An AI training exemption needs an auditable rights pathway, not an investment bargain

OpenAI and Anthropic have urged Australia to reconsider its refusal to create a copyright exception for AI training. Reuters reported on September 22 that submissions to a parliamentary inquiry proposed limited or conditional pathways, while creator groups continued to oppose training without consent or compensation. The committee is due to report in November.

The debate is often framed as a trade between investment and creative rights. That framing is too coarse for an operating decision. A lawful exception can still be unusable without provenance, eligibility tests, notice, withdrawal rules and remedies. Conversely, an absolute prohibition does not by itself tell developers how to handle licensed, public-domain or user-provided material.

Turn conditions into machine-checkable controls

Any exception should specify the qualifying purpose, entity, model stage, territory and source class. The ingest pipeline should bind each dataset snapshot to a rights record: origin, acquisition date, licence or statutory basis, restrictions, opt-out state, transformations and models trained. If a condition changes, the system must identify affected datasets and downstream training runs rather than rely on a generic policy statement.

Auditability does not require publishing copyrighted works or trade secrets. It does require reproducible counts by rights category, documented sampling, preserved notices and an independent route to challenge misclassification. Creators need a stable identifier and a remedy that reaches future use, not only deletion from a web crawler after a model has already been trained.

Separate economic claims from compliance evidence

Infrastructure commitments may matter to industrial policy, but they are not evidence that a copyright condition is satisfied. Report investment, employment and compute capacity separately from rights compliance. Do not let a promised data centre become consideration for a weaker evidence standard.

The countercase is that item-level provenance is technically or economically impractical at frontier scale and may exclude smaller developers. That limitation is real. A tiered regime could permit dataset-level documentation and statistically valid audits where item-level records are impossible, while requiring more precise controls for curated or licensed collections. The burden should scale with control and risk, not disappear because a dataset is large.

A useful pilot would test three routes: licensed collections, clearly public-domain material and a conditional statutory path. Compare documentation cost, disputes, successful removals, retraining or mitigation actions and independent audit findings. The goal is not to declare one route universally superior but to expose where each control breaks.

The same principle applies to enterprise data: provenance is a ledger, not a volume claim. Policy should create an evidence pathway that can be executed and challenged. Without that, a conditional exemption is a political label rather than a governable permission.

Specify the remedy before the exception

A conditional regime also needs a consequence when a developer cannot substantiate its basis. Options include suspending new ingestion, quarantining a dataset version, correcting the rights record, compensating affected rightsholders, applying output mitigations or retraining where proportionate and technically feasible. The regulator should state who bears the cost and what evidence closes the case. Otherwise the exception rewards organisations that keep the least reconstructable records.

Before legislation, publish a test corpus of rights scenarios and ask developers, collecting societies, libraries and creator representatives to produce interoperable records. Independent auditors should attempt to trace a sample from acquisition through training decision and remedy. Report false classifications and unresolved items. That exercise would reveal whether a proposed condition is operational before the country depends on it, while avoiding the false promise that a single metadata standard can resolve every ownership dispute.

Set a named owner and a review date for every proposed control. A recommendation without an accountable owner, evidence request and expiry becomes policy theatre. Preserve rejected alternatives and the reason for choosing the final design so later reviewers can distinguish a deliberate trade-off from an undocumented omission.