Is agreeing philosophical principles on data, copyright and AI necessary?

I was at parliament this week discussing AI and the Data (Use & Access) Bill. Thanks Baroness Beeban Kidron, Tim Flagg & UKAI for the invitation. A confession:

I didn’t recognise Björn Ulvaeus. Apparently, he’s a bit of a big deal. Even without being a fan, I thought he spoke very eloquently, including the comment that:

“Artists don’t fear AI, they fear losing their [livelihoods] because of AI”

His words illustrate my own observation: as important as philosophical discussions are, artists and other rights-holders will only be convinced about any changes to UK copyright rules, if they are shown a real, positive impact to their individual livelihoods or their organisation’s revenue.

Wednesdays’ announcement of an “economic assessment” (https://lnkd.in/esgxvrWz) is an interesting step. Some thoughts below on what more the UK government could do.

Quality training data are vital for machine learning, but getting access is often difficult, especially for early-stage companies like ones I work on.

Government support here is positive, but while it reduces cost & complexity for tech companies, weakening copyright would harm other rights-holders financially.

So, what’s the alternative? Well the UK government aren’t just legislators. They are also data providers (https://datalibrary.uk/); with the ability to test access and ‘licensing’ frameworks for AI / ML use without the financial risks that make rights-holders cautious.

They could work on model agreements and payment mechanisms, using their own data to show tech companies and artists / rights-holders how they could work in practice. This would be even more useful than an economic impact assessment on its own.

To be more specific, someone asked Tim Flagg, “what do UKAI’s members want?”. My answer would be:

• ➡️ Simple and easy access to data
• ➡️ Clarity on data quality (and value) before investing
• ➡️ Freedom to experiment
• ➡️ Certainty on future access (i.e. no risk of unilateral removal)
• ➡️ Flexibility to pursue different commercial paths

For balance, the equivalent for rights-holders (I also work with data providers) is an agreement that allows them to:

• ➡️ Measure the value of their data for model training and be paid for it
• ➡️ Ensure data provision is a long-term relationship not a single transaction
• ➡️ Understand downstream use of a model, its impact and be paid for it. (The Guardian gave a good example of LLM summaries reducing site traffic)