by Gautam Hazari
The Ship of Theseus and Identity in the Agentic AI world
The Ship of Theseus, and all the interesting variations like the Grandfather’s Axe, Trigger’s Broom and similarities like the Buddhist text Da Zhidu Lun, is a fascinating thought experiment, one I’ve talked about a few times in the past – https://www.linkedin.com/pulse/ship-theseus-identity-paradox-gautam-hazari/ and https://sekura.id/ship-of-theseus/
The Ship of Theseus presents an interesting dilemma, posing a question without necessarily answering it, which makes it even more intriguing.
The quest for almost everything we do has been to find answers. Still, with the commodification of information and knowledge, and now the same with intelligence, especially with GenAI models, the status quo is being challenged.
Questions represent creativity; questions have more degrees of freedom than answers. They have been one of the most critical tools of our species, which has led us to where we are now. And when ChatGPT has all the answers, who creates the questions?
So, that brings me to: What’s the pertinent question in the Ship of Theseus thought experiment?
Theseus is presented as a mythical king and the founder of Athens, a Greek hero. The ship was sailed by Theseus through a battle and was kept in a harbour as a museum. As time passes, the wooden parts of the ship deteriorate and rot, and are then repaired and replaced with new ones.
The paradoxical question raised here is about the identity of the Ship of Theseus over time: Is it the same ship after its repaired? Is the identity of the ship the same after the parts have been replaced with new ones? Has the identity of the ship persisted over time?
At the very core of this paradox is a fundamental question about the dynamism of identity and highlighting a binary contrast between this dynamism over time and permanence, along with a silent assertion on the pluralism of identity, the ship being the composition of its parts.
In one of my earlier writeups (https://sekura.id/ship-of-theseus/), I proposed the key Principles of Identity against the backdrop of the Ship of Theseus:
- Identity is plural
- Identity is a composition of plurality into a singularity
- Identity is contextual
- Identity is dynamic
- Identity is invisible
Identity has always been a critical unsolved problem in the digital world, as the Internet was never created with an identity layer; this is the identity crisis of the Internet, as I always call it.
These identity principles are even more critical now in the post-ChatGPT digital world, where AI has evolved significantly from Predictive AI to Generative AI and now into Agentic AI, and in the near future – Physical AI.
Let’s explore a bit more about the constituents of identity, particularly digital Identity. Identity, as narrated in ISO/IEC 24760 and referred to in ISO 29115 clause 3.13 as the “set of attributes related to an entity”.
In the case of digital identity, generally these attributes are first name, last name, age, gender, address, nationality, ethnicity, marital status, etc. and could be several other attributes based on the context in which the identity needs to be established.
These attributes have been at the core of various processes in the digital world, from identity proofing (KYC) to credit scores and identity verification.
The KYC process can be traced back to the passage of the US Bank Secrecy Act in 1970. However, the objective was somewhat specific – for financial institutions to have systems in place to detect and report suspicious activities. The Bank of England introduced the first proper KYC guidelines in the early ‘90s.
The interesting thing is that the process, or rather the attributes in the process with establishing the identity, has remained more or less the same. However, one can argue that these attributes are obvious, and so also is the persistence of the attributes.
One critical observation of most of these attributes about identity is that they are mostly static concerning time. With the acceleration of AI and ML models evolving with multiple architectures, many of these static attributes can be generated by training these models using various accessible data sources.
Let’s look into some of the approaches and mechanisms used to generate some of these attributes.
One of the key approaches is MiVOLO (https://arxiv.org/abs/2307.04616).
Introduced in July 2023 is one of the most popular approaches for age and gender identification from images. MiVOLO (Multi Input VOLO) is an extension to the VOLO (Vision Outlooker for Visual Recognition) (https://arxiv.org/pdf/2106.13112), published in 2021, which used the transformer architecture for vision recognition. The accuracy is achieved up to 98.3% for gender and 68.69% for age depending on the dataset used.
There have also been various approaches used by using the streamed data from the sensors in mobile devices to estimate the age group, sensors like the accelerometer, gyroscope and orientation sensors are used to train ML models. One such approach was published in a paper titled: “Age Group Detection Using Smartphone Motion Sensors” (https://new.eurasip.org/Proceedings/Eusipco/Eusipco2017/papers/1570346241.pdf).
Although, this was pre-GenAI world then, the approach managed to achieve around 89% accuracy in some experimental setups, using simplistic classification algorithms like Random Forest Tree, Logistic Regression etc.
Interesting approaches have been used to predict demographics attributes using music. A paper titled: “Predicting user demographics from music listening information” (https://link.springer.com/article/10.1007/s11042-018-5980-y).
The approach achieved more than 95% accuracy for age prediction, 81% accuracy in gender prediction, 70% accuracy in prediction the country. Considering, this was done in 2018, – the pre-ChatGPT/pre-Gen AI world – the potential of the approach is remarkable.
There have been approaches used to predict many other attributes, including education, ethnicity, income, parental status and even political preference using accessible data like website traffic data and Twitter (now X) feeds: https://jair.org/index.php/jair/article/view/10984.
Predicting demographic attributes – which are typically stable or change in predictable ways – has been done for years. With advances in AI, machine learning, and computing power, those predictions are only getting more accurate.
Here are some of the approaches from the pre-ChatGPT/pre-GenAI/pre-Agentic AI era:
- 2012 – Prediction of gender, marital status and other attributes – Demographic Prediction Based on User’s Mobile Behaviours (https://www.idiap.ch/project/mdc/publications/files/mdc-final241-ying.pdf)
- 2015 – Prediction of gender and other attributes using the data on installed apps on the mobile device (https://dl.acm.org/doi/abs/10.1145/2721896.2721908)
- 2016 – Prediction of age, race and income based on installed apps on the mobile device (https://ojs.aaai.org/index.php/icwsm/article/view/14776)
- 2017 – Prediction of gender, age, ethnicity, level of education, disabilities, employment and socio economic status using smartphone sensor data (https://eecs.wsu.edu/~holder/msgm/pdfs/AkterNDA2017.pdf)
- 2018 – Predicting demographic attributes using just the name and Twitter handle (https://pure.johnshopkins.edu/en/publications/predicting-twitter-user-demographics-from-names-alone)
Many of the attributes which are generally static in nature in relation to time, can be generated using ML models and can even be generated with acceptable accuracy using GenAI models.
Also, in the Agentic AI world – there can be AI agents created which can autonomously look for the most appropriate data sources which could be beyond common sense and then let the GenAI models to train the models and also act as external knowledge bases using RAG (Retrieval Augmented Generation).
One important aspect to note here is if these attributes can be generated, then they can be manipulated as well, giving rise to multiple dimensions of identity-related issues, from synthetic identity fraud to privacy issues, as these attributes can be “known” without the explicit or even implicit knowledge of the user/entity.
Synthetic identity fraud has cost £300 million in UK alone, as per UK Finance (https://www.ukfinance.org.uk/news-and-insight/blog/why-synthetic-identity-fraud-detection-must-go-beyond-onboarding). Lexis-Nexis predicts a rise to £4.2 billion by 2027.
KPMG suggests it’s a $6 billion problem.
So now, let’s now come back to the principles of dynamism of identity, and, as was posed as the paradoxical question in the Ship of Theseus, dynamism of identity with respect to time needs to be centre stage.
This is where the attributes and assets of the mobile networks become much more critical.
Many of the attributes are dynamic and temporal; although many of these attributes may not be traditional identity attributes, these dynamic attributes can redefine identity attributes and add practical confidence in the age of GenAI and Agentic AI.
Attributes like SIM Swap date-timestamp (when did the last SIM Swap happen?), Device Swap (when did the last device change happen?), Call Forward status (is there a conditional or unconditional call forward active?), Account Tenure (how long the user/entity has been using the account for), Device Type (what device the user/entity has been using at the moment, does this device fit within the normal usage of the user?) etc. are all dynamic with respect to time and are extremely useful attributes in this post-Gen AI and Agentic AI world.
With the industry’s focus on network APIs – including the entire supply chain ecosystem – from the MNOs as the ultimate suppliers, to the channel partners like Sekura.id/XConnect and also the consumers’ ecosystems, including the hyperscalers like Google and Microsoft, the GSMA Open Gateway initiative along with CAMARA from Linux Foundation, and infusing acceleration by the joint ventures like Aduna – there has not been a better time to solve the evolved identity crisis of the digital world using the dynamic attributes set, and make the digital world a better and safer place.
It is fascinating, then, that a thought experiment from the 1st century AD, the Ship of Theseus, which presented a paradoxical question, can be used as a backdrop to help solve some of the critical problems of the digital world, created without an Identity layer.
We at Sekura.id/XConnect are passionately working towards solving the identity crisis of the Internet, adapting to the future of the Gen AI and the Agentic AI worlds.
Let’s continue to work towards making the digital world a SAFr place, together.
Gautam Hazari – CTO, Sekura.id
Gautam Hazari is the architect behind mobile identity APIs and a driving force in digital trust. Like the Ship of Theseus, his work poses a vital question: if every part of our digital identity changes, what still makes it us? At Sekura.id, he fuses deep technical expertise with a human-first philosophy to deliver seamless, privacy-led identity solutions for the modern world.
