apply to CSAM perceptual hash and URL matching, but does not exclude fraud keyword detection) also appears potentially relevant to the fraud keyword matching recommendation: the service provider should have systems or processes designed to swiftly take down illegal content of which it is aware (mirroring the statutory obligation in S.10(3)).

Recommendation 4A goes on that for that purpose when the provider has reason to suspect that content may be illegal content, the provider should either make an illegal content judgement in relation to the content and, if it determines that the content is illegal, swiftly take it down; or do the same where its terms of service prohibit the type of illegal content in question and the content is in breach. 

Ofcom comments in relation to Recommendation 4A that:

"The design of this option is not prescriptive as to whether services use wholly or mainly human or automated content moderation processes." [Volume 12, para 12.50]

Thus there appears to be the potential four-way interaction between internal content moderation policies, the statutory takedown obligation, the Recommendation 4A takedown recommendation, and the provider's public terms of service. 

How these might mesh with each other is not immediately clear to this reader. In part this could depend on questions of interpretation of the Act, such as whether awareness for purposes of the statutory takedown obligation requires human awareness or can be satisfied by an automated system, and if so whether awareness equates to reasonable grounds to infer under S.192.    

Overall, the scope for arbitrary interference on user rights of freedom of expression appears to be greater for fraud keyword detection than with CSAM hash and URL matching.

The question of safeguards for proactive, automated detection systems is due to raise its head again. Ofcom has said that it is planning an additional consultation later this year on how automated tools, including AI, can be used to proactively detect illegal content and content most harmful to children – including previously undetected child sexual abuse material.

30 July 2024. Correction to description of 'high hurdle'.