'Ship-then-patch' is becoming the default governance model for Big Tech's AI rollouts. Critically evaluate the adequacy of pre-launch safety testing for generative AI features.

'Ship-then-patch' means releasing a generative AI feature publicly and repairing harms afterwards. Google's withdrawal of its Google Earth image-generation tool within about a day of its July 2026 launch [1] shows the model is commercially efficient but inadequate where the underlying data is trusted as evidence.

Where the model has merit

• Scale reveals what labs miss: the flaw surfaced only when outside investigators generated fake satellite images of a nuclear plant and refugee camps, exposing gaps internal testing had cleared [2].
• Reversibility as a control: rapid rollback and staged "experimental" releases limit exposure when a defect appears [1].
• Avoiding over-caution: legitimate uses Google cited — heritage reconstruction, urban-planning and teaching mock-ups — would stall under indefinite pre-release testing [1].

Where it is inadequate

• Irreversible informational harm: fabricated imagery, once circulated, cannot be recalled; worse, it lets authentic satellite evidence be dismissed as AI-made, corroding open-source intelligence used in conflict and disaster verification [2].
• Self-certification: safety claims rest on internal red-teaming with no independent pre-deployment audit or published results.
• Guardrails bolted on, not built in: the company conceded it was adding "stronger guardrails" only after rollback [2].
• Externalised costs: detection burden shifts to journalists, researchers and regulators, not the releasing firm.

The regulatory correction underway

• India's IT (Intermediary Guidelines) Amendment Rules, 2026 require labelling and traceable metadata for synthetically generated information [3].
• The EU AI Act, Article 50, applicable from August 2026, mandates machine-readable marking of AI outputs and disclosure of deepfakes [4].
• NITI Aayog's Responsible AI framework similarly stresses accountability designed into deployment, not retrofitted [5].

Pre-launch testing is therefore necessary but presently insufficient — adequate for low-stakes creative tools, unsafe for features layered over evidentiary data. A calibrated path forward is risk-tiered clearance: mandatory independent red-teaming and provenance watermarking before release for high-trust domains, with lighter iteration elsewhere. Innovation and public trust are then complements, not trade-offs.