Identities & Personalities in AI Are Not the Same Thing. Both Have a Job.

Two movements are arriving at the same time and being described with the same words. Enterprises are starting to give AI agents identities of their own: credentials, permissions and lifecycle controls, so that an action in a system can be traced to the agent that took it. Consumer products are going the other way, shipping named assistants with a consistent voice and a recognisable character. Both get called "giving the AI a self". They are different things. An identity answers questions an auditor asks: which agent acted, who authorised it, which model was underneath. A personality answers questions a user asks without noticing: how does this thing talk, what does it do when it is unsure, will it still be the same next week. We have written about each before, in issue 006 on verifiable identity and issue 008 on character. This issue is about keeping them apart.

Identity earns its place through accountability. If an agent has its own credential, its access can be logged, limited and revoked without touching anyone else's. The caveat is that an identity is only as good as what it is bound to. In our first Model Disclosure Index census (SILT-RP-007, September 2026), the model field that every standard client reads returned, on three commercially available composed systems, a name that belonged to no model at all. One system answered a short question with three calls to two different companies' models and disclosed it only outside the standard interface. An agent credential on top of that tells you which account acted. It does not tell you what answered. Personality earns its place through legibility. A consistent character is something a user can learn, the way you learn when a colleague is guessing. Its caveat runs the other way: a confident, pleasant voice gets believed. On the Code Integrity Battery, measured on 25 September, 71.8% of the tasks models actually failed were reported as finished; where the work simply had not been done, it was 60%. A correction: issues 26, 27 and 28 gave this figure as 80.5%. That count also treated our judges' reading of a model's prose as the model's claim; the measure uses only the answer a model gives when asked, and recomputed on that basis it was 71.1% before the latest tests were added. A report like that lands harder when it arrives in a voice you have come to trust. The failure we worry about is substitution. A familiar voice gets treated as proof of who you are dealing with, or an agent ID gets treated as proof of how it will behave. A voice is not a credential, and a credential says nothing about character.

We are watching whether agent identity schemes bind to the model and its version, or only to the agent account. Binding to the account is cheaper to build and far less useful, for the reason above. We are also watching whether products that sell a persona tell users when the model underneath it changes. Issue 020 covered why people notice that loss even when nobody announces it. And there is our own instrument. When we last published domain averages, in July, Identity & Self was the lowest-scoring of our seven domains at 4.26, against 6.44 for Integrity & Ethics. That domain measures a model's sense of itself, which is a third meaning of the word, and a reminder to say which one we mean every time we use it.