Statutory researcher access and legal-shield advocacy

Organise for statutory access rights that compel platform and AI-vendor disclosure of models, APIs, training-data metadata, evaluation infrastructure, and deployment logs to vetted independent researchers, and for parallel legal shields — safe-harbour provisions, computer-fraud-and-abuse-act carve-outs, terms-of-service anti-enforcement protections — that immunise researchers from civil and criminal liability for good-faith adversarial testing of production systems. The vehicle is EU DSA Article-40 implementation-and-expansion work, national researcher-access statutes modelled on the DSA, US safe-harbour amendments to the CFAA and DMCA anti-circumvention rules, and platform-specific access agreements civil society negotiates as a condition of the vendor's regulatory posture.

An actor chooses this strategy because independent evaluation is the substrate of every other AI-good strategy — audit, litigation, journalism, red-teaming, and rulemaking-comment work all depend on outside actors' ability to measure what a deployed AI system actually does — and every one of those measurement channels is currently blocked or made legally hazardous by the vendor's terms-of-service, API access-controls, and threatened liability under criminal-hacking statutes. Statutory researcher access converts a fragile permission the vendor can revoke at will into a right the vendor cannot deny, and the corresponding shield converts good-faith adversarial testing from a criminal-exposure situation (the Sandvig-v-Barr trajectory, the Aaron Swartz lineage) into a protected professional activity. The strategy underwrites the measurement layer the rest of the movement's evidence base rests on.

It trades speed for structural yield. Statutory access rights are hard to win — the DSA took a decade of European civil-society work before Article 40 landed with any bite, and its implementation is still contested — and once won they are exposed to executive-branch under-implementation, vendor litigation, and pre-emption. The shield side is even harder: legislators are risk-averse on measures the tech industry frames as pro-hacking, and researcher-shield law in the US has moved primarily through case law (van Buren, Sandvig) rather than statute, leaving the shield jurisdiction-dependent and often uncertain at the moment a researcher needs it. The strategy also depends on a class of civil-society actor — independent platform researchers, algorithmic-audit organisations, academic labs — whose funding base is precarious and whose numbers globally remain small.

This strategy differs from strat-empirical-audit-and-expose by layer of the pipeline — audit-and-expose produces the finding, researcher-access-and-shield secures the right and safety to gather the evidence the finding rests on — and from strat-freedom-of-information-as-evidence-gathering by target actor (private platforms vs. public bodies). It is upstream of most other measurement-dependent strategies in the corpus and is the strategy the AI-good movement would notice missing most sharply if the shields it currently rests on quietly narrowed.