Lord Stevens of Birmingham (CB)

Essentially, I agree with the right reverend Prelate the Bishop of Hereford. I could almost leave it there, but I will briefly say, in the spirit of the amendments, that the tablers are right to raise general concerns about the possibility of abuse through bias—as we heard from the noble Baroness, Lady O’Loan—and hallucination. After all, we have had the first high-profile resignation of a public sector leader in the form of the chief constable of the West Midlands praying in aid the fabrication of a non-existent football match as the reason why Parliament was misled. In addition to bias and hallucination, there is the risk of what is called scheming. The results from some of the LLMs—published, for example, in the journal Nature in October—show some pretty disturbing examples. In the article, headed “AI Models that Lie, Cheat, and Plot Murder”, there are examples where models have attempted to write self-propagating worms, fabricate legal documentation and leave hidden notes to future instances of themselves. The punchline, essentially, is that, in regard to some of these technologies, “the world is in a lucky period in which models are smart enough to scheme but not smart enough to escape monitoring”. That is scary because, in five years’ time, that may no longer be true. So there are good reasons for generalised concerns about AI and wanting to circumscribe the role it might play in this legislation. However, for the reasons that others have mentioned— I suspect the noble Baroness, Lady Coffey, herself would accept this—this probing amendment is written too broadly. It says: “Artificial intelligence must not be used to carry out any functions in any section or schedule of this Act”. Given that, for example, under Clauses 5(5) and 12(2), a doctor has to discuss with a person their diagnosis, their prognosis, any treatments available, the likely effects of them, and palliative, hospice and other care, it is highly likely that those will be informed by machine learning. It will interpret, for example, CT scans or MRIs, and AI tools will personalise and optimise therapies, potentially with predictive AI for better prognosis. So, were this to come back on Report, there would be a good case for ensuring greater precision in the firepower that is aimed at this particular concern. However, all that should not in any way excuse or divert us from an equivalent worry: we must not kid ourselves that the gold standard is human expert judgment on many of the questions posed by the Bill. As we discussed, Clause 2(1)(b) requires an assessment of whether somebody with a terminal illness will live longer than six months. Unfortunately, as we have heard, that turns out to be a clinically irrelevant threshold that is very hard for expert judgment to get right. I have just pulled the data from a large study looking at 98,000 people across London over the last decade and at prognostic accuracy, and the answer was that clinicians were able to be accurate about whether somebody was going to live for two weeks with about 74% accuracy, and they were able to be accurate about whether somebody was going to live more than a year with 83% accuracy, but, in terms of being able to predict whether somebody is going to survive for weeks or months, accuracy was only 32%. So, whatever our concerns about AI, human expert judgment, which underpins the Bill, is itself highly fallible.