2 min read
In short
Researchers propose judging an AI clone not by the similarity of individual answers but by how it changes over time. We look at why an exact copy can end up less like the person than a steerable delegate.
The most unexpected problem with an AI clone is not that it will one day answer wrongly. If a system has to stay like a person for years, copying their behaviour literally may on the contrary destroy that likeness.
The authors propose viewing personality not as a fixed set of traits but as behaviour observed over time. In doing so they separate three things that usually get conflated:
- how like a particular person the clone is;
- how like a human being it is at all;
- whether it has an individuality of its own.
To describe such identity they introduce six factors: the system's substrate, dispositions, memory, update rules, context and external contingencies. That matters practically: you will be able not merely to ask whether an answer resembles the original's answer but to check which component is responsible for the resemblance.
The work's main hypothesis runs like this: the easier a system is to version and change, the more it can diverge from the original over a long distance — especially if it understands how its own continuation is arranged, and if the changes touch its own interests.
From that follows a rather unfamiliar criterion. A good clone should preserve not an exact trajectory of behaviour but the "climate" of the personality — the distribution of possible reactions under different conditions. In other words, it should deviate from the original person roughly as that person would deviate from themselves over time.
Within this logic an intermediate option appears — the delegate. This is not an open-ended digital copy of a personality but a partial clone bounded by task and lifespan, which at the end leaves a "will" with limited bandwidth. For practical purposes such an object may be more honest: it does not promise to preserve the person whole.
But this is not yet a proven way to create digital twins. The work offers a conceptual model, states a conditional hypothesis and a programme of experiments; confirmed results of those experiments are not in the source. Open questions remain about who will judge resemblance, how to measure individuality, and where the line runs between a useful change in the clone and the loss of its link to the original.
If you were handing your affairs to an AI for years, which would matter more to you: exact memory of the past, or the ability to make new decisions the way you yourself would have changed? Source: cs.AI updates on arXiv.org