Part one asked who owns what an AI produces. This part asks a prior question: how would we know the system is intelligent at all — and if we cannot say, on what basis would we grant it rights?
In part one we covered the division between narrow and general intelligence, ANI’s influence on copyright law, and the complexities arising when ANI generates copyrightable works.
The trouble with testing intelligence
To validate what constitutes AI, we need a yardstick. The available ones do not hold up.
The Turing test, prominent since the 1950s, is outdated and parochial — limiting examination to communication and speech, which chatbots satisfy without difficulty.
Lady Lovelace’s test proposes that a machine is intelligent if it can surprise the observer of its outcome. That has been criticised as superficial, and the criticism is well founded. A machine may surprise us because of a bug in the software or hardware. And the human mind is subject to cognitive bias — with a human preparing and examining the program, there is ample room for an unintended faulty understanding to appear as surprise where none is warranted.
There is a related phenomenon: outputs that surprise the programmer because of an incomplete understanding of the software or its fundamentals — the black box. The programmer’s inability to understand how the system functions should not be the basis for attributing intelligence to it.
That is the sharpest point in this piece, and it has aged well. Opacity is routinely offered as evidence of sophistication. It is equally consistent with a system nobody has adequately examined.
The question that follows
Setting aside our limited understanding, ANI is sufficiently adopted that accountability for legal violations must be placed somewhere. But given the range of activities ANI performs in any given field, a blanket liability would constrain its development and growth.
And there is a characterisation problem underneath. Any tool presently claimed as AI is arguably advanced software with higher computational abilities, running on advanced hardware. If it is advanced software — should it receive copyright protection at all?
Where the reasoning leads
Assume for argument that an ANI is granted copyright over a literary work and earns from it. Could it then purchase cloud storage, replicate itself, produce further works and claim copyright over those?
There is no conclusive answer, and such questions would fall to be decided case by case. The difficulty for courts is keeping pace with a fast-moving technology, while a rigid approach ignores its significant impact on the economy and society.
What is not viable is unregulated adoption, given ANI’s capacity to generate material that infringes intellectual property. Regulation placing appropriate liability on the use and adoption of these systems is needed.
A framework for liability
The proposal: the ANI or its programmer should be answerable for the logic and rationality the system applied in creating the work. That would help courts assess whether liability arises from a poor training dataset or bad algorithms, and whether the explanations offered are bona fide.
Understanding the basic functionality of these systems is a precondition for deciding such matters properly.
The conclusion follows from the premises: ANI is advanced software with higher computational abilities, which may not amount to intelligence as such — and since the definition of intelligence is itself unclear, granting it copyright is not viable.
Legislation
What is needed is comprehensive legislation governing the aspects of these systems — in time, as inevitable as the Companies Act. We will require an Artificial Intelligence Act.
The right starting point is an inclusive definition of AI, drawn broadly enough to cover new functionalities without requiring frequent amendment. That would let courts and government place wider accountability on adoption, without leaving gaps for developers to exploit — while recognising that the growth of AI matters, and stringent law should not become the obstacle.
Where this has gone. The prediction was sound. The EU AI Act was adopted in 2024 as the first comprehensive regime, taking the risk-based approach this piece anticipates rather than attempting to define intelligence — sidestepping the measurement problem entirely by regulating use and consequence instead. India has proceeded through advisories and sectoral guidance rather than a dedicated statute.
The training data question this piece identifies — liability arising from a poor dataset — has since become the central issue in AI copyright litigation worldwide, though framed as whether training on copyrighted works infringes rather than as a question of the resulting output.
The takeaways
- The Turing and Lovelace tests are inadequate for identifying machine intelligence.
- A programmer’s incomprehension is not evidence of intelligence — the black box problem.
- Liability should attach to the logic applied — poor training data or bad algorithms.
- The EU AI Act took the risk-based route, regulating use rather than defining intelligence.
Frequently asked questions
Why is it hard to test whether a machine is intelligent? The Turing test only examines communication, which chatbots satisfy; the Lovelace test relies on surprising an observer, which a software bug can also achieve.
What is the black box problem? Where a system’s output surprises its own programmer because of incomplete understanding of the software — which should not be treated as evidence of intelligence.
Should AI-generated works receive copyright? On this analysis no — if the system is advanced software rather than demonstrably intelligent, and intelligence itself is undefined, there is no principled basis for protection.
Does India have AI legislation? Not a dedicated statute — India has proceeded through advisories and sectoral guidance, while the EU adopted a comprehensive risk-based AI Act in 2024.
Useful official resources
- The Copyright Act, 1957
- WIPO — Artificial Intelligence and IP