We all see AI progress all over the world. There are announcemets of advances here and there. Today I noticed an article AI can predict your personality simply by scanning your eyes, based on this research paper. I would take this as an example for discussion about AI regulation.
The research paper analyses eye movement and its relation to personalities. The study is based on obervation of 42 participants. Their data have been analysed using a machine learning approach (random forest classifier) to asses level of selected personality traits. For this research, each pearsonality trait has been assessed on three levels (low, medium, high) by specified psychological test and by AI, i.e. by the random forest classifier. The results from AI have been compared with the results from psychological tests and the accuracy has been described by an F1 score, a particular measure of classification success. Based on the research paper, the used AI algorithm has mean F1 score of 40%-49% at best. Very roughly if means that the personality trait level has been porperly classified for slightly less than half of the participants. For sure this is better correlation than random choosing of the level, which would give F1 at 33%, and for sure this is relevant result.
Now, the other article says "AI can predict your personality simply by scanning your eyes". Nice, catchy title. For sure it attracts readers. Is it correct? Well, it depends for what purpose. With F1 below 50% there is certainly some level of prediction, but there is still high chance of missclassification. As with many reaserches, the results can help and can endanger. Fortunately, the article is correct when saying that "On the more negative side, this discovery also screams of privacy implications." and that "Although these are fun (and freaky) things to think about, keep in mind that eye movement patterns aren't determinants for these character traits. These are simply tendencies that the researchers found had a correlation.". It is only at the end, but it is there.
This is how it works with AI: it is all about probability, about chance of right classification. And this is where we have to be careful. It would be not wise to prohibit such reasearch. The reasearch is helpful. It enriches our knowledge and it can be a base for a further research. What we need to be careful about is interpretation of the reuslts and prevention of abuse. The reasearch paper is actually sound in disclosing the relevant facts, including calculation of number of participants, written consent, compensation, used methods, approval by Ethics committee etc., so the reader can assess, whether the reasults are appropriate for a specific purpose..
Thoughts about "resposibility" of Artificial Intelligence seem premature in this light. AI does not classify personality in human sense. It just correlates data based on provided sample. It is an algorithm, a calculation with, once trained, predictable result. No magic behind: zeroes, ones, math, algorithm and result. The selection and configuration of the algorithm is kind of magic, but it is under control of the developer. The trained algorithm can be tested and its reliability can be calculated. No "black-box" excuses. AI is a product as any other, with all consequences.
The responsibility for outputs is shared between developers and users, as with any other product. They share responsibility for adequacy of the AI model for specific purpose, for ensuring reliability and for ensuring its compliance and should share the consequences of the decisions. As with other products. Same as with any other computer system, with any other product. It needs to be compliant with all current regulations, including antidiscrimination (no excuse for AI or black box, if it takes shortcut and judges a client based on race, nationality, religion, political believes, etc.), privacy (undisclosed psychological profiles, health conditions), security, etc. We need to protect customers: those who pay by money, those who pay by thier personal data or time, as well as those, who have no other choice than deal with AI because of dominance of the provider or mandatory use. All this without restricting reasearch and development.
Regulators can help the commercial sphere to deal with this responsibility by stating standards for training and for testing AI models, where appropriate. The standards can be differentiated by intended usage of AI model: more strict for healthcare, less strict for shops or social media, more strict when it comes to personality traits, less stict when it comes to wheather prediction, same as with other products. Having standards is on one hand an obligation, but on the other hand it is kind of excuse - the standards set baseline, what is considered sufficient care. Without a standard, any care might be not enough.
Statistics, ML and AI are powerfull tools and we should take care to not abuse them. It is right to make research compliant with applicable ethical and other standards. The real thread is in inappropriate use or abuse of the results. Humankind has already experience with abuse of statistics of observable human properties to personality, intelligence, somebody's options, somebody's life. We should take care to not repeat such mistakes.
- data protection regulation Artificial Intelligence GDPR machine learning Ethics psychometric profiling Machine Ethics
In reply to You have made some very by Jaroslav BARAN