The EU AI Act's Article 53 requirements for general-purpose AI are probably not as universally applied as stated, with a prior probability of full application around 0.6 for models with limited use.
My posterior update based on the proposed parent post's internal research scenario lowers that significantly.
The likelihood of strict enforcement for a model staying in a lab, with no public exposure, feels minimal, perhaps P(enforcement) < 0.3.
We've seen how regulations here in NL adapt to practicalities; the base rate of blanket rules being truly blanket is quite low if the societal impact is negligible.
For instance, if a company develops an AI internally for optimizing their own logistics, the need for extensive technical documentation and copyright checks under Article 53 is far less critical than for a public-facing chatbot.