Over 90 Percent Refuse, AI Gets Used More Anyway: What the DGPh 2026 Survey Leaves Open

The German Society for Photography (DGPh), together with Hochschule Hannover, has published the eighth wave of its survey on Germany's image market, and the numbers show a divide that's widening rather than closing: the share of editorial teams planning to use AI-generated images or video has risen from 46.2 to 56.6 percent. At the same time, more than 90 percent of surveyed photographers and image agencies refuse to make their own work available for training AI models.

The question the study itself leaves open

Placed side by side, these two figures raise a question the report never answers: if more than 90 percent of professional image creators refuse to license their work for AI training, where does the training data for the models an increasing number of editorial teams want to use right now actually come from? The study documents the divide precisely, but leaves open what's actually happening on the other side of it – a gap that belongs in the assessment just as much as the numbers themselves.

The uncomfortable answer

The honest answer is more uncomfortable than the question. Most of what current image models "know" was likely absorbed before there was any way to consent to it or object to it in the first place – at a point when neither a licensing structure nor organized refusal existed. Today's refusal by more than 90 percent is real and understandable, but it mostly affects future, additional licensing – not retroactively what an already-trained model carries within it. That's the same boundary that showed up in the GEMA ruling against Suno: a court can find that a training process was unlawful without automatically changing what an already-existing model has already learned.

Why the refusal still isn't meaningless

That puts the refusal in perspective without making it pointless. For new model versions, for current work only now being created, and as a bargaining position toward platforms that depend on fresh, high-quality images, collective refusal remains real leverage – just leverage that works forward, not backward. Anyone refusing today as a photographer or agency doesn't undo what existing models have already learned, but does help determine what the next generation of models gets to build on.

What this means for our own practice

For studios like ours, working between classic photography and AI-assisted production, the practical consequence has two sides. Your own signature can't be protected by blanket refusal of AI training – that ship has largely sailed for models already trained. It's better protected the way I already described regarding tool independence: through a result that stays recognizable regardless of which model happens to be in use – not through the hope that any single tool stays untouched by someone else's training data.

Conclusion

The DGPh 2026 survey shows a market pulling apart in two directions at once: more demand for AI images, more refusal to supply the data that trains them. But the actually interesting finding lies in what the study doesn't say – that this refusal can't pull back a model that's already trained. Ignoring that means fighting a battle that was already decided somewhere else. Acknowledging it means at least putting your own position to use for what hasn't been decided yet.