A recent communication from Quilt, the us-based company offering an insurance underwriting based upon a selfie, has been massively shared (especifically in France on this very good blog). More shared anyway than this article published in june about Lapetus, the underlying technology! It is the opportunity to come back to the operating model, the advantages and the drawbacks to keep in mind.
Selfie-insurance: what is it all about?
Lapetus has built an artificial intelligence robot (Chronos) capable, analysing a selfie picture, to identify the major medical risks of the individual.
More precisely, on the basis of some questions (“bio demography” for Lapetus), the tool determines a life expectancy. Then on the basis of the facial analysis, the tool determines the sex, some physiological data, the BMI (Body Mass Index). It can also identify visible markers of more complex diseases, such as diabetes, genetic disorders, or signs of depression, cigarette or drug use.
For an insurer, this is the opportunity, in a photo and some questions (in other words, in a few seconds), to price a health risk. The prospects are very interesting!
Chronos relies on a corpus of 1.5million selfies that have been supplemented with medical or socio-economic data ( Note: I can not find any information on this crucial step and the sources of data provided </ em >). Every day, the information submitted by the candidates completes the modeling.
In other words, the tool captures a certain number of data as input, then compares them with the characteristics of the individuals present in the database and ranks the individual in a category, a risk profile. This profile can then be used by an insurer to determine a rate.
In terms of technology, this is reminiscent of Bayesian Filters (to the origin of anti-spam engines from our mailboxes). These filters, which learn over time, classify the mails in a spam folder if they identify the probability of such a case based on an occurrence of words.
In other words, for a lambda individual in France:
“lucrative” + “investment” + “bamako” + “moneytransfer” = probably spam
“rendez-vous” + “monday” + “2pm” + “office” = Probably not spam
In this case, the shape of the face, the size of the eyes, and so on will probably be identified.
The interest of this technology for insurers is threefold:
- Time saving: As we said, in just a few seconds the insurer gets as much information as it did several minutes ago, or more if medical examinations were needed. </ li>
- Accuracy of Risk Identification : The law of large numbers probably makes the analysis engine quite reliable.
- Fraud Limitation : An individual may lie in his statement, but his physical characteristics can not deceive the engine.
The limits of insurance selfie pricing
As in any use case of artificial intelligence, the Lapetus engine relies on algorithms. These are intended to determine the probability that an individual (in this case his selfie) is suffering from one or more ills.
However, this raises questions!
Like any algorithm, the engine ranks people according to a certain level of probability. Thus, he will be interested in the most common cases. In other words, it will categorize unitarily individuals spread evenly around a normal law. It will therefore tend to harmonize practices and smooth the specificities of each to move from continuous data to discrete data, and therefore to limit the number of output profiles.
More precisely, algorithmic management therefore tends to reduce or even exclude divergences. In this sense, and for the case that interests us, this means that all the individuals who will not have a specific risk will be able to benefit from the service, while the others will have to undergo a premium or return to standard paper process. It’s a risk-selection method that could leave lots of people on the side.
Margin of error
The use of statistical principles is based on 4 cases:
- Positive cases: those for which the engine is certain of the attribution to a profile
- False positives: those who will be assigned to a profile wrongly
- False negatives: those who will not be assigned to a profile, wrongly
- Negatives: those who are not assigned to a profile and for whom we are sure that they should not be part of it.
The goal is to limit false (positive or negative) to avoid mistakes. The law of big numbers again should reduce the risks, but it could reveal methods of anti-selection, counter productive!
Ugly face offence!
Voluntarily provocative, this title is a difficulty. Will it be necessary to take out an insurance tomorrow only a day where one slept well (to avoid having dark circles), absolutely not a day after party, surely not on Monday morning and not the 3rd Monday of January ? What if you do not respond to the canons of beauty and that the engine finds that you look terribly like a compulsive addict lined with a notorious depressive alcoholic?
The difficulty lies once again in the management of false positives (or false negatives, depending on what you are looking for).
Conclusion of selfie insurance
This solution opens up extremely interesting perspectives. Care must be taken, however, to take care of the management of all people who may be at risk. The whole value would be to implement prevention or support procedures to reduce the risk of these individuals, which would improve the user experience of these people.
What do you think? Are you worried or in a hurry to implement this kind of technology?
Contact me to brainstorm with you on your use cases and identify the best ones solutions for your market!