Abstract
This paper starts from section 2.6 of the Stanford Encyclopedia of Philosophy's Turing Test entry, where Lovelace's objection is set against Turing's reply. The debate is often framed as a question about whether machines can create ex nihilo. That framing is too strong. The useful question is whether machines can exhibit originativity: novelty relative to a reference class, value in a domain, causal independence from mere replay of human instructions, and adaptive revision under feedback. Lovelace's famous claim that the Analytical Engine "has no pretensions to originate anything" is best read as a sourcehood claim, not a proof of impossibility. Turing's reply, especially his appeal to surprise and learning machines, points in the right direction, but surprise is observer-relative and therefore not enough on its own. Contemporary creativity theory and current AI evidence support a graded conclusion: machine creativity is task-relative, empirically testable, and not captured by a binary can/cannot-think test. Machines can think in the functional sense relevant to this debate, while consciousness and moral status remain separate questions.
Keywords: machine intelligence, creativity, Lovelace objection, Turing test, computational creativity
This paper builds from section 2.6 of the Stanford Encyclopedia of Philosophy's Turing Test entry, where Lovelace's objection is set against Turing's reply. The Lovelace-Turing debate survives because it compresses several questions into one: can machines generate genuinely new artifacts, can they learn, and can they think at all? Section 2.6 makes the key issue clear: Lovelace's objection is usually summarized as the claim that machines can only do what we know how to order them to do, while Turing's reply is that machines can surprise us and that learning machines complicate the picture (Oppy and Dowe 2021). That is the right starting point, but it is not yet a resolution.
My claim is simple. Lovelace was right that machine creativity should not be treated as magic. Turing was right that a machine's inability to originate something ex nihilo does not show that machines cannot think. The mistake is to treat "origination" as a yes/no property. A better account is graded: a system can be more or less originative depending on whether it produces outputs that are novel, valuable, causally autonomous, and able to change with learning.
That matters because the question is not whether machines can duplicate human imagination. It is whether they can generate outputs through causal organization that is sufficiently independent of simple replay to count as originative. That is a tractable empirical question.
Lovelace's famous line that the Analytical Engine has "no pretensions to originate anything" is a claim about sourcehood. Read carefully, it does not deny that the machine might transform science or produce downstream novelty; it denies that the machine is the autonomous source of what it outputs. Turing's response shifts the frame. He treats novelty as relative to an observer and points to learning systems that become less predictable even to their designers (Turing 1950).
That shift is decisive. If "novelty" means "new to the designer," then many systems already qualify. If it means "new in the history of the world," then almost nothing does. The problem is to avoid both extremes. We need a criterion that can be tested across systems, tasks, and evaluators.
Call the relevant property originativity. A system is originative when it produces outputs that are:
- novel relative to a relevant reference class,
- valuable or adequate within a domain,
- not a simple replay of stored human instructions or examples, and
- capable of adaptive revision under feedback or self-modification.
Formally:
Originative(S, D, R) = Novel(S, D, R) && Valuable(S, D) && CausalIndependence(S, R) && AdaptiveRevision(S, D)
This is not a reduction of creativity to a single number. It is a way to make the question measurable. A Turing-style imitation game, a Lovelace-style originality challenge, and a psychology-style creativity benchmark each sample different parts of the construct. No single test can settle the issue. A serious conclusion requires convergent evidence from multiple tests and interventions.
Following Boden, it is useful to distinguish historical creativity from psychological creativity, or P-creativity (Boden 2004). The Lovelace-Turing debate is mainly about P-creativity: whether a system can generate outputs that are new relative to its own history and inputs. On that standard, machine creativity is already real in some domains.
Recent empirical work is mixed, which is exactly what one should expect if creativity is graded rather than binary. Koivisto and Grassini found that humans still outperformed current AI on one creative divergent-thinking task (Koivisto and Grassini 2023). Hubert, Awa, and Zabelina found the opposite on another divergent-thinking setup, reporting that current generative language models were more creative than humans on those tasks (Hubert, Awa, and Zabelina 2024). The mixed results do not cancel each other out; they show that creativity depends on the task, the metric, and the reference class.
That is the key point. The question is not whether machines are universally creative. The question is whether some systems, in some settings, can generate novel and valuable outputs through sufficiently autonomous causal organization. On current evidence, the answer is yes.
This also explains why the strongest form of Lovelace's objection is no longer tenable as a universal impossibility claim. A machine can be genuinely sourceful without being source-free. Its novelty may be distributed across model design, training data, prompting, and feedback, but distributed sourcehood is still sourcehood.
The lower bound for machine research is not "prove something no model can prove." That standard is too narrow and too brittle. Machines will prove correct things, boring things, and wrong things. The human research skill is upstream: choosing the right problems, defining them well, and judging which results matter.
So the Lovelace-Turing debate resolves in a qualified Turingian direction. Machines can think in the functional sense relevant here because thinking is an organized causal process that can be realized, at least in part, by machines. That does not settle consciousness or moral status. It does settle the narrower question in section 2.6: machines can do more than replay what humans already know how to tell them.
Bringsjord, Selmer, Paul Bello, and David Ferrucci. 2001. "Creativity, the Turing Test, and the (Better) Lovelace Test." Minds and Machines 11(1): 3-27.
Boden, Margaret A. 2004. The Creative Mind: Myths and Mechanisms. 2nd ed. London: Routledge.
Hubert, Kent F., Kim N. Awa, and Darya L. Zabelina. 2024. "The Current State of Artificial Intelligence Generative Language Models Is More Creative than Humans on Divergent Thinking Tasks." Scientific Reports 14: 3440.
Koivisto, Mika, and Simone Grassini. 2023. "Best Humans Still Outperform Artificial Intelligence in a Creative Divergent Thinking Task." Scientific Reports 13: 13601.
Kronfeldner, Maria. 2021. "The Freedom We Mean: A Causal Independence Account of Creativity and Academic Freedom." European Journal for Philosophy of Science 11, article 58.
Lovelace, Ada. 1843. "Notes by the Translator." In Luigi Menabrea, Sketch of the Analytical Engine Invented by Charles Babbage, Esq. Scientific Memoirs 3: 666-731.
Oppy, Graham R., and David L. Dowe. 2021. "The Turing Test." Stanford Encyclopedia of Philosophy.
Paul, Elliot Samuel, and Dustin Stokes. 2023. "Creativity." Stanford Encyclopedia of Philosophy.
Riedl, Mark O. 2014. "The Lovelace 2.0 Test of Artificial Creativity and Intelligence." arXiv 1410.6142.
Ritchie, Graeme. 2007. "Some Empirical Criteria for Attributing Creativity to a Computer Program." Minds and Machines 17(1): 67-99.
Turing, Alan M. 1950. "Computing Machinery and Intelligence." Mind 59(236): 433-460.