On AI: Memory, Identity, and Lock-In

As generative AI reaches a fever pitch of investment, product releases, and hype, most of us have ignored a profound flaw as we march relentlessly toward The Next Big Thing. Our most dominant AI products and services (those from OpenAI, Google, and Microsoft, for example) are deployed in the cloud via a “client-server” architecture — “a computing model where resources, such as applications, data, and services, are provided by a central server, and clients request access to these resources from the server.”

Now, what’s wrong with that? Technically, nothing. A client-server approach isn’t controversial; in fact, it’s an efficient and productive approach for a company offering data-processing products and services. The client — that’s be you and your device — provides input (a prompt, for example) which is relayed to the server. The server takes that input, processes it, and delivers an output back to the client.

Non-controversial, right? Well, sure, if the “server” in question is a neutral platform that’s only in the business of processing your data so you can use the services it offers. Banks, for example, use neutral client-server architectures to provide online financial services, as do most health care providers. The data you share with them isn’t used for anything other than the provision of services.