Mr Blottisham entered the Senior Common Room carrying a report on the growing use of artificial intelligence in scientific research.
“I believe we can now settle the question of AI and scientific creativity,” he announced. “If artificial intelligence tends to produce similar ideas, we need only ensure that scientists select the best ones. The problem, if there is one, is a failure of selection.”
Professor Quillibrace looked up from his book.
“And who, precisely, will select the best ones?”
“Scientific reviewers, naturally.”
“Using what criteria?”
“The established criteria of scientific excellence.”
Miss Elowen Stray put down her cup. “And if those criteria favour ideas that resemble what science already accepts?”
Blottisham looked at her with mild astonishment.
“Why should that be a problem? We have established standards for a reason.”
“Of course,” she said. “But standards that help us recognise good work may also make it harder to recognise work that challenges the assumptions on which those standards rest.”
“An idea that challenges established knowledge must still demonstrate its worth.”
“Certainly. But consider what happens if an unfamiliar idea is difficult to formulate in the accepted vocabulary, difficult to assess by familiar methods, and difficult to justify within established expectations. It may be rejected before its significance can be properly understood.”
Quillibrace closed his book. “The question is not whether science should abandon its standards. It is whether the conditions under which those standards operate permit genuinely different possibilities to be explored.”
Blottisham shook his head.
“I think you are making a straightforward matter unnecessarily complicated. Suppose an AI system generates a thousand hypotheses. Researchers test them, reject the weak ones, and retain the strongest. Science advances.”
“Perhaps,” said Quillibrace. “But suppose the system generates hypotheses that tend to follow familiar patterns, because those patterns are prominent in its training. Suppose researchers favour suggestions that fit established expectations. Suppose reviewers then reward papers that conform to those expectations. What might happen?”
“Science would become more efficient.”
“At what?”
Blottisham paused.
“At identifying good hypotheses.”
“Even if the process makes unfamiliar hypotheses less likely to be generated, pursued or accepted?”
Blottisham considered this objection.
“Then we should instruct the AI to generate unfamiliar hypotheses.”
“An entirely reasonable experiment,” said Elowen. “But how would we recognise the unfamiliar possibilities that matter? We cannot simply instruct a system to produce novelty and assume that the result will be significant. Nor can we assess every possibility without regard to evidence.”
“Then we must devise better criteria for novelty.”
“And who devises those criteria?” she asked. “Researchers working within the same scientific practices that have already shaped their expectations?”
Blottisham looked dissatisfied.
“You seem determined to make every solution part of the problem.”
“Not every solution,” Elowen replied. “But solutions change the conditions in which further problems arise. That is why we must examine the whole process.”
Quillibrace leaned forward.
“Consider the difference between asking whether an individual AI-generated idea is good and asking how AI changes the distribution of ideas that science encounters. The first is a question of evaluation. The second concerns the organisation of scientific inquiry itself.”
“Distribution?” Blottisham asked.
“Which ideas are generated, which become visible, which attract attention, which are investigated, and which are ultimately accepted. These are different stages, and each can shape what follows.”
Elowen continued.
“An idea that is never generated cannot be evaluated. An idea that is generated but never pursued cannot contribute evidence. An idea that is investigated but cannot be expressed in an accepted form may never receive a fair hearing. And an idea that challenges the framework used to assess it may be judged inadequate precisely because it does not fit.”
“Surely,” said Blottisham, “you are not suggesting that scientific reviewers should accept ideas merely because they are unconventional.”
“Certainly not. Unconventional ideas can be mistaken. The point is that familiarity and merit are not the same thing.”
Quillibrace nodded.
“Nor does the problem require anyone to behave irrationally. Each participant may act reasonably according to the immediate task: the system produces plausible suggestions, the researcher selects promising ones, the reviewer applies established standards, and the journal favours work its readership can assess. Yet their combined activity may systematically narrow the range of possibilities science explores.”
Blottisham stared at him.
“You mean the entire process could produce an undesirable result even when every participant is doing precisely what is expected of them?”
“An unfortunate feature of organised activity,” said Quillibrace, “and by no means peculiar to science.”
Elowen glanced at the report.
“AI need not create this tendency. Scientific institutions already have incentives that can favour familiar approaches, fashionable questions and readily publishable results. AI may amplify those tendencies if its outputs and the systems that evaluate them reinforce one another.”
“Then AI is not the culprit,” Blottisham said.
“Not necessarily. It may be a contributing factor, an amplifier, or in some circumstances a source of genuinely different possibilities. We need to investigate what happens in particular ecologies of research rather than assume a single outcome.”
Blottisham appeared relieved.
“I thought for a moment that you were about to recommend banning artificial intelligence from scientific inquiry.”
“Would that restore every possibility that science has neglected?” Elowen asked.
“Probably not.”
“Would it remove the influence of established expectations, institutional incentives or disciplinary habits?”
“Not entirely.”
“Then a ban would not address the whole problem.”
Blottisham sat down again.
“Very well. We need a more comprehensive remedy. Perhaps a committee to ensure that every scientific field considers a sufficient number of unconventional ideas.”
Quillibrace sighed.
“Mr Blottisham, you have just proposed standardising the production of unconventionality.”
“I fail to see the difficulty. The committee could establish quotas.”
“Would the quota apply to the number of unconventional ideas generated, the number investigated, or the number accepted?”
“All three, perhaps.”
“And how would the committee determine whether an idea was genuinely unconventional without relying on the very expectations we are trying to examine?”
Blottisham opened his mouth, then closed it.
Elowen smiled sympathetically.
“The difficulty is not that we should abandon evaluation. It is that we must remain alert to the possibility that our ways of evaluating ideas can change what becomes available for evaluation in the first place.”
“And how,” asked Blottisham, “are we to remain alert to possibilities we cannot yet recognise?”
“That,” said Quillibrace, “is a much better question.”
For a moment, the room was quiet.
Elowen looked again at the report. “Perhaps we should ask whether our research practices provide ways for unexpected results to alter the questions we ask, the methods we use and the standards by which we judge the answers. We need not accept every challenge. But we must allow evidence and inquiry to change the conditions under which possibilities are assessed.”
Blottisham rose, apparently satisfied that the discussion had at last reached a practical conclusion.
“I shall recommend a committee to examine whether our committees are sufficiently open to possibilities they cannot yet recognise.”
Quillibrace watched him leave.
“Do you suppose he understands the problem?”
“Perhaps not entirely,” Elowen replied. “But he has begun to ask a different question.”
“And that,” said Quillibrace, “may be how a possibility first enters an ecology that has not yet made room for it.”
From the corridor came Blottisham's voice.
“One further matter! Should the committee include artificial intelligence?”
Elowen and Quillibrace exchanged glances.
“Only,” said Quillibrace, “if we are prepared to consider what its participation might change.”
A pause followed.
“And who,” Blottisham called, “will be responsible for deciding that?”
Quillibrace returned to his book.
“I believe,” he said, “we have our first agenda item.”
No comments:
Post a Comment