On 29 September 2026 the European Commission opened a targeted consultation on whether further EU copyright measures are needed for generative AI. It runs to 3 November 2026 and covers the use of copyright-protected content in AI systems and the application of copyright to scientific research, along with live-event piracy and music performers' remuneration, according to the Commission's announcement. For Estonia, a small digital-first state with a small-language problem, the answer matters more than the procedural label suggests.
The case for tightening the rules
The strongest argument for new binding measures is a fair one. Under the 2019 DSM Directive, a commercial AI developer may mine lawfully accessible works unless the rightsholder has reserved its rights. That puts the burden on creators, many of whom are individuals or small organisations with no technical team to deploy machine-readable reservations. A system that is lawful in theory can fail in practice if opt-outs are unevenly honoured, hard to detect, or applied after a model has already been trained. Rightsholders also argue, reasonably, that AI outputs can compete with the works that trained them, and that voluntary licensing is only as strong as the legal leverage behind it. Regulators who ask whether the 2019 compromise still fits 2026 technology are doing their job.
What Estonia actually has
Estonia transposed the DSM Directive's two text-and-data-mining (TDM) exceptions into its Copyright Act. Reproduction of a lawfully accessible work for TDM is permitted without consent or royalties, and for uses outside scientific research the author may expressly exclude it. Estonian rightsholders already use this. The Estonian Actors' Association, for instance, publishes a notice excluding the performances of the actors it represents from TDM for AI development outside scientific research, and says anyone wanting to mine them needs its prior written licence. That is the opt-out system working as designed: a collective body reserved its rights in public, and a would-be developer knows where to go.
This matters for the consultation because it is evidence. Brussels is asking whether the current framework needs binding reinforcement. Member states with working reservation practice can show that the existing rule produces identifiable licensing counterparties without a new mandate.
The Estonian-language wrinkle
The Ministry of Justice and Digital Affairs has been preparing an amendment under which publicly available Estonian texts that carry no opt-out mark could be read by AI for research and development. As ERR News reported in March 2025, Justice Minister Liisa Pakosta framed AI "thinking" as comparable to the way researchers think, and Cultural Affairs Committee chair Liina Kersna argued the state has a constitutional obligation to ensure AI speaks Estonian well, noting that closed models are significantly more accurate in English than in Estonian. ERR said the amendment's submission date was undetermined, and we have found no confirmation that it has since been adopted.
The Kersna point is the heart of the innovation case. Estonian is spoken by about one million people. The corpus of high-quality Estonian text is tiny compared with English, so every barrier to lawful use falls disproportionately on a language community that cannot out-publish its way to model quality. If compliance costs make Estonian-capable models uneconomic, the result is not stronger creators' rights. It is weaker Estonian-language AI and heavier reliance on foreign systems trained on whatever Estonian text happened to leak into English-centred datasets.
Where the consultation could go wrong
The risk is a Europe-wide fix designed for the largest creative markets and applied to everyone. Three possible directions deserve scrutiny.
- Mandatory remuneration or licensing for training. This would turn a reservable exception into a payment obligation, favouring the few firms able to negotiate with every rightsholder and shutting out startups, universities and small-language projects.
- A narrowed research exception. The consultation explicitly lists copyright in scientific research. The Commission's own page points to a July 2026 feasibility study on an EU-level registry of TDM opt-outs, which suggests the Commission is weighing infrastructure rather than prohibition. A shared registry could help both sides if it is cheap, optional to use, and does not make registration a condition of protection.
- Vague transparency duties. Disclosure of training sources can be valuable, but only if it is specific enough to be audited and light enough for small developers to meet.
Of these, the registry is the most promising. It addresses the real defect in the current system, which is discoverability of reservations, without changing the underlying balance.
What Estonia should say
The consultation closes 35 days after it opened, so national governments and Estonian stakeholders have a short window. A sensible Estonian position has four parts.
- Defend the existing opt-out structure as proportionate: it has produced concrete reservations, such as the actors' notice, without blocking research.
- Support standardised, machine-readable reservation signals and an optional EU registry, so that opting out is easy and honouring opt-outs is verifiable.
- Protect the research exception, since universities and language institutes are the ones building national-language resources.
- Tell Brussels plainly that any new obligations must carry a small-language and SME carve-out or lighter regime, and that Estonian-language data policy is a cultural-sovereignty issue, not a loophole.
The domestic amendment should be judged on the same test. Letting AI read unmarked public Estonian text is defensible only if the opt-out is real: easy to apply, clearly signalled, and honoured. That also requires care over what counts as "publicly available" and as an acceptable mark, since a rule that treats silence as consent invites exactly the unevenness that critics of the opt-out model worry about.
The broader point
The DSM Directive's bargain was imperfect but coherent: research gets a floor, commercial users get access, and creators get a veto. Evidence that the veto is hard to exercise justifies better tools, not a rewrite. Estonia has both a stake in cheap, lawful access to its language data and a creative sector that has shown it can use the veto. That makes it a good witness for a calibrated outcome. If the Commission concludes the current rules need strengthening, the proportionate path is to improve reservation infrastructure and transparency while leaving lawful access intact. The alternative is a licensing regime built for global giants that Europe's smallest languages cannot afford.