Online Safety Bill

Lord Allan of Hallam Excerpts

Committee stage
Thursday 22nd June 2023 Lords Chamber

[The Deputy Chairman of Committees (Baroness Barker) (LD)]

My Lords, I must advise the Committee that if Amendment 230 is agreed to then I cannot call Amendment 231 because of pre-emption.

[Lord Allan of Hallam (LD)]

My Lords, we are reaching the end of our Committee debates, but I am pleased that we have some time to explore these important questions raised by the noble Lord, Lord Knight of Weymouth.

I have an academic friend who studies the internet. When asked to produce definitive answers about how the internet is impacting on politics, he politely suggests that it may be a little too soon to say, as the community is still trying to understand the full impact of television on politics. We are rightly impatient for more immediate answers to questions around how the services regulated by this Bill affect people. For that to happen, we need research to be carried out.

A significant amount of research is already being done within the companies themselves—both more formal research, often done in partnership with academics, and more quick-fix commercial analyses where the companies do their own studies of the data. These studies sometimes see the light of day through publication or quite often through leaks; as the noble Lord, Lord Knight, has referred to, it is not uncommon for employees to decide to put research into the public domain. However, I suggest that this is a very uneven and suboptimal way for us to get to grips with the impact on services. The public interest lies in there being a much more rigorous and independent body of research work, which, rightly, these amendments collectively seek to promote.

The key issues that we need to address head-on, if we are actively to promote more research, lie within the data protection area. That has motivated my Amendment 233A—I will explain the logic of it shortly—and is the reason why I strongly support Amendment 234.

A certain amount of research can be done without any access to personal data, bringing together aggregated statistics of what is happening on platforms, but the reality is that many of the most interesting research questions inevitably bring us into areas where data protection must be considered. For example, looking at how certain forms of content might radicalise people will involve looking at what individual users are producing and consuming and the relationships between them. There is no way of doing without it for most of the interesting questions around the harms we are looking at. If you want to know whether exposure to content A or content B led to a harm, there is no way to do that research without looking at the individual and the specifics.

There is a broad literature on how anonymisation and pseudonymisation techniques can be used to try to make those datasets a little safer. However, even if the data can be made safe from a technical point of view, that still leaves us with significant ethical questions about carrying out research on people who would not necessarily consent to it and may well disagree with the motivation behind the sorts of questions we may ask. We may want to see how misinformation affects people and steers them in a bad direction; that is our judgment, but the judgment of the people who use those services and consume that information may well be that they are entirely happy and there is no way on earth that they would consent to be studied by us for something that they perceive to be against their interests.

Those are real ethical questions that have to be asked by any researcher looking at this area. That is what we are trying to get to in the amendments—whether we can create an environment with that balance of equity between the individual, who would normally be required to give consent to any use of their data, and the public interest. We may determine that, for example, understanding vaccine misinformation is sufficiently important that we will override that individual’s normal right to choose whether to participate in the research programme.

My Amendment 233A is to Amendment 233, which rightly says that Ofcom may be in a position to say that, for example, vaccine misinformation is in the overriding public interest and we need research into it. If it decides to do that and the platforms transfer data to those independent researchers, because we have said in the amendment that they must, the last thing we want is for the platforms to feel that, if there is any problem further down the track, there will be comeback on them. That would be against the principle of natural justice, given that they have been instructed to hand the data over, and could also act as a barrier.

--- Later in debate ---

[Lord Parkinson of Whitley Bay (Con)]

We are very aware that we are not the only jurisdiction looking at the important issues the Bill addresses. The Government and, I am sure, academic researchers will observe the implementation of the European Union’s Digital Services Act with interest, including the provisions about researchers’ access. We will carefully consider any implications of our own online safety regime. As noble Lords know, the Secretary of State will be required to undertake a review of the framework between two and five years after the Bill comes into force. We expect that to include an assessment of how the Bill’s existing transparency provisions facilitate researcher access.

[Lord Allan of Hallam (LD)]

I do not expect the Minister to have an answer to this today, but it will be useful to get this on the record as it is quite important. Can he let us know the Government’s thinking on the other piece of the equation? We are getting the platforms to disclose the data, and an important regulatory element is the research organisations that receive it. In the EU, that is being addressed with a code of conduct, which is a mechanism enabled by the general data protection regulation that has been approved by the European Data Protection Board and creates this legal framework. I am not aware of equivalent work having been done in the UK, but that is an essential element. We do not want to find that we have the teeth to persuade the companies to disclose the data, but not the other piece we need—probably overseen by the Information Commissioner’s Office rather than Ofcom—which is a mechanism for approving researchers to receive and then use the data.

[Lord Parkinson of Whitley Bay (Con)]

We are watching with interest what is happening in other jurisdictions. If I can furnish the Committee with any information in the area the noble Lord mentions, I will certainly follow up in writing.

--- Later in debate ---

[Lord Moylan (Con)]

I did not say that it would be impossible for a court to do this; I said it was likely to lead to high levels of inconsistency. We are dealing with what is likely to be very specialist cases. You can imagine this in the context of people feeling non-trivially psychologically harmed by statements about gender, climate, veganism, and so forth. These are the things where you see this happening. The idea that there is going to be consistency across the courts in dealing with these issues is, I think, very unlikely. It will indeed have a chilling effect on people being able to express views that may be controversial but are still valid in an open society.

[Lord Allan of Hallam (LD)]

My Lords, I want to reflect on the comments that the noble Lord, Lord Moylan, has just put to us. I also have two amendments in the group; they are amendments to the government amendment, and I am looking to the Minister to indicate whether it is helpful for me to explain the rationale of my amendments now or to wait until he has introduced his. I will do them collectively.

First, the point the noble Lord, Lord Moylan, raised is really important. We have reached the end of our consideration of the Bill; we have spent a lot of time on a lot of different issues, but we have not spent very much time on these new criminal offences, and there may be other Members of your Lordships’ House who were also present when we discussed the Communications Act back in 2003, when I was a Member at the other end. At that point, we approved something called Section 127, which we were told was essentially a rollover of the dirty phone call legislation we had had previously, which had been in telecoms legislation for ever to prevent that deep-breathing phone call thing.

--- Later in debate ---

Amendment 268AZB aims to apply business disruption enforcement measures to any internet service that “persistently fails to prevent”, or indeed allows, the illegal encouragement of self-harm. As I mentioned earlier in Committee, the Bill significantly reduces the likelihood of users encountering this material on internet sites. It requires all user-to-user services to remove this content and search services to minimise users’ access to it. I hope that that reassures the noble Lords in relation to their amendments to my amendment.

[Lord Allan of Hallam (LD)]

I completely accept that, yes, by requiring the regulated services to prevent access to this kind of content, we will make a significant difference, but it is still the case that there will be—we know there will be, because they exist today—these individual websites, blogs or whatever you want to call them which are not regulated user-to-user services and which are promoting self-harm content. It would be really helpful to know what the Government think should happen to a service such as that, given that it is outside the regulation; it may be persistently breaking the law but be outside our jurisdiction.

--- Later in debate ---

[Baroness Kidron (CB)]

My Lords, I thank the noble Lord, Lord Knight, for laying out the amendment and recognise that there was a very thoughtful debate on the subject of machine-generated content on Amendment 125 in my name on a previous day of Committee.

I appreciate that the concept of labelling or watermarking machine-generated material is central to recent EU legislation, but I am equally aware that there is more than one school of thought on the efficacy of that approach among AI experts. On the one hand, as the noble Lord, Lord Knight, beautifully set out—with the help of his artificial friend—there are those who believe that visibly marking the division of real and altered material is a clue for the public to look more carefully at what they are seeing and that labelling it might provide an opportunity for both creators and digital companies to give greater weight to “human-created material”. For example, it could be that the new BBC Verify brand is given greater validity by the public, or that Google’s search results promote it above material labelled as machine-generated as a more authentic source. There are others who feel that the scale of machine-generated material will be so vast that this labelling will be impossible or that labelling will downgrade the value of very important machine-generated material in the public imagination, when in the very near future it is likely that most human activity will be a blend of generated material and human interaction.

I spent the first part of this week locked in a room with others at the Institute for Ethics in AI in Oxford debating some of these issues. While this is a very live discussion, one thing is clear: if we are to learn from history, we must act now before all is certain, and we should act with pragmatism and a level of humility. It may be that either or both sets of experts are correct.

Industry has clearly indicated that there is an AI arms race, and many companies are launching services that they do not understand the implications of. This is not my view but one told to me by a company leader, who said that the speed of distribution was so great that the testing was confined to whether deploying large language models crashed the platforms; there was no testing for safety.

The noble Lord, Lord Stevenson, says in his explanatory statement that this is a probing amendment. I therefore ask the Minister whether we might meet before Report and look once again at the gaps that might be covered by some combination of Amendment 125 and the amendment in front of us, to make certain that the Bill adequately reflects the concerns raised by the enforcement community and reflects the advice of those who best understand the latest iterations of the digital world.

The Communications Act 2003 made a horrible mistake in not incorporating digital within it; let us not do the same here. Adding explicit safety duties to AI and machine learning would not slow down innovation but would ensure that innovation is not short-sighted and dangerous for humanity. It is a small amendment for what may turn out to be an unimaginably important purpose.

[Lord Allan of Hallam (LD)]

My Lords, it is a pleasure to follow the noble Baroness, Lady Kidron. I will try to keep my remarks brief.

It is extremely helpful that we have the opportunity to talk about this labelling question. I see it more as a kind of aperitif for our later discussion of AI regulation writ large. Given that it is literally aperitif hour, I shall just offer a small snifter as to why I think there may be some challenges around labelling—again, perhaps that is not a surprise to the noble Baroness.

When we make rules, as a general matter we tend to assume that people are going to read them and respond in a rationalist, conformist way. In reality, particularly in the internet space, we often see that there is a mixed environment and there will be three groups. There are the people who will look at the rules and respond in that rational way to them; a large group of people will just ignore them—they will simply be unaware and not at all focused on the rules; and another group will look for opportunities to subvert them and use them to their own advantage. I want to comment particularly on that last group by reference to cutlery and call centres, two historic examples of where rules have been subverted.

On the cutlery example, I am a Sheffielder, and “Made in Sheffield” used to mean that you had made the entire knife in Sheffield. Then we had this long period when we went from knives being made in Sheffield to bringing them to Sheffield and silver-plating them, to eventually just sharpening them and putting them in boxes. That is relevant in the context of AI. Increasingly, if there is an advantage to be gained by appearing to be human, people will look at what kind of finishing you need, so: “The content may have been generated by AI but the button to post it was pushed by a human, therefore we do not think it is AI because we looked at it and posted it”. On the speech of the noble Lord, Lord Knight, does the fact that my noble friend intervened on him and the noble Lord had to use some of his own words now mean that his speech in Hansard would not have to be labelled “AI-generated” because we have now departed from it? Therefore, there is that question of individuals who will want something to appear human-made even if it was largely AI-generated, and whether they will find the “Made in Sheffield” way of bypassing it.

Interestingly, we may see the phenomenon flipping the other way, and this is where my call centres come in. If people go to a popular search engine and type in “SpinVox”, they will see the story of a tech company that promised to transcribe voicemails into written text. This was a wonderful use of technology, and it was valued on the basis that it had developed that fantastic technology. However, it turned out—or at least there were claims, which I can repeat here under privilege—that it was using call centres in low-cost, low-wage environments to type those messages out. Therefore, again, we may see, curiously, some people seeing an advantage to presenting content as AI-generated when it is actually made by humans. That is just to flag that up—as I say, it is a much bigger debate that we are going to have. It is really important that we are having it, and labelling has a role to play. However, as we think about it, I urge that we remember those communities of people who will look at whatever rules we come up with and say, “Aha! Where can I get advantage?”, either by claiming that something is human when it is generated by AI or claiming that it is generated by AI if it suits them when it was actually produced by humans.

--- Later in debate ---

[Lord Knight of Weymouth (Lab)]

My Lords, I am grateful to everyone for that interesting and quick debate. It is occasionally one’s lot that somebody else tables an amendment but is unavoidably detained in Jerez, drinking sherry, and monitoring things in Hansard while I move the amendment. I am perhaps more persuaded than my noble friend might have been by the arguments that have been made.

We will return to this in other fora in response to the need to regulate AI. However, in the meantime, I enjoyed in particular the John Booth quote from the noble Baroness, Lady Bennett. In respect of this Bill and any of the potential harms around generative AI, if we have a Minister who is mindful of the need for safety by design when we have concluded this Bill then we will have dealt with the bits that we needed to deal with as far as this Bill is concerned.

[Lord Allan of Hallam (LD)]

Can the noble Lord confirm whether he generated those comments himself, or was he on his phone while we were speaking?

[Lord Knight of Weymouth (Lab)]

I do not have an invisible earpiece feeding me my lines—that was all human-generated. I beg leave to withdraw the amendment.