How Much will AI Regulation Cost?
The European Union's AI Act, in the last stages of approval, will impose an array of obligations on companies, much as the General Data Protection Regulation (GDPR) did for personal data.
In 2021, a preliminary analysis for the EU estimated industry compliance costs at around €29,000 for each model, or €1.6 and €3.3 billion in total annually (assuming 10% of systems fall in the "high-risk" category). However, that doesn't include performing the required conformity assessments, which the study estimated at around €20,000 per model for an external audit and considerably more with a full internal process.
The Center for Data Innovation, a US-based industry funded thinktank, estimated around the same time that the AI Act would impose compliance costs of around $6 billion per year. The EU vehemently objected to the analysis, but all-in, it doesn't sound substantially higher than its own numbers.
I suspect both sets of numbers significantly under-estimate the total investment firms will make in people, processes, training, and systems. And they likely represent only required compliance costs for companies operating in Europe, not the much greater scope of follow-on compliance activity worldwide.
Are the costs too high? Well, that depends on the other side of the ledger. How do these compliance costs match up with the financial returns firms get from their AI investments? And how great are the benefits of the protections the regulations will achieve? What direct and indirect costs will be avoided because firms invest up front in systems and processes of accountability due to regulation? These are difficult questions to answer, especially at this stage of market development. But they are important to consider. It's not helpful to ignore the costs, nor on the other side, to make sweeping statements that regulation will excessively chill innovation.
If nothing else, we should be asking whether there are ways to achieve the important goals of AI regulation with similar effectiveness at lower cost. Regulation isn't an either/or for something as broad as AI; it's a set of options along multiple dimensions. In the coming years, we will see exploration of many paths to AI regulation. Let's hope those involved are open to learning from the results.
领英推荐
If you enjoy my thoughts about AI regulation and policy, please subscribe to my new Substack on Accountable AI: https://accountableai.substack.com/