One of the newly dominant questions of our age is the degree to which, or even whether, human judgement should take precedence over machine-generated output across different fields. The existence of this debate is itself an admission of the unusual reasoning capabilities now exhibited by some forms of machine-generated analysis, particularly with the rise of advanced AI tools such as large language models. Rather than asking which form of insight should take precedence over the other, this development should provoke a different way of thinking: that “cognitive cooperation” between human and machine intelligence will be required to produce both intellectual and practical breakthroughs across a wide array of fields in the future. To understand why this is the case, it is useful to return to the historical debate over the relationship and supposed competition between natural and artificial processes, a debate steeped in the intellectual developments of the seventeenth century.
In his Novum organum, published in 1620, Francis Bacon, the English polymath, produced the epochal statement that symbolised a radical epistemological change in how knowledge about the natural world could be conceived, one which still resonates strongly with our time:
“…the secrets of nature reveal themselves more readily under the vexations of art than when they go their own way.”
With “art” in the above statement meaning artifice or experiments, Bacon, often dubbed the “father of experimental philosophy”, was part of a wider network of scholars who, in the period of the so-called scientific revolution, helped establish the validity of artificially constructed experiments as a means of acquiring empirical knowledge of the natural world.
The rise of experimentation was one of the distinctive features of this transformation in scientific method and remains central to scientific practice today. But the elevation of experimentation also formed part of a wider seventeenth-century debate over the relationship between “natural” processes and “artificial”, human-constructed ones. In a different form, that debate has resurfaced with greater existential stakes in the age of artificial intelligence, particularly with the recent dominance of large language models. Essentially, the spirit, if not the letter, of the seemingly “modern” debate over the supposed hierarchy between “human” or natural processes and artificially produced or machine-generated ones has its intellectual antecedents in the seventeenth century. By affirming that knowledge derived from experimentation mattered in interpreting the physical qualities of nature, the experimental philosophers challenged one of the fundamental distinctions of Aristotelian philosophy, the dominant paradigm of knowledge in medieval natural philosophy.
Aristotle was himself known to perform experiments, such as establishing through filtration and distillation that saltwater was not the natural state of water. Yet his clearest views on the epistemic status of art and experiments were evident in his Physics, where he wrote that “generally art partly completes what natures cannot bring to a finish and partly imitates her”. The implication was that natural processes possessed a superiority that human art or technology could only imitate or supplement. Bacon opposed this tradition. In the Descriptio globi intellectualis, he directly attacked the notion of “art … solely as a kind of supplement to nature; that is, its power is this, that it can either finish what nature has begun or put it right when it has gone astray, but can in no way alter it radically and shake it to its depths”. Bacon described this view as a “subtle evil”, which “has brought profound despair into human affairs”.
This seventeenth-century dispute resonates strongly with the contemporary division over the epistemic status of sophisticated large language models. There are some, like Microsoft AI CEO Mustafa Suleyman, who, in a recent essay, describe AI models primarily as mechanical instruments, rather like Aristotle’s conception of art as something that “partly imitates nature”. In a similar vein, Simon Baron-Cohen, the renowned psychologist, at a recent event compared LLMs to “machines spitting out answers based on pattern recognition” and described them as “no different to hand calculators”. Then there are others, such as Dario Amodei, Anthropic’s CEO, who has repeatedly called for a slowdown in the development of frontier AI systems because of their potentially transformative risks. His position leans more strongly towards the Baconian possibility that artificial technology can alter nature radically or “shake it to its depths”.
What is evident from recent developments, however, is that large language models both “imitate nature”, in the sense that artificial neural networks are modelled on natural neural processes, and have displayed a capacity to “shake nature to its depths” that goes well beyond that of passive mechanical tools. Their unusual reasoning abilities have already triggered wider debates over machine intelligence, consciousness and the future relationship between humans and machines. Just as in the seventeenth century, what is now required is a new way of experimentation, one whose bedrock rests on “cognitive cooperation” between humans and machine-generated analysis, rather than simply on human insight guiding passive AI tools.
A fascinating aspect of this is that, unlike the barometer or the air pump, which were functional tools that pioneered experimentation in the seventeenth century, large language models possess internal neural networks that are difficult to understand. Many of the engineers who develop these models have acknowledged that they do not fully understand how the models “reason”. Unlike traditional computer programs, these networks are trained or cultivated rather than programmed in strict algorithmic detail.
This opacity has created the emerging field of mechanistic interpretability, which attempts to discover and understand how large language models work internally. Although recent work by interpretability pioneer Nick Cammarata, focused more extensively on vision models than on language models such as ChatGPT, suggests that neural networks may be less opaque than previously thought, it does not lessen the need for interpretability research. The field may yet prove immensely important given its relevance to AI safety. Anthropic’s research into one of Claude’s internal reasoning spaces, known as the J-space, offers another example of what this field can reveal.
Inspired by global workspace theory in neuroscience and reviewed by the neuroscientists who developed this theory, Anthropic has suggested that Claude has developed a small collection of internal neural patterns that functions as a mental workspace, allowing the model to process concepts without necessarily expressing that processing in its final output. In its experiments, intermediate stages of reasoning became active in the J-space even when they were absent from Claude’s final answer. The discovery of this mental workspace is very important because it offers a remarkable insight into how Claude “reasons”, even when the process is not immediately visible to its lay users. This window into Claude’s internal reasoning process could potentially serve as a bedrock for the future alignment of LLMs with ethical human values.
Crucially, Anthropic stated that the J-space was not designed or programmed by its engineering team but instead emerged on its own during training. The significance of this research lies not in proving that Claude is conscious, something Anthropic itself explicitly rejects, but in revealing that there remain aspects of large language models that are yet to be fully understood and which may radically reshape our understanding of both human and machine intelligence. Anthropic’s J-space experiments show Claude working through intermediate steps of reasoning that do not appear in its final answer, a capacity the barometer and air pump did not possess in the seventeenth century.
It is this possibility that makes collaborative intelligence a useful way to understand the role of LLMs in scientific inquiry. Indeed, over the past decade, LLMs have proved to be among the most accurate tools for understanding human intelligence, for example by predicting how the human brain responds to language. Recent work in generative causal testing at Microsoft, for example, has used models to distil brain-prediction systems into short verbal explanations of what particular regions of the cortex respond to.
These examples are not simple cases of human insight interpreting machine-generated analysis from passive, non-reasoning mechanical tools. Rather, they are instances of “cognitive cooperation” between human scientists and advanced reasoning models working together to explicate aspects of nature or, in this case, the human brain that have not yet been fully understood. This is broadly in agreement with the Baconian formulation that “the secrets of nature reveal themselves more readily under the vexations of art than when they go their own way”. Dario Amodei similarly described this form of intellectual co-reasoning in his Machines of Loving Grace essay, where he imagined LLMs functioning as research assistants capable of contributing cognitively throughout the full research cycle and potentially accelerating scientific discovery.
The human future, especially with the rise of artificial intelligence, will, like the past, prove to be a “foreign country”. Things will certainly be done differently there. But in navigating that future, it is important to adopt the spirit of the intellectual revolutionaries who elevated experimentation as a hallmark of scientific practice in the seventeenth century. Their achievement lay partly in recognising that artifice or experiments could reveal truths about nature that unaided observation could not. We need a similar openness today by not treating machine intelligence as opposed to human insight, but recognising the possibility of genuine “cognitive cooperation” between the two. As AI systems become more capable, supporting pioneering research areas like interpretability will become increasingly important in helping to mitigate the immense risks that come with the revolutionary possibilities of this technology.
