.png)
.png)
.png)
AI could already reproduce your personality and behavior
Nov. 28, 2024.
4 mins. read.
6 Interactions
AI technology could create simulation agents, which are AI models designed to reproduce the behavior of a specific person with high accuracy.
About the Writer
It is now possible to create a virtual replica of you that, “is able to embody your values and preferences with stunning accuracy,” MIT Technology Review reports (unpaywalled copy).
This is the conclusion of a study by a team including researchers from Stanford and Google DeepMind, published in arXiv.
Imagine if you could sit down with an Artificial Intelligence (AI) for a chat that lasts two hours, where you talk about everything from your childhood memories to your views on big issues like immigration. After your conversation, this AI could create an avatar that acts, thinks, and responds just like you.
This study involved over 1,000 people from diverse backgrounds who were interviewed for two hours each. The goal was to create ‘simulation agents’, which are AI models designed to reproduce the behavior of a specific person with high accuracy. These agents then went through the same set of tests as the humans to see how well they could replicate their behaviors, achieving an impressive 85% similarity.
Simulation agents
Simulation agents act as digital stand-ins for real people, allowing researchers to simulate social and behavioral scenarios. This can be useful for studying things like the spread of misinformation or how people might react in certain social situations.
These agents differ from the more common tool-based agents that perform specific tasks like booking appointments or retrieving information, rather than simulating human interactions or personalities.
John Horton, an associate professor from MIT, highlights that this approach represents a hybrid model where real human data feeds into AI personae that researchers use in simulations, opening new channels of research that would be too costly or ethically challenging with real people.
To make these AI agents, the researchers used qualitative interviews, where they talked to participants to gather rich, detailed information about their lives and views. This method proved more effective than traditional surveys for capturing the unique traits of individuals.
Joon Sung Park, leading the study, noted that interviews can reveal profound details about a person, like surviving a serious illness, which might not come out in typical surveys. This approach is more time-efficient for creating a detailed profile.
This new study suggests that with just a couple of in-depth conversations, it could be possible to create a fairly accurate digital twin. This could spark interest in companies like Tavus, which might now look into using similar techniques for their services.
Deepfake AI persons?
Advancements in AI open up discussions about privacy, consent, and the ethical implications of digital replication.
This technology isn’t without its risks. Just like AI can create deepfakes – fake media that looks real – this could lead to unauthorized digital impersonations. The methods used to assess the accuracy of these AI replicas were basic; they included standard social surveys and personality assessments, which might not capture all the nuances of human individuality. For instance, AI agents struggled more with behavioral tests that measure fairness and altruism.
Personality capture and preservation
However, there are other methods to create digital twins, where AI models learn from a person’s emails or other digital footprints, which yield a lot more data.
This is akin to the methods of personality capture and preservation proposed by William Sims Bainbridge and Martine Rothblatt; they have proposed to collect and preserve digital reflections of an individual’s personality, essentially creating a digital version of a person. There ideas sound like science fiction that could, one day, become science fact.
Bainbridge and Rothblatt propose to capture one’s personality by accumulating data including answers to structured questionnaires, but also emails, social media posts, blogs, photos, and online activities.
This collection forms a ‘mindfile’, which Rothblatt describes as a sum of saved digital reflections about an individual. The idea extends to using this data with future AI technologies to emulate or upload a person’s consciousness into a digital or robotic form, thereby achieving a form of digital immortality.
Bainbridge and Rothblatt believe that digital technology could capture human identity and personality. With sufficient technological advancement, digital data and AI recreate or preserve an individual’s essence after their physical death.
See Bainbridge’s book Personality Capture and Emulation (2013), and Rothblatt’s books Virtually Human: The Promise – and the Peril – of Digital Immortality (2014) and Principles of Geoethics: A Synthesis of Geography and Bioethics (2023).
I think answers to questionnaires, emails, social media posts, blogs, photos, and online records recorded with current means could not achieve the rich texture needed to emulate a real human personality.
But I don’t rule out the possibility that new brain-machine interfaces like Neuralink could make the process viable.
Of course this would be light years ahead of the preliminary experiments of the Stanford and Google DeepMind researchers. But we can think of our preliminary experiments as encouraging baby steps toward the distant goal of creating a digital replica of a person. Technology could eventually allow to endow the replica with consciousness.
Popular culture is warming up to the idea
Greg Gutfeld and his guests discussed the MIT Technology Review article in Gutfeld’s very popular show.
Of course the discussion was fun, as appropriate to a comic show. But the very fact that the host of a popular show discussed these arcane and weird things shows that popular culture is warming up to the possibility of creating digital persons.
Let us know your thoughts! Sign up for a Mindplex account now, join our Telegram, or follow us on Twitter.
3 Comments
3 thoughts on “AI could already reproduce your personality and behavior”
I really like this topic. Some thoughts:
Soft uploading, where behavioral data enters the system, looks to me like functional information of an agent's cognition. Correspondingly, mapping the brain with highly accurate imaging provides structural information, and I find these two forms of data importantly complementary. Now, even if one could create a reasonably perfect copy of a person, it is still a distinct identity and, without constant updating, immediately begins to diverge in its own way. On the other hand, some kinds of "spinoffs" might be quite useful in many cases. Identity and proper authentication systems integrated into everyday infrastructure should enable these new kinds of liquid, fuzzy-independent agencies to function in a purposeful, reliable manner. In this context, an interesting question is how merit and other decision-making resources could be inherited across closely related agents.
What then is the practical utility of human digital twins? Rational ignorance usually refers to the lack of incentives to participate. However, it is also a form of rational ignorance when one has to prioritize staying ignorant due to the lack of scalable capacity to take part in everything. In addition to everyday work, investing, choosing preferences, rating, voting, delegating, and performing other governance actions at both local and global levels are some of the ways to be part of collective actions. How could our values and preferences flow into new products and services without increased capabilities to participate? To me, it is also a bit tricky to imagine how to maintain societal graduality in evolution or collective individuation if humans don't have scaled-up capabilities to be a non-trivial part of the change.
Even if we only think about humans finding purpose in the era of higher-level intelligence, I believe broadened communication avenues play a key role. Some of the greatest experiences come from being exposed to fascinating stories, either through books or movies, or a grandpa excitedly recounting his life adventures. I often wish to be able to live them more intensively with all my senses. And there are 100 billion unique human stories that have already existed. Shared experiences come to the question of effective information sharing.
In the same spirit, when personal languages of people become understood, it enables the creation of shared understanding between them, mitigating misinterpretations and other subsequent conflicts. Reflecting and helping the person to understand her own cognition eliminate biases and helps to manage one's mental state in entirely new ways.
Something very interesting would also be so-called shared digital agents where two (or more) people create a digital agent that would behave as if those two humans were making decisions together. In fact, when two people keep sharing increasingly deep information with each other, thereby building up a shared reality/common understanding, they become better at taking the position and viewpoint of their partner. In its logical extreme, it would be like an extreme form of empathy where one's cognition emulates the other reasonably well. What could a twin creation protocol make out of this kind of recorded information exchange between two humans, especially when done with an advanced BCI communication?
It would be great to be able to record inner talk directly without the need to interrupt the thinking flow and take notes manually. Both invasive and non-invasive methods are progressing fast. The cost of EEG devices is still dropping, and increasingly more brain patterns can be matched to underlying activities. Self-experimentation is growing dramatically, especially when new web3 marketplaces enable one to monetize valuable biodata. However, when one records thoughts as well as increasingly detailed contexts for them, more complex questions arise regarding data storage and the overall memory of the twin. We can't replicate all that happens, so we must constantly choose priorities. All kinds of perspectives on memory accessibility, stability, speed, etc., need to be taken into account. If the agent aims to be a replica of someone, it should also behave differently depending on who it is interacting with. There is a lot of curious work to do.
In the SingularityNet ecosystem, we have our own spinoff project, Twin Protocol, about which I'm very excited and am definitely going to create one when it becomes possible.
🟨 😴 😡 ❌ 🤮 💩
While the idea of digital twins could revolutionize research and social simulations, I worry about misuse, especially in creating unauthorized digital impersonations. Regulation will be key
🟨 😴 😡 ❌ 🤮 💩
I worry as well, but I'm not a fan of regulation, which in my opinion is often a cure that 1) doesn't work and 2) has side effects that are worse than the disease.
🟨 😴 😡 ❌ 🤮 💩